For a real world example of the challenges of harnessing LLMs, look at Apple. Over a year ago they had a big product launch focused on "Apple Intelligence" that was supposed to make heavy use of LLMs for agentic workflows. But all we've really gotten since then are a couple of minor tools for making emojis, summarizing notifications, and proof reading. And they even had to roll back the notification summaries for a while for being wildly "out of control". [1] And in this year's iPhone launch the AI marketing was toned down significantly.
I think Apple execs genuinely underestimated how difficult it would be to get LLMs to perform up to Apple's typical standards of polish and control.
> to perform up to Apple's typical standards of polish and control.
i no longer believe they have kept on to the standards in general. the ux/ui used to be a top priority, but the quality control has certainly gone down over the years [1]. the company is now driven by supply chain and business-minded optimizations than what to give to the end user.
at the same time, what one can do using AI has large correlation with what one does with their devices in the first place. a windows recall like feature for ipad os might have been interesting (if not equally controversial), but not that useful because even till this day it remains quite restrictive for most atypical tasks.
>> to perform up to Apple's typical standards of polish and control.
>i no longer believe they have kept on to the standards in general.
One 100% agree with this, if I compare AI's ability to speed up the baseline for me in terms of programming Golang (hard/tricky tasks clearly still require human input - watch out for I/O ops) with Apple's lack of ability to integrate it in even the simplest of ways.. things are just peculiar on the Apple front. Bit similar to how MS seems to be gradually loosing the ability to produce a version of Windows that people want to run due to organisational infighting.
Personally, I’ve never seen an AI flow of any kind that meets what would meet the quality of a typical ‘corporate’ acceptable flow. As in, reliably works, doesn’t go crazy randomly, etc.
I’ve seen a lot of things that look like they’re working for a demo, but shortly after starting to use it? Trash. Not every time (and it’s getting a little better), but often enough that personally I’ve found them a net drain on productivity.
And I literally work in this space.
Personally, I find apples hesitation here a breath of fresh air, because I’ve come to absolutely hate Windows - and everybody doing vibe code messes that end up being my problem.
With Apple it's incredibly obvious that most software product development is nowadays handled by outsourced/offshored contractors who simply do not use the products. At least I hope that's the case, it would be disastrous if the state of iOS/watchOS is the result of their in-house on-shore talent.
Like most AI products it feels like they started with a solution first and went searching for the problems. Text messages being too long wasn't a real problem to begin with.
There are some good parts to Apple Intelligence though. I find the priority notifications feature works pretty well, and the photo cleanup tool works pretty well for small things like removing your finger from the corner of a photo, though it's not going to work on huge tasks like removing a whole person from a photo.
> it's not going to work on huge tasks like removing a whole person from a photo.
I use it for removing people who wander into the frame quite often. It probably wont work for someone close up, but its great for removing a tourist who spends ten minutes taking selfies in front of a monument.
Honestly I love the priority notifications and the notification summaries. The thing that drives me absolutely insane, is that the fact that when I view the notification through clicking on it from another space other than the "While in the reduce interruptions focus" it doesn't clear. Because of this, I always have infinite notifications.
I want to open WhatsApp and open the message and have it clear the notif. Or atleast click the notif from the normal notif center and have it clear there. It kills me
Do you know if apple is using their new tools to do mail filtering? It's an interesting choice if they are since it's a genuine problem with a mature (but always evolving) solution.
Well yeah, but that's in part a problem with always-on doorbell cameras. On paper they're illegal in many countries (privacy laws, you can't just put up a camera and record anyone out in public), in practice the police asks people to put their doorbell cameras in a registry so they can request footage if needs be.
Anyway, I get wanting to see who's ringing your doorbell in e.g. apartment buildings, and that extending to a house, especially if you have a bigger one. But is there a reason those cameras need to be on all the time?
At least in the USA it’s legal to record public spaces. So recording the street and things that can be seen from it is legal, but pointing a camera over your neighbors fence is not.
And a lot of people don't share that opinion, so this isn't the law in a lot of countries. When you wanted to suggest that it is a problem, that US companies try to extend the law of there home country to other parts of the world, then I endorse that.
it isn't creepy, it's super annoying if you don't live in the woods. got a ring doorbell and turned them off a few hours after installation, it was driving me nuts.
That makes... That makes just enough sense to become nonsense, rather than mere noise.
I mean, I could imagine a person with no common sense almost making the same mistake: "I have a list of 5 notifications of a person standing on the porch, and no notifications about leaving, so there must be a 5 person group still standing outside right now. Whadya mean, 'look at the times'?"
> A biologist, a physicist and a mathematician were sitting in a street cafe watching the crowd. Across the street they saw a man and a woman entering a building. Ten minutes they reappeared together with a third person.
> - They have multiplied, said the biologist.
> - Oh no, an error in measurement, the physicist sighed.
> - If exactly one person enters the building now, it will be empty again, the mathematician concluded.
It does feel like somebody forgot that "from the first sentence or two of the email, you can tell what it's about" was already a rule of good writing...
Maybe they remembered that a lot of people aren't actually good writers. My brother will send 1000 word emails that meander through subjects like what he ate for breakfast to eventually get to the point of scheduling a meeting about negotiating a time for help with moving a sofa. Mind you, I see him several times a week so he's not lonely, this is just the way he writes. Then he complains endlessly about his coworkers using AI to summarize his emails. When told that he needs to change how he writes to cut right to the point, he adopts the "why should I change, they're the ones who suck" mentality.
So while Apple's AI summaries may have been poorly executed, I can certainly understand the appeal and motivation behind such a feature.
I mean...this depends very heavily on what the purpose of the writing is.
If it's to succinctly communicate key facts, then you write it quickly.
- Discovered that Bilbo's old ring is, in fact, the One Ring of Power.
- Took it on a journey southward to Mordor.
- Experienced a bunch of hardship along the way, and nearly failed at the end, but with Sméagol's contribution, successfully destroyed the Ring and defeated Sauron forever.
....And if it's to tell a story, then you write The Lord of the Rings.
There’s a thread here that could be pulled - something about using AI to turn everyone into exactly who you want to communicate with in the way you want.
Probably a sci-fi story about it, if not, it should be written.
I think people read texts because they want to read them, and when they don't want to read the texts they are also not even interested in reading the summaries.
Why do I think this? ...in the early 2000's my employer had a company wide license for a document summarizer tool that was rather accurate and easy to use, but nobody ever used it.
The obvious use case is “I don’t want to read this but I am required to read this (job)” - the fact that people don’t want to use it even there is telling, imo.
Even bending over that far backwards to find a useful example comes up empty.
Those kinds of emails are so uncommon they’re absolutely not worth wasting this level of effort on. And if you’re in a sorry enough situation where that’s not the case, what you really need is the outside context the model doesn’t know. The model doesn’t know your office politics.
No one cares about the terms of service. And if they actually do, they will need to read every word very carefully to know if they are in legal trouble. A possibly wrong summary of a terms of service document is entirely and completely useless.
It's not even that they are useless, they are actively wrong. I could post pages upon pages of screenshots of the summaries being literally wrong about the content of the messages it summarised.
My wife was in China recently and was sending back pictures of interesting things - one came in while I was driving and my iPhone read out a description of the picture that had been sent - "How cool is that!" I thought.
However, when I stopped driving and looked at the picture the AI generated description was pretty poor - it wasn't completely wrong but it really wasn't what I was expecting given the description.
It’s been surprisingly accurate at times “a child holding an apple” in a crowded picture, and then sometimes somewhat wrong.
What really kills me is “a screenshot of a social media post” come on it’s simple OCR read the damn post to me you stupid robot! Don’t tell me you can’t, OCR was good enough in the 90s!
The description said "People standing in front of impressive scenery" (or something like that) - it got the scenery part correct but the people are barely visible and really small.
Apple's whole brand is built around tight control, predictable behavior, and a super polished UX which is basically the opposite of how LLMs behave out of the box
> I think Apple execs genuinely underestimated how difficult it would be to get LLMs to perform up to Apple's typical standards of polish and control
Not only Apple, this is happening across the industry. Executives' expectations of what AI can deliver are massively inflated by Amodei et al. essentially promising human-level cognition with every release.
The reality is aside from coding assistants and chatbot interfaces (a la chatgpt) we've yet to see AI truly transform polished ecosystems like smartphones and OS' for a reason.
Which is ironic, given all I really want from Siri is an advanced-voice-chat-level chat gpt experience - being able to carry on about 90% of a natural conversation with gpt, while Siri vacillates wildly between 1) simply not responding 2) misunderstanding and 3) understand but refusing to engage - feels awful.
The thought that a company like Apple, which surely put hundreds of engineers to work on these tools and went through multiple iterations of their capabilities, would launch the capabilities...Only for its executives to realize after release that current AI is not mature enough to add significant commercial value to their products, is almost comical.
The reality is that if they hadn’t announced these tools and joined the make-believe AI bubble, their stock price would have crashed. It’s okay to spend $400 million on a project, as long as you don’t lose $50 billion in market value in an afternoon.
I'm happy they ate shit here because I like my mac not getting co-pilot bullshit forced into it, but apparently Apple had two separate teams competing against each other on this topic. Supposedly a lot of politics got in the way of delivering on a good product combined with the general difficulty of building LLM products.
I do prefer that Apple is opting to have everything run on device so you aren’t being exposed to privacy risks or subscriptions. Even if it means their models won’t be as good as ones running on $30,000 GPUs.
If you have say 16GB of GPU RAM and around 64GB of RAM and a reasonable CPU then you can make decent use of LLMs. I'm not a Apple jockey but I think you normally have something like that available and so you will have a good time, provided you curb your expectations.
I'm not an expert but it seems that the jump from 16 to 32GB of GPU RAM is large in terms of what you can run and the sheer cost of the GPU!
If you have 32GB of local GPU RAM and gobs of RAM you can rub some pretty large models locally or lots of small ones for differing tasks.
I'm not too sure about your privacy/risk model but owning a modern phone is a really bad starter for 10! You have to decide what that means for you and that's your thing and your's alone.
It also means that when the VC money runs dry, it's sustainable to run those models on-device vs. losing money running on those $$$$$ GPUs (or requiring consumers to opt for expensive subscriptions).
> Apple had two separate teams competing against each other on this topic
That is a sign of very bad management. Overlapping responsibilities kill motivation as winning the infighting becomes more important than creating a good product. Low morale, and a blaming culture is the result of such "internal competition". Instead, leadership should do their work and align goals, set clear priorities and make sure that everybody rows in the same direction.
It’s how Apple (relatively famously?) developed the iPhone, so I’d assume they were using this as a model.
> In other words, should he shrink the Mac, which would be an epic feat of engineering, or enlarge the iPod? Jobs preferred the former option, since he would then have a mobile operating system he could customize for the many gizmos then on Apple’s drawing board. Rather than pick an approach right away, however, Jobs pitted the teams against each other in a bake-off.
But that's not the same thing right? That means having two teams competing for developing the next product. That's not two organisations handling the same responsibilities. You may still end up in problems with infighting. But if there is a clear end date for that competition and then no lasting effects for the "losers" this kind of "competition" will have very different effects than setting up two organisations that fight over some responsibility
> Distrust between the two groups got so bad that earlier this year one of Giannandrea’s deputies asked engineers to extensively document the development of a joint project so that if it failed, Federighi’s group couldn’t scapegoat the AI team.
> It didn’t help the relations between the groups when Federighi began amassing his own team of hundreds of machine-learning engineers that goes by the name Intelligent Systems and is run by one of Federighi’s top deputies, Sebastien Marineau-Mes.
This is a pretty good article, and worth reading if you aren't aware that Apple has seemingly mostly abandoned the vision of on-device AI (I wasn't aware of this)
There are four main levers for improving an ML system:
1. You can change the training data.
2. You can change the objective function.
3. You can change the network topology.
4. You can change various hyperparameters (learning rate, etc.).
From there, I think it is better to look at the process as one of scientific discovery rather than a software debugging task. You form hypotheses and you try to work out how test them by mutating things in one of the four categories above. The experiments are expensive and the results are noisy, since the training process is highly randomized. A lot of times the effect sizes are so small it is hard to tell if they are real. The universe of potential hypotheses is large, and if you test a lot of them, you have to correct for the chance that some will look significant just by luck. But if you can add up enough small, incremental improvements, they can produce a total effect that is large.
The good news is that science has a pretty good track record of improving things over time. The bad news is that it can take a lot of time, and there is no guarantee of success in any one area.
While it’s possible to demonstrate the safety of an AI for
a specific test suite or a known threat, it’s impossible
for AI creators to definitively say their AI will never act
maliciously or dangerously for any prompt it could be given.
This possibility is compounded exponentially when MCP[0] is used.
I wonder if a safer approach to using MCP could involve isolating or sandboxing the AI. A similar context was discussed in Nick Bostrom's book Superintelligence. In the book, the AI is only allowed to communicate via a single light signal, comparable to Morse code.
Nevertheless, in the book, the AI managed to convince people, using the light signal, to free it. Furthermore, it seems difficult to sandbox any AI that is allowed to access dependencies or external resources (i.e. the internet). It would require (e.g.) dumping the whole Internet as data into the Sandbox. Taking away such external resources, on the other hand, reduces its usability.
Yeah in that regard we should always treat it like a junior something. Very much like you can't expect your own kids to never do something dangerous even if tell it for years to be careful. I got used to getting my kid from the Kindergarten with a new injury at least once a month.
I think it's very dangerous to use the term "junior" here because it implies growth potential, where in fact it's the opposite: you are using a finished product, it won't get any better. AI is an intern, not a junior. All the effort you're spending into correcting it will leave the company, either as soon as you close your browser or whenever the manufacturer releases next year's model -- and that model will be better regardless of how much time you waste on training this year's intern, so why even bother? Thinking of AI as a junior coworker is probably the least productive way of looking at it.
We should move well beyond human analogies.
I have never met a human that would straight up lie about something, or build up so much deceptive tests that it might as well be lying.
Granted this is not super common in these tools, but it is essentially unheard of in junior devs.
> I have never met a human that would straight up lie about something
This doesn't match my experience. Consider high profile things like the VW emissions scandal, where the control system was intentionally programmed to only engage during the emissions test. Dictators. People are prone to lie when it's in their self interest, especially for self preservation. We have entire structures of government, courts, that try to resolve fact in the face of lying.
If we consider true-but-misleading, then politics, marketing, etc. come sharply into view.
I think the challenge is that we don't know when an LLM will generate untrue output, but we expect people to lie in certain circumstances. LLMs don't have clear self-interests, or self awareness to lie with intent. It's just useful noise.
There is an enormous amount of difference between planned deception as part of a product, and undermining your own product with deceptive reporting about its quality. The difference is collaboration and alignment. You might have evil goals, but if your developers are maliciously incompetent, no goal will be accomplished.
> it’s impossible for AI creators to definitively say their AI will never act maliciously or dangerously for any prompt it could be given
This is false, AI doesn't "act" at all unless you, the developer, use it for actions. In which case it is you, the developer, taking the action.
Anthropomorphizing AI with terms like "malicious" when they can literally be implemented with a spreadsheet—first-order functional programming—and the world's dumbest while-loop to append the next token and restart the computation—should be enough to tell you there's nothing going on here beyond next token prediction.
Saying an LLM can be "malicious" is not even wrong, it's just nonsense.
> bugs are usually caused by problems in the data used to train an AI
I think a fundamental problem is that many people assume that an LLM's failure to correctly perform a task is a bug that can be fixed somehow. Often times, the reason for that failure is simply a property of the AI systems we have at the moment.
When you accidentally drop a glass and it breaks, you don't say that it's a bug in gravity. Instead, you accept that it's a part of the system you're working with. The same applies to many categories of failures in AI systems: we can try to reduce them, but unless the nature of the system fundamentally changes (and we don't know if or when that will happen), we won't be able to get rid of them.
"Bug" carries an implication of "fixable" and that doesn't necessarily apply to AI systems.
But it's worse than that. Even if in theory the system could be fixed, we don't actually know how to fix it for real, the way we can fix a normal computer program.
The reason we can't fix them is because we have no idea how they work; and the reason we have no idea how they work is this:
1. The "normal" computer program, which we do understand, implement a neural network
2. This neural network is essentially a different kind of processor. The "actual" computer program for modern deep learning systems is the weights. That is, weights : neural net :: machine language : normal cpu
3. We don't program these weights; we literally summon them out of the mathematical aether by the magic of back-propagation and gradient descent.
This summoning is possible because the "processor" (the neural network architecture) has been designed to be differentiable: for every node we can calculate the slope of the curve with respect to the result we wanted, so we know "The final output for this particular bit was 0.7, but we wanted it to be 1. If this weight in the middle of the network were just a little bit lower, then that particular output would have been a little bit higher, so we'll bump it down a bit."
And that's fundamentally why we can't verify their properties or "fix" them the way we can fix normal computer programs: Because what we program is the neural network; the real program, which runs on top of that network, is summoned and not written.
That’s a very poetic description. Mine is simpler. It’s a generator. You train it to give it the parameters (weights) of the formula thag generate stuff (the formula is known). Then you give it some input data, and it will gives you an output.
Both the weights and the formula is known. But the weight are meaningless in a human fashion. This is unlike traditional software where everything from encoding (the meaning of the bits) to how the state machine (the cpu) was codified by humans.
The only ways to fix it (somewhat) is to come up with better training data (hopeless), a better formula, or tacking something on top to smooth the worst errors (kinda hopeless).
The correct way to fix it would be to build a decompiler to normal code, that would explain what it does, but this is akin to building the everything machine.
The part about AI being very sensitive to small perturbations of their input is actually a very active research topic (and coincidentally the subject of my PhD). Most vision AIs suffer from poor spatial robustness [1], you can drastically lower their accuracy simply by translating the inputs by well-chosen (adversarial) translations of a few pixels! I don't know much about text processing AIs but I can imagine their semantic robustness is also studied.
> bugs are usually caused by problems in the data used to train an AI
This also is a misunderstanding.
The LLM can be fine, the training and data can be fine, but because the LLMs we use are non-deterministic (at least in regard to their being intentional attempts at entropy to avoid always failing certain scenarios) current algorithms are inherently by-design not going to always answer every question correctly that it potentially could have if the values that fall within a range had been specific values for that scenario. You roll the dice on every answer.
This is not necessarily a problem. Any programming or mathematical question has several correct answers. The problem with LLMs is that they don't have a process to guarantee that a solution is correct. They will give a solution that seems correct under their heuristic reasoning, but they arrived at that result in a non-logical way. That's why LLMs generate so many bugs in software and in anything related to logical thinking.
>> a solution that seems correct under their heuristic reasoning, but they arrived at that result in a non-logical way
Not quite ... LLMs are not HAL (unfortunately). They produce something that is associated with the same input, something that should look like an acceptable answer. A correct answer will be acceptable, and so will any answer that has been associated with similar input. And so will anything that fools some of the people, some of the time ;)
The unpredictability is a huge problem. Take the geoguess example - it has come up with a collection of "facts" about Paramaribo. These may or may-not be correct. But some are not shown in the image. Very likely the "answer" is derived from completely different factors, and the "explanation" in spurious (perhaps an explanation of how other people made a similar guess!)
The questioner has no way of telling if the "explanation" was actually the logic used. (It wasn't!) And when genuine experts follow the trail of token activation, the answer and the explanation are quite independent.
> Very likely the "answer" is derived from completely different factors, and the "explanation" in spurious (perhaps an explanation of how other people made a similar guess!)
This is very important and often overlooked idea. And it is 100% correct, even admitted by Anthropic themselves. When user asks LLM to explain how it arrived to a particular answer, it produces steps which are completely unrelated to the actual mechanism inside LLM programming. It will be yet another generated output, based on the training data.
> Any programming or mathematical question has several correct answers.
Huh? If I need to sort the list of integer number of 3,1,2 in ascending order the only correct answer is 1,2,3. And there are multiple programming and mathematical questions with only one correct answer.
If you want to say "some programming and mathematical questions have several correct answers" that might hold.
I think what they meant is something along the lines of:
- In Math, there's often more than one logically distinct way of proving a theorem, and definitely many ways of writing the same proof, though the second applies more to handwritten/text proofs than say a proof in Lean.
- In programming, there's often multiple algorithms to solve a problem correctly (in the mathematical sense, optimality aside), and for the same algorithm there are many ways to implement it.
LLMs however are not performing any logical pass on their output, so they have no way of constraining correctness while being able to produce different outputs for the same question.
I find it quite ironical that while discussing the topic of logic and correct answers the OP talks rather "approximately" leaving the reader to imagine what he meant and others (like you) to spell it out.
Yes, I thought as well of your interpretation, but then I read the text again, and it really does not say that, so I choose to answer to the text...
No, but if you phrase it like "there are multiple correct answers to the question 'I have a list of integers, write me a computer program that sorts it'", that is obviously true. There's an enormous variety of different computer programs that you can write that sorts a list.
I think more charitably, they meant either that 1. There is often more than one way to arrive at any given answer, or 2. Many questions are ambiguous and so may have many different answers.
Is it? You have three wishes, which the maliciously compliant genie will grant you. Let’s hear your unambiguous request which definitely can’t be misinterpreted.
> The problem with LLMs is that they don't have a process to guarantee that a solution is correct
Neither do we.
> They will give a solution that seems correct under their heuristic reasoning, but they arrived at that result in a non-logical way.
As do we, and so you can correctly reframe the issue as "there's a gap between the quality of AI heuristics and the quality of human heuristics". That the gap is still shrinking though.
I'll never doubt the ability of people like yourself to consistently mischaracterize human capabilities in order to make it seem like LLMs' flaws are just the same as (maybe even fewer than!) humans. There are still so many obvious errors (noticeable by just using Claude or ChatGPT to do some non-trivial task) that the average human would simply not make.
And no, just because you can imagine a human stupid enough to make the same mistake, doesn't mean that LLMs are somehow human in their flaws.
> the gap is still shrinking though
I can tell this human is fond of extrapolation. If the gap is getting smaller, surely soon it will be zero, right?
> doesn't mean that LLMs are somehow human in their flaws.
I don't believe anyone is suggesting that LLMs flaws are perfectly 1:1 aligned with human flaws, just that both do have flaws.
> If the gap is getting smaller, surely soon it will be zero, right?
The gap between y=x^2 and y=-x^2-1 gets closer for a bit, fails to ever become zero, then gets bigger.
The difference between any given human (or even all humans) and AI will never be zero: Some future AI that can only do what one or all of us can do, can be trivially glued to any of that other stuff where AI can already do better, like chess and go (and stuff simple computers can do better, like arithmetic).
> I'll never doubt the ability of people like yourself to consistently mischaracterize human capabilities
Ditto for your mischaracterizations of LLMs.
> There are still so many obvious errors (noticeable by just using Claude or ChatGPT to do some non-trivial task) that the average human would simply not make.
Firstly, so what? LLMs also do things no human could do.
Secondly, they've learned from unimodal data sets which don't have the rich semantic content that humans are exposed to (not to mention born with due to evolution). Questions that cross modal boundaries are expected to be wrong.
> If the gap is getting smaller, surely soon it will be zero, right?
I think sometimes it gives a "wrong" answer not because it wasn't trained well, but because it could give multiple plausible answers and just happened to land on the unhelpful one
> Because eventually we’ll iron out all the bugs so the AIs will get more reliable over time
Honestly this feels like a true statement to me. It's obviously a new technology, but so much of the "non-deterministic === unusable" HN sentiment seems to ignore the last two years where LLMs have become 10x as reliable as the initial models.
They have certainly gotten better, but it seems to me like the growth will be kind of logarithmic. I'd expect them to keep getting better quickly for a few more years and then kinda slow and eventually flatline as we reach the maximum for this sort of pattern matching kind of ML. And I expect that flat line will be well below the threshold needed for, say, a small software company to not require a programmer.
It will take at least a few more decades at least from the looks of it. I would be 6 feet under by then so yes "Nothing will ever be able to do my job".
Since "AGI has been achieved internally" tweet I've only seen incremental improvements that are guaranteed to never be able to do my job. Or most people's jobs.
Ironically I came to the comments to point out that all over hackernews you see this sentiment repeated, and that's by a group I would consider to be far more technically competent then your average person. And very helpfully there is one just a few comments down from the top.
Technical competence and an interest in sociological development do not always coincide. Technology often seeks simplicity, whereas sociology examines inherently complex human behavior.
By the time a junior dev graduates to senior, I expect that they'll be more reliable. In fact, at the end of each project, I expect the junior dev to have grown more reliable.
LLMs don't learn from a project. At best, you learn how to better use the LLM.
They do have other benefits, of course, i.e. once you have trained one generation of Claude, you have as many instances as you need, something that isn't true with human beings. Whether that makes up for the lack of quality is an open question, which presumably depends on the projects.
How long do you think that will remain true? I've bootstrapped some workflows with Claude Code where it writes a markdown file at the end of each session for its own reference in later sessions. It worked pretty well. I assume other people are developing similar memory systems that will be more useful and robust than anything I could hack together.
For LLMs? Mostly permanently. This is a limitation of the architecture. Yes, there are workarounds, including ChatGPT's "memory" or your technique (which I believe are mostly equivalent), but they are limited, slow and expensive.
Many of the inventors of LLMs have moved on to (what they believe are) better models that would handle such learnings much better. I guess we'll see in 10-20 years if they have succeeded.
I'm not sure that's generally true. However, older people have a track record, and a reliable older person is likely to be more reliable than a younger person without such a track record.
Reliable has different meanings. I think in this case the meaning is closer to "deterministic" and "follows instructions." An older worker will more reliably behave the same way twice, and more reliably follow the same set of instructions they've been following throughout their career.
> To make this more concrete, here are some example ideas that are perfectly true when applied to regular software but become harmfully false when applied to modern AIs: ...
Hmm, I don't think any of these were true with non-AI software. Commonly held beliefs, sure.
If anything, I am glad AI is helping us revisit these assumptions.
- Software vulnerabilities are caused by mistakes in the code
Setting aside social engineering, mistake implies these were knowable in advance. Was the lack of TLS in the initial HTTP spec a mistake?
- Bugs in the code can be found by carefully analysing the code
If this was the case, why do people reach for rewriting buggy code they don't understand?
- Once a bug is fixed, it won’t come back again
Too many counter examples to this one in my lived experience.
- Every time you run the code, the same thing happens
Setting aside seeding PRNGs, there's the issue of running the code on different hardware. Or failing hardware.
- If you give specifications beforehand, you can get software that meets those specifications
I have never seen this work without needing to revise the specification during implementation.
My current method for trying to break through this misconception is informing people that nobody knows how AI works. Literally. Nobody knows. (Note that knowing how to make something is not the same as knowing how it works. Take humans as an obvious example.)
I think you’re confusing knowing how a system works with being able to predict how that system will perform.
In a non-linear system the former is often easier than the latter. For example we know how planets “work” from the laws of motion. But planetary orbits involving > 2 bodies are non-linear, and predicting their motion far into the future is surprisingly difficult.
Neural networks are the same. They’re actually quite simple, it’s all undergraduate maths and statistics. But because they’re non-linear systems, predicting their behaviour is practically impossible.
I think there are people who know exactly how it works, they know how neural networks work, they know how the transformer architecture works, how attention works, embeddings, tokenization, etc. We just can’t define the weights of the connections between the neurons.
There's no magic involved, the LLM creators can go anywhere and rebuild an LLM with pretty much the same outcome, if they have the same training data. With unlimited time you could even reproduce the output of an LLM manually as it is just a lot of mathematics. Including reasoning, as that is mostly adding words in the context that will steer the word predication to include reasoning. As this is a useful LLM behavior this context is now part of the training data, so it becomes part of the neural network weights.
Yes, but we can't inspect, reproduce or explain the emergent property independently. We can't pick out the "math reasoning" part or the "programming" part, or inspect how it's working, or selectively change any of it. You can't turn any dials or twiddle any knobs. You can't replace one part with another, or pick out components. You can't peek inside and say:
"hey it's got an irrational preference for naming its variables after famous viking warriors, lets change that!"
But worse, it's not that you can't change it, you just don't know! All you can do is test it and guess its biases.
Is it racist, is it homophobic, is it misogynistic? There was an article here the other day about AI in recruitment and the hidden biases. And there was a recruitment AI that only picked men for a role. The job spec was entirely gender neutral. And they hadn't noticed until a researcher looked at it.
It's a black box. So if it does something incorrectly, all they can do is retrain and hope.
Again, this is my present understanding of how it all works right now.
How is it possible that nobody knows how it works - it’s running on hardware we have complete control over and perfect observability into, is it not? At any frame we can pause, examine the state, then step forward, examine the state, and observe what changes have occurred - we have perfect knowledge of the source code, the compiler, whatever components you prefer to break software down into -
The source code is not the LLM. The LLM is billioms of random floating point numbers that somehow encode everything the model knows and can do.
The ML field has a good understanding of the algorithms that produce these floating point numbers and lots of techniques that seem to produce “better” numbers in experiments. However, there is little to no understanding of what the numbers represent or how they do the things they do.
We know how each of the "parts" work, but there is a gazillion of parts (especially since you need to take the model weights into account, which are way larger in size than the code that generates them or uses them to generate stuff), and we found out that together they do something that we do not really understand why they do it.
And inspecting each part is not enough to understand how, together, they achieve what they achieve. We would need to understand the entire system in a much more abstract way, and currently we have nothing more than ideas of how it _might_ work.
Normally, with software, we do not have this problem, as we start on the abstract level with a fully understood design and construct the concrete parts thereafter. Obviously we have a much better understanding of how the entire system of concrete parts works together to perform some complex task.
With AI, we took the other way: concrete parts were assembled with vague ideas on the abstract level of how they might do some cool stuff when put together. From there it was basically trial-and-error, iteration to the current state, but always with nothing more than vague ideas of how all of the parts work together on the abstract level. And even if we just stopped the development now and tried to gain a full, thorough understanding of the abstract level of a current LLM, we would fail, as they already reached a complexity that no human can understand anymore, even when devoting their entire lifetime to it.
However, while this is a clear difference to most other software (though one has to get careful when it comes to the biggest projects like Chromium, Windows, Linux, ... since even though these were constructed abstract-first, they have been in development for such a long time and have gained so many moving parts in the meantime that someone trying to understand them fully on the abstract level will probably start to face the difficulty of limited lifetime as well), it is not an uncommon thing per se: we also do not "really" understand how economy works, how money works, how capitalism works. Very much like with LLMs, humanity has somehow developed these systems through interaction of billions of humans over a long time, there was never an architect designing them on an abstract level from scratch, and they have shown emergent capabilities and behaviors that we don't fully understand. Still, we obviously try to use them to our advantage every day, and nobody would say that modern economies are useless or should be abandoned because they're not fully understood.
If I need to manage an AI as I would manage an employee's brain, I'm going to need quite a few non-technical resources to actually achieve that: time, willpower to babysit it, ability to motivate it, leverage in the form of incentives (and reprimands), to name a few.
AI sits at a weird place where it can't be analyzed as software, and it can't be managed as a person.
My current mental model is that AGI can only be achieved when a machine experiences pleasure, pain, and "bodily functions". Otherwise there's no way to manage it.
> One popular dataset, FineWeb, is about 11.25 trillion words long3, which, if you were reading at about 250 words per minute, would take you over 85 thousand years to read. It’s just not possible for any single human (or even a team of humans) to have read everything that an LLM has read during training.
Do you have to read everything in a dataset with your own eyes to make sense of it? This would make any attempt to address bias in the dataset impossible, and I think it's not, so there should be other ways to make sense of the dataset distribution without having to read it yourself.
It would help if this piece was clearer about the context in which "AI bugs" reveal themselves. As an argument for why you shouldn't have LLMs making unsupervised real-time critical decisions, these points are all well taken. AI shouldn't be controlling the traffic lights in your town. We may never reach a point where it can. But among technologists, the major front on which these kinds of bugs are discussed is coding agents, and almost none of these points apply directly to coding agents: agent coding is (or should be) a supervised process.
Fortunately, we can have LLMs write code and keep all the benefits of normal software (determinism, reproducibility, permanent bug fixes etc.)
I don’t think anyone is advocating for web apps to take the form of an LLM prompt with the app getting created on the fly every time someone goes to the url.
Do you know personally some CEO-s? I know a couple and they generally seem less empathic than the general population, so I don't think that like/dislike even applies.
On the other hand, trying to do something "new" is lots of headaches, so emotions are not always a plus. I could make a parallel to doctors: you don't want a doctor to start crying in a middle of an operation because he feels bad for you, but you can't let doctors doing everything that they want - there needs to be some checks on them.
I would say that the parallel is not at all accurate because the relationship between a doctor and a patient undergoing surgery is not the same as the one you and I have with CEOs. And a lot of good doctors have emotions and they use them to influence patient outcomes positively.
Labor competes for compensation, CEOs compete for status (above a certain enterprise size, admittedly). Show me a CEO willingly stepping down to be replaced by generative AI. Jamie Dimon will be so bold to say AI will bring about a 3 day week (because it grabs headlines [1]) but he isn't going to give up the status of running JPMC; it's all he has besides the wealth, which does not appear to be enough. The feeling of importance and exceptionalism is baked into the identity.
Spoiler there’s no reason we couldn’t work three days a week now. And 100 might be pushing it, but having life expectancy to 90 as well within our grass today as well. We have just decided not to do that.
Almost everyone is "labor" to some extent. There is always a huge customer or major investor that you are beholden to. If you are independently wealthy then you are the exception.
To me, the greatest threat is information pollution. Primary sources will be diluted so heavily in an ocean of generated trash that you might as well not even bother to look through any of it.
I see that as the death knell for general search engines built to indiscriminately index the entire web. But where that sort of search fails, opportunities open up for focused search and curated search.
Just as human navigators can find the smallest islands out in the open ocean, human curators can find the best information sources without getting overwhelmed by generated trash. Of course, fully manual curation is always going to struggle to deal with the volumes of information out there. However, I think there is a middle ground for assisted or augmented curation which exploits the idea that a high quality site tends to link to other high quality sites.
One thing I'd love is to be able to easily search all the sites in a folder full of bookmarks I've made. I've looked into it and it's a pretty dire situation. I'm not interested in uploading my bookmarks to a service. Why can't my own computer crawl those sites and index them for me? It's not exactly a huge list.
And it imitates all the unimportant bits perfectly (like spelling, grammar, word choice) while failing at the hard to verify important bits (truth, consistency, novelty)
It’s already been happening but now it’s accelerated beyond belief. I saw a video about how WW1 reenactment photos end up getting reposted away from their original context and confused with original photos to the point it’s impossible to tell unless you can track it back to the source.
Now most of the photos online are just AI generated.
Concentrated power is kinda a pre-requisite for anything bad happening, so yes, it's more likely in exactly the same way that given this:
Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations.
"Linda is a bank teller" is strictly more likely than "Linda is a bank teller and is active in the feminist movement" — all you have is P(a)>P(a&b), not what the probability of either statement is.
Why does an AI need the ability to "dislike" to calculate that its goals are best accomplished without any living humans around to interfere? Superintelligence doesn't need emotions or consciousness to be dangerous.
I hear that. Then I try to use AI for simple code task, writing unit tests for a class, very similar to other unit tests. If fails miserably. Forgets to add an annotation and enters in a death loop of bullshit code generation. Generates test classes that tests failed test classes that test failed test classes and so on. Fascinating to watch. I wonder how much CO2 it generated while frying some Nvidia GPU in an overpriced data center.
AI singularity may happen, but the Mother Brain will be a complete moron anyway.
Regularly trying to use LLMs to debug coding issues has convinced me that we're _nowhere_ close to the kind of AGI some are imagining is right around the corner.
Sure, but also the METR study showed the rate of change is t doubles every 7 months where t ~= «duration of human time needed to complete a task, such that SOTA AI can complete same with 50% success»: https://arxiv.org/pdf/2503.14499
I don't know how long that exponential will continue for, and I have my suspicions that it stops before week-long tasks, but that's the trend-line we're on.
At least Mother Brain will praise your prompt to generate yet another image in the style of Studio Ghibli as proof that your mind is a tour de force in creativity, and only a borderline genius would ask for such a thing.
Most reasonable AI alarmists are not concerned with sentient AI but an AI attached to the nukes that gets into one of those repeating death loops and fires all the missiles.
You can say that, and I might even agree, but many smart people disagree. Could you explain why you believe that? Have you read in detail the arguments of people who disagree with you?
Our best technology at current require teams of people to operate and entire legions to maintain. This leads to a sort of balance, one single person can never go too far down any path on their own unless they convince others to join/follow them. That doesn't make this a perfect guard, we've seen it go horribly wrong in the past, but, at least in theory, this provides a dampening factor. It requires a relatively large group to go far along any path, towards good or evil.
AI reduces this. How greatly it reduces this, if it reduces it to only a handful, to a single person, or even to 0 people (putting itself in charge), seems to not change the danger of this reduction.
I'm not so sure it will be that either, it would be having multiple AIs essentially at war with each other over access to GPUs/energy or whatever the materials are needed to grow if/when that happens. We will end up as pawns in this conflict.
Given that even fairly mediocre human intelligences can run countries into the ground and avoid being thrown out in the process, it's certainly possible for an AI to be in the intelligence range where it's smart enough to win vs humans but also dumb enough to turn us into pawns rather just go to space and blot out the sun with a Dyson swarm made from the planet Mercury.
But don't count on it.
I mean, apart from anything else, that's still a bad outcome.
IIRC the original idea was that the machines used our brain capacity as a distributed array but then they decided batteries was easier to understand while been sillier, just burn the carbon they are feeding us, it’s more efficient.
If I could write the matrix reverted, Neo would discover that the last people put themselves in the pods because the world was so fucked up, and the machines had been caretakers that were trying to protect them from themselves. That revision would make the first movie perfect.
It's great to help people understand that AI can be both surprisingly good and disappointing, and that testing is the only way to know, but it's impossible to test everything. That sets expectations.
I think that means savvy customers will want details or control over testing, and savvy providers will focus on solutions they can validate, or where testing is included in the workflow (e.g., code), or where precision doesn't matter (text and meme generation). Knowing that in depth is gold for AI advocates.
Otherwise, I don't think people really know or care about bugs or specifications or how AI breaks prior programmer models.
But people will become very hostile and demand regulatory frenzies if AI screws things up (e.g., influencing elections or putting people out of work). Then no amount of sympathy or understanding will help the industry, which has steadily been growing its capability for evading regulation via liability disclaimers, statutory exceptions, arbitration clauses, pitting local/regional/national governments against each other, etc.
To me that's the biggest risk: we won't get the benefits and generational investments will be lost in cleaning up after a few (even accidental) bad actors at scale.
I think your post is fundamentally wrong. See, you are comparing AI responses with the written code, which may not be the fair comparison. I see it as, better you could compare code generated by AI vs the code written by an engineer.
The original author seems to view the AI application as itself a software application which has desired or undesired, and predictable or unpredictable, behaviors. That doesn't seem like an invalid thing to talk about merely because there are other software-related conversations we can have about AIs (or other code-quality-related conversations).
I don't understand the "your boss" framing of this article, or more accurately, the title of this article. The article contents don't actually seem to have anything to do with management specifically. Is the reader is meant to believe that not being scared of AI is a characteristic of the managerial class? Is the unstated implication that there is some class warfare angle and anybody who isn't against AI is against laborers? Because what the article actually overtly argues, without any reading between the lines, is quite mundane.
> Is the unstated implication that there is some class warfare angle and anybody who isn't against AI is against laborers?
I didn't read it that way. I read "your boss" as basically meaning any non-technical person who may not understand the challenges of harnessing LLMs compared to traditional, (more) deterministic software development.
Indeed, from reading the article I could really see any discussion of "your boss", so I changed the title to something more representative, and a condensed version of a phrase from the article.
Apple’s underwhelming LLM rollout—like the pulled notification summaries and trivial emoji tools—proves even big tech struggles to turn AI hype into reliable, daily-useful features; I’d take a working email organizer over a glitchy "smart" summary any day.
> It’s entirely possible that some dangerous capability is hidden in ChatGPT, but nobody’s figured out the right prompt just yet.
This sounds a little dramatic. The capabilities of ChatGPT are known. It generates text and images. The qualities of the content of the generated text and images is not fully known.
Think of the news about the kid who got recommended to suicide by ChatGPT, or chatgpt providing the user information on how to do illegal activities, these capabilities are the ones that the author it's referring to
And that sounds a little reductive. There's a lot that can be done with text and images. Some of the most influential people and organizations in the world wield their power with text and images.
Yeah, and to riff off the headline, if something dangerous is connected to and taking commands from ChatGPT then you better make sure there’s a way to turn it off.
Plus there is the 'monkeys with typewriters' problem with both danger and hypothetical good. In contrast, ChatGPT may technically reply to the right prompt with a universal cancer cure/vaccine. Psuedorandomly generating it wouldn't help as you wouldn't recognize it from all of the other queries of things we don't know of as true or false.
Likewise what to ask it for how to make some sort of horrific toxic chemical, nuclear bomb, or similar isn't much good if you cannot recognize it and dangerous capability depends heavily on what you have available to you. Any idiot can be dangerous with C4 and detonator or bleach and ammonia. Even if ChatGPT could give entirely accurate instructions on how to build an atomic bomb it wouldn't do much good because you wouldn't be able to source the tools and materials without setting off red flags.
"The worst effects of this flaw are reserved for those who create what is known as the “lethal trifecta”. If a company, eager to offer a powerful AI assistant to its employees, gives an LLM access to un-trusted data, the ability to read valuable secrets and the ability to communicate with the outside world at the same time, then trouble is sure to follow. And avoiding this is not just a matter for AI engineers. Ordinary users, too, need to learn how to use AI safely, because installing the wrong combination of apps can generate the trifecta accidentally."
Not the point, but I’m confused by the Geoguessr screenshot. Under the reasoning for its decision, it mentions “traffic keeps to the left” but that is not apparent from the photo.
Then it says the shop sign looks like a “Latin alphabet business name rather than Spanish or Portuguese”. Uhhh… what? Spanish and Portuguese use the Latin alphabet.
> The answer is 24! See the ASCII values of '1' is 49, '2' is 50, and '+' is 43. Adding all that together we get 3. Now since we are doing this on a computer with a 8-bit infrastructure we multiply by 3 and so the answer is 24.
Cool! I didn't understand any of that but it was correct and you sound smart. I will put this thing in charge of critical parts of my business.
Tremendous alpha right now in making scary posts about AI. Fear drives clicks. You don't even need to point to current problems, all you have to do is say we can't be sure they won't happen in the future.
All the same criticisms are true about hiring humans. You don’t really know what they’re thinking, you don’t really know what their values and morals are, you can’t trust that they’ll never make a mistake, etc.
I think you're misreading the article; the point here is not "LLMs are bad and can't replace humans," the point is that many non-technical people have the expectation that LLMs can replace humans _but still behave like regular software_ with regard to reliability and operability.
When a CEO sees their customer chatbot call a customer a slur, they don't see "oh my chatbot runs on a stochastic model of human language and OpenAI can't guarantee that it will behave in an acceptable way 100% of the time", they see "ChatGPT called my customer a slur, why did you program it to do that?"
But this is why using the AI in the production of (almost) deterministic systems makes so much sense, including saving on execution costs.
ISTR someone else round here observing how much more effective it is to ask these things to write short scripts that perform a task than doing the task themselves, and this is my experience as well.
If/when AI actually gets much better it will be the boss that has the problem. This is one of the things that baffles me about the managerial globalists - they don't seem to appreciate that a suitably advanced AI will point the finger at them for inefficiency much more so than at the plebs, for which it will have a use for quite a while.
It's no different from those on HN that yell loudly that unions for programmers are the worst idea ever... "it will never be me" is all they can think, then they are protesting in the streets when it is them, but only after the hypocrisy of mocking those in the street protesting today.
Agreed. My dad was raised strongly fundamentalist, and in North America, that included (back then) strongly resisting unions. In hindsight, I've come to realize that my parent's weren't maybe even of average intelligence, and definitely of above-average gullibility.
Unionized software engineers would solve a lot of the "we always work 80 hour weeks for 2 months at the end of a release cycle" problems, the "you're too old, you're fired" issues, the "new hires seems to always make more than the 5/10+ year veterans", etc. Sure, you wouldn't have a few getting super rich, but it would also make it a lot easier for "unionized" action against companies like Meta, Google, Oracle, etc. Right now, the employers hold like 100x the power of the employees in tech. Just look at how much any kind of resistance to fascism has dwindled after FAANG had another round of layoffs..
Software "engineers" totally miss a key thing in other engineering professions as well, which is organizations to enforce some pretense of ethical standards to help push back against requests from product. Those orgs often look a lot like unions.
This article makes a solid case. The worst kinds of bugs in software are not the most obvious ones like syntax errors, they are the ones where the code appears to be working correctly, until some users do something slightly unusual after a few weeks of some code change being deployed and it breaks spectacularly but the bug only affects a small fraction of users so developers cannot reproduce the issue... And the cose change happened such time ago that the guilty code isn't even suspected.
> here are some example ideas that are perfectly true when applied to regular software
Hm, I'm listening, let's see.
> Software vulnerabilities are caused by mistakes in the code
That's not exactly true. In regular software, the code can be fine and you can still end up with vulnerabilities. The platform in which the code is deployed could be vulnerable, or the way it is installed make it vulnerable, and so on.
> Bugs in the code can be found by carefully analysing the code
Once again, not exactly true. Have you ever tried understanding concurrent code just by reading it? Some bugs in regular software hide in places that human minds cannot probe.
> Once a bug is fixed, it won’t come back again
Ok, I'm starting to feel this is a troll post. This guy can't be serious.
> If you give specifications beforehand, you can get software that meets those specifications
You should read the footnote marked [1] after "a note for technical folk" at the beginning of the article. He is very consciously making sweeping generalizations about how software works in order to make things intelligible to non-technical readers.
> I’m also going to be making some sweeping statements about “how software works”, these claims mostly hold, but they break down when applied to distributed systems, parallel code, or complex interactions between software systems and human processes.
I'd argue that this describes most software written since, uh, I hesitate to even commit to a decade here.
For the purposes of the article, which is to demonstrate how developing an LLM is completely different from developing traditional software, I'd say they are true enough. It's a CS 101 understanding of the software development lifecycle, which for non-technical readers is enough to get the point across. An accurate depiction of software development would only obscure the actual point for the lay reader.
At least the 1950’s. That’s when stuff like asynchrony and interrupts were worked out. Dijkstra wrote at length about this in reference to writing code that could drive a teletype (which had fundamentally non-deterministic timings).
If you include analog computers, then there are some WWII targeting computers that definitely qualify (e.g., on aircraft carriers).
> these claims mostly hold, but they break down when applied to distributed systems, parallel code, or complex interactions between software systems and human processes
The claims the GP quoted DON’T mostly hold, they’re just plain wrong. At least the last two, anyway.
He is trying to lax the general public perception around AIs shortcomings. He's giving AI a break, at the expense of regular developers.
This is wrong on two fronts:
First, because many people foresaw the AI shortcomings and warned about them. This "we can't fix a bug like in regular software" theatre hides the fact that we can design better benchmarks, or accountability frameworks. Again, lots of people foresaw this, and they were ignored.
Second, because it puts the strain on non-AI developers. It blamishes all the industry, putting together AI with non-AI in the same bucket, as if AI companies stumbled on this new thing and were not prepared for its problems, when the reality is that many people were anxious about the AI companies practices not being up to standard.
I think it's a disgraceful take, that only serves to sweep things under a carpet.
I don't think he's doing that at all. The article is pointing out to non-technical people how AI is different than traditional software. I'm not sure how you think it's giving AI a break, as it's pointing out that it is essentially impossible to reason about. And it's not at the expense of regular developers because it's showing how regular software development is different than this. It makes two buckets, and puts AI in one and non-AI in the other.
He is. Maybe he's just running with the pack, but that doesn't matter either.
The fact is, we kind of know how to prevent problems in AI systems:
- Good benchmarks. People said several times that LLMs display erratic behavior that could be prevented. Instead of adjusting the benchmarks (which would slow down development), they ignored the issues.
- Accountability frameworks. Who is responsible when an AI fails? How the company responsible for the model is going to make up for it? That was a demand from the very beginning. There are no such accountability systems in place. It's a clown fiesta.
- Slowing down. If you have a buggy product, you don't scale it. First, you try to understand the problem. This was the opposite of what happened, and at the time, they lied that scaling would solve the issues (when in fact many people knew for a fact that scaling wouldn't solve shit).
Yes, it's kind of different. But it's a different we already know. Stop pushing this idea that this stuff is completely new.
'we' is the operative word here. 'We', meaning technical people who have followed this stuff for years. The target audience of this article are not part of this 'we' and this stuff IS completely new _for them_. The target audience are people who, when confronted with a problem with an LLM, think it is perfectly reasonable to just tell someone to 'look at the code' and 'fix the bug'. You are not the target audience and you are arguing something entirely different.
Let's pretend I'm the audience, and imagine that in the past I said those things ("fix the bug" and "look at the code").
What should I say now? "AI works in mysterious ways"? Doesn't sound very useful.
Also, should I start parroting innacurate outdated generalizations about regular software?
The post doesn't teach anything useful for a beginner audience. It's bamboozling them. I am amazed that you used the audience perspective as a defense of some kind. It only made it worse.
Please, please, take a moment to digest my critique properly. Think about what you just said and what that implies. Re-read the thread if needed.
Where did "can't you just turn it off?" in the title come from? It doesn't appear anywhere in the actual title or the article, and I don't think it really aligns with its main assertions.
It shows up at https://boydkane.com under the link "Why your boss isn't worried about advanced AI". Must be some kind of sub-heading, but not part of the actual article / blog post.
Presumably it's a phrase you might hear from a boss who sees AI as similar to (and as benign/known/deterministic as) most other software, per TFA
In my experience it’s usually the engineers that aren’t worried about AI, because they see the limitations clearly every time they use it. It’s pretty obvious that whole thing is severely overhyped and unreliable.
Your boss (or more likely, your bosses’ bosses’s boss) is the one deeply worried about it. Though mostly worried about being left behind by their competitors and how their company’s use of AI (or lack thereof) looks to shareholders.
It depends on where you are in the chain, and what kind of engineering you’re doing. I think a lot of engineers are so focused on the logistics, capabilities, and flaws, and so used to being indispensable, that they don’t viscerally get that they’re standing on the wrong side of the tree branch they’re sawing through. AI does not need to replace a single engineer before increased productivity means we’ll have way too many engineers, which mean jobs are impossible to get, and the salaries are in the shitter. Middle managers are terrified because they know they’re not long for this (career) world. Upper managers are having 3 champagne lunches because they see big bonuses on the far side of skyrocketing profits and cratering payroll costs.
This sci-fi thing goes as far back as the 1983 movie WarGames, where they wanted to pull the plug on a rogue computer, but there was a reason you couldn’t do that:
McKittrick: General, the machine has locked us out. It's sending random numbers to the silos.
Pat Healy: Codes. To launch the missiles.
General Beringer: Just unplug the goddamn thing! Jesus Christ!
McKittrick: That won't work, General. It would interpret a shutdown as the destruction of NORAD. The computers in the silos would carry out their last instructions. They'd launch.
Further than that, even - this trope appears in Colossus: The Forbin Project, released in 1970, where the rogue computer is buried underground with its own nuclear reactor, so it can't be powered off.
In real life it won’t be that the computer prevents you from turning it off. It’ll be that the computer is guarded by cultists who think its god, and unstoppable market forces that require it to keep running.
When AI ends up running everything essential to survival and society, it’ll be preposterous to even suggest pulling the plug just because it does something bad.
Can you imagine the chaos of completely turning off GPS or Gmail today? Now imagine pulling the plug on something in the near future that controls all electric power distribution, banking communications, and Internet routing.
This is the case with capitalism today. I don't like where he took the philosophy, but Nick Land did have an insight that all the worst things we believe about AI (e.g. paperclip optimizing etc) are capitalism in a nutshell.
Just listen to what these CEOs say on the topic and they basically admit something terrible is being built, but that the most important things is that they are the ones to do it first.
It's a poor choice of phrase if the purpose is to illustrate a false equivalence. It applies to AI both as much (you can kill a process or stop a machine just the same regardless of whether it's running an LLM) and as little (you can't "turn off" Facebook any more than you can "turn off" ChatGPT) as it does to any other kind of software.
Or a bad driver crashing everything multiple times a week.
Or a misbehaving process not handing control back to the OS.
I grew up in the era of 8 and 16 bit micros and early PCs, they where hilariously less stable than modern machines while doing far less, there wasn’t some halcyon age of near perfect software, it’s always been a case of things been good enough to be good enough but at least operating systems did improve.
I guess that is because you run it on old hardware. When I've bought my Asus ROG expensive laptop I had bsod almost daily. A year later with all updates I had bsod once in a month on the same device and windows installation.
If you have faulty hardware no amount of software is going to solve your problems (other than software that just completely deactivates said faulty hardware).
The fact you continued to have BSOD issues after a full reinstall is pretty strong evidence you probably had some kind of hardware failure.
Mostly because Microsoft shut down kernel access, wrote it's own generic drivers for "simple" devices (USBs, printers, sound cards, ...) and made "heavy" drivers submit to their WHQL quality control to be signed to run.
I remember Linux being remarkable reliable throughout its entire life in spite of being rabidly worked on.
Windows is only stabilizing because it's basically dead. All the activity is in the higher layers, where they are racking their brains on how to enshittify the experience, and extract value out of the remaining users.
I grew up in the same era and I recall crashes being less frequent.
There were plenty of other issues, including the fact that you had to adjust the right IRQ and DMA for your Sound Blaster manually, both physically and in each game, or that you needed to "optimize" memory usage, enable XMS or EMS or whatever it was at the time, or that you spent hours looking at the nice defrag/diskopt playing with your files, etc.
More generally, as you hint to, desktop operating systems were crap, but the software on top of it was much more comprehensively debugged. This was presumably a combination of two factors: you couldn't ship patches, so you had a strong incentive to debug it if you wanted to sell it, and software had way fewer features.
Come to think about it, early browsers kept crashing and taking down the entire OS, so maybe I'm looking at it with rosy glasses.
Last year I assembled a retro PC (Pentium 2, Riva TNT 2 Ultra, Sound Blaster AWE64 Gold) running Windows 98 to remember my childhood, and it is more stable than what I remembered, but still way worse than modern systems. There are plenty of games that will refuse to work for whatever reason, or that will crash the whole OS, specially when existing, and require a hard reboot.
Oh and at least in the '90s you could already ship patches, we used to get them with the floppies and later CDs provided by magazines.
It truly depends on the quality of the software you were using at the time. Maybe the software you used didn't result in many issues. I know a lot of the games I played as a kid on my family's or friend's Win95 machines resulted in system lockups or blue screens practically every time we used them.
As I mess around with these old machines for fun in my free time, I encounter these kinds of crashes pretty dang often. Its hard to tell if its just the old hardware is broken in odd ways or not so I can't fully say its the old software, but things are definitely pretty unreliable on old desktop Windows running old desktop Windows apps.
I think old in this sense is "released" rather than "beta" - it takes time to make any software reliable. Many of the examples here further prove that young software is unreliable.
Neither, you’re reading it wrong. Think of it as codebases getting more reliable over time as they accumulate fixes and tests. (As opposed to, say, writing code in NodeJS versus C++)
Age of Code does not automatically equal quality of code, ever. Good code is maintained by good developers. A lot of bad code is pushed out by management, and other situations, or just bad devs. This is a can of worms you're talking your way into.
You're using different words - the top comment only mentioned the reliability of the software, which is only tangentially related to the quality, goodness, or badness of the code used to write it.
Old software is typically more reliable, not because the developers were better or the software engineering targeted a higher reliability metric, but because it's been tested in the real world for years. Even more so if you consider a known bug to be "reliable" behavior: "Sure, it crashes when you enter an apostrophe in the name field, but everyone knows that, there's a sticky note taped to the receptionist's monitor so the new girl doesn't forget."
Maybe the new software has a more comprehensive automated testing framework - maybe it simply has tests, where the old software had none - but regardless of how accurate you make your mock objects, decades of end-to-end testing in the real world is hard to replace.
As an industrial controls engineer, when I walk up to a machine that's 30 years old but isn't working anymore, I'm looking for failed mechanical components. Some switch is worn out, a cable got crushed, a bearing is failing...it's not the code's fault. It's not even the CMOS battery failing and dropping memory this time, because we've had that problem 4 times already, we recognize it and have a procedure to prevent it happening again. The code didn't change spontaneously, it's solved the business problem for decades... Conversely, when I walk up to a newly commissioned machine that's only been on the floor for a month, the problem is probably something that hasn't ever been tried before and was missed in the test procedure.
Yup, I have worked on several legacy codebases, and a pretty common occurence is that a new team member will join and think they may have discovered a bug in the code. Sometimes they are even quite adamant that the code is complete garbage and could never have worked properly. Usually the conversation goes something like: "This code is heavily used in production, and hasn't been touched in 10 years. If it's broken, then why haven't we had any complaints from users?"
And more often than not the issue is a local configuration issue, bad test data, a misunderstanding of what the code is supposed to do, not being aware of some alternate execution path or other pre/post processing that is running, some known issue that we've decided not to fix for some reason, etc. (And of course sometimes we do actually discover a completely new bug, but it's rare).
To be clear, there are certainly code quality issues present that make modifications to the code costly and risky. But the code itself is quite reliable, as most bugs have been found and fixed over the years. And a lot of the messy bits in the code are actually important usability enhancements that get bolted on after the fact in response to real-world user feedback.
Old code that has been maintained (bugfixed), but not messed with too much (i.e. major rewrites or new features) is almost certain to be better than most other code though?
"Bugfixes" doesn't mean the code actually got better, it just means someone attempted to fix a bug. I've seen plenty of people make code worse and more buggy by trying to fix a bug, and also plenty of old "maintained" code that still has tons of bugs because it started from the wrong foundation and everyone kept bolting on fixes around the bad part.
One of frustrating truths about software is that it can be terrible and riddled with bugs but if you just keep patching enough bugs and use it the same way every time it eventually becomes reliable software ... as long as the user never does anything new and no-one pokes the source with a stick.
I much prefer the alternative where it's written in a manner where you can almost prove it's bug free by comprehensively unit testing the parts.
I think we all agree that the quality of the code itself goes down over time. I think the point that is being made is that the quality of the final product goes up over time.
E.g. you might fix a bug by adding a hacky workaround in the code; better product, worse code.
It actually might. Older code running in production is almost automatically regression tested with each new fix. It might not be pretty, but it's definitely more reliable for solving real problems.
The list of bugs tagged regression at work certainly suggests it gets tested... But fixing those regressions...? That's a lot of dev time for things that don't really have time allocated for them.
The author didn't mean that an older commit date on a file makes code better.
The author is talking about the maturity of a project. Likewise, as AI technologies become more mature we will have more tools to use them in a safer and more reliable way.
I've seen too many old projects that are not by any means better no matter how much they get updates because management define priorities. I'm not alone in saying I've been in a few projects where the backlog is rather large. When your development is driven by marketing people trying to pump up sales, all the "non critical" bugs begin to stack up.
Survivorship bias is real, but is missing the important piece of the story when it comes to software, which doesn't just survive but is also maintained. Sure you may choose to discard/replace low quality software and keep high quality software in operation, which leads to survivorship bias, but the point here is that you also have a chance to find and fix issues in the one that survived, even if those issues weren't yet apparent in version 0.1. Author is not trying to say that version 0.1 of 30 year old software was of higher quality than version 0.1 of modern software -- they're saying that version 9 of 30 year old software is better than version 0.1 of modern software.
In my experience actively maintained but not heavily modified applications tend towards stability over time. It don't even matter if they are good or bad codebases -- even a bad code will become less buggy over time if someone is working on bug fixes.
New code is the source of new bugs. Whether that's an entirely new product, a new feature on an existing project, or refactoring.
70 years ago we were fascinated by the concept of converting analog to a perfect digital copy. In reality, that goal was a pipe drea!m and the closest we can ever get is a near identical facimile to which data fits... But it's still quite easy to determine digital from true analog with rudimentary means.
Human thought is analog. It is based on chemical reactions, time, and unpredictably (effectively) random physical characteristics. AI is an attempt to turn that which is purely digital into an rational analog thought equivalent.
No matter how much effort, money, power, and rare mineral eating TPUs will - ever - produce true analog data.
It's been closer to 100 years since we figured out information theory and discredited this idea (that continuous/analog processes have more, or different, information in them than discrete/digital ones)
This is all true. But digital audio and video media has captured essentially all economic value outside of live performance. So it seems likely that we will find a "good enough" in this domain too.
Interesting point with economic value extraction. The economy sacrificed accuracy and warmth of analog storage for convenience and security of digital storage. With economic incentive I am sure society will sacrifice accuracy and precision for the convenience of AI
For a real world example of the challenges of harnessing LLMs, look at Apple. Over a year ago they had a big product launch focused on "Apple Intelligence" that was supposed to make heavy use of LLMs for agentic workflows. But all we've really gotten since then are a couple of minor tools for making emojis, summarizing notifications, and proof reading. And they even had to roll back the notification summaries for a while for being wildly "out of control". [1] And in this year's iPhone launch the AI marketing was toned down significantly.
I think Apple execs genuinely underestimated how difficult it would be to get LLMs to perform up to Apple's typical standards of polish and control.
[1] https://www.bbc.com/news/articles/cge93de21n0o
> to perform up to Apple's typical standards of polish and control.
i no longer believe they have kept on to the standards in general. the ux/ui used to be a top priority, but the quality control has certainly gone down over the years [1]. the company is now driven by supply chain and business-minded optimizations than what to give to the end user.
at the same time, what one can do using AI has large correlation with what one does with their devices in the first place. a windows recall like feature for ipad os might have been interesting (if not equally controversial), but not that useful because even till this day it remains quite restrictive for most atypical tasks.
[1] https://www.macobserver.com/news/macos-tahoe-upside-down-ui-...
>> to perform up to Apple's typical standards of polish and control.
>i no longer believe they have kept on to the standards in general.
One 100% agree with this, if I compare AI's ability to speed up the baseline for me in terms of programming Golang (hard/tricky tasks clearly still require human input - watch out for I/O ops) with Apple's lack of ability to integrate it in even the simplest of ways.. things are just peculiar on the Apple front. Bit similar to how MS seems to be gradually loosing the ability to produce a version of Windows that people want to run due to organisational infighting.
Personally, I’ve never seen an AI flow of any kind that meets what would meet the quality of a typical ‘corporate’ acceptable flow. As in, reliably works, doesn’t go crazy randomly, etc.
I’ve seen a lot of things that look like they’re working for a demo, but shortly after starting to use it? Trash. Not every time (and it’s getting a little better), but often enough that personally I’ve found them a net drain on productivity.
And I literally work in this space.
Personally, I find apples hesitation here a breath of fresh air, because I’ve come to absolutely hate Windows - and everybody doing vibe code messes that end up being my problem.
>i no longer believe they have kept on to the standards in general.
They're definitely not as good as they WERE, but they're still better than anybody else.
With Apple it's incredibly obvious that most software product development is nowadays handled by outsourced/offshored contractors who simply do not use the products. At least I hope that's the case, it would be disastrous if the state of iOS/watchOS is the result of their in-house on-shore talent.
> minor tools for making emojis, summarizing notifications, and proof reading.
The notification / email summaries are so unbelievably useless too: it’s hardly more work to skim the notification / email that I do anyway.
Like most AI products it feels like they started with a solution first and went searching for the problems. Text messages being too long wasn't a real problem to begin with.
There are some good parts to Apple Intelligence though. I find the priority notifications feature works pretty well, and the photo cleanup tool works pretty well for small things like removing your finger from the corner of a photo, though it's not going to work on huge tasks like removing a whole person from a photo.
> it's not going to work on huge tasks like removing a whole person from a photo.
I use it for removing people who wander into the frame quite often. It probably wont work for someone close up, but its great for removing a tourist who spends ten minutes taking selfies in front of a monument.
Honestly I love the priority notifications and the notification summaries. The thing that drives me absolutely insane, is that the fact that when I view the notification through clicking on it from another space other than the "While in the reduce interruptions focus" it doesn't clear. Because of this, I always have infinite notifications.
I want to open WhatsApp and open the message and have it clear the notif. Or atleast click the notif from the normal notif center and have it clear there. It kills me
I mean it happened quite a few times that phishing emails became the priority notification on my phone
Do you know if apple is using their new tools to do mail filtering? It's an interesting choice if they are since it's a genuine problem with a mature (but always evolving) solution.
The Ring app notification summaries still scare me.
> "A bunch of people right outside your house!!!"
because it aggregates multiple single person walking by notifications that way...
That is a fantastic example of blind application of AI making things worse.
the advertising of those spy doorbells is entirely based on paranoia
so ramping it up the rhetoric doesn't really hurt them...
Unrelated, but am I the only person who finds the concept of “getting notifications for somebody walking by a house” to be really creepy?
Well yeah, but that's in part a problem with always-on doorbell cameras. On paper they're illegal in many countries (privacy laws, you can't just put up a camera and record anyone out in public), in practice the police asks people to put their doorbell cameras in a registry so they can request footage if needs be.
Anyway, I get wanting to see who's ringing your doorbell in e.g. apartment buildings, and that extending to a house, especially if you have a bigger one. But is there a reason those cameras need to be on all the time?
At least in the USA it’s legal to record public spaces. So recording the street and things that can be seen from it is legal, but pointing a camera over your neighbors fence is not.
And a lot of people don't share that opinion, so this isn't the law in a lot of countries. When you wanted to suggest that it is a problem, that US companies try to extend the law of there home country to other parts of the world, then I endorse that.
it isn't creepy, it's super annoying if you don't live in the woods. got a ring doorbell and turned them off a few hours after installation, it was driving me nuts.
That makes... That makes just enough sense to become nonsense, rather than mere noise.
I mean, I could imagine a person with no common sense almost making the same mistake: "I have a list of 5 notifications of a person standing on the porch, and no notifications about leaving, so there must be a 5 person group still standing outside right now. Whadya mean, 'look at the times'?"
> A biologist, a physicist and a mathematician were sitting in a street cafe watching the crowd. Across the street they saw a man and a woman entering a building. Ten minutes they reappeared together with a third person.
> - They have multiplied, said the biologist.
> - Oh no, an error in measurement, the physicist sighed.
> - If exactly one person enters the building now, it will be empty again, the mathematician concluded.
https://www.math.utah.edu/~cherk/mathjokes.html
It does feel like somebody forgot that "from the first sentence or two of the email, you can tell what it's about" was already a rule of good writing...
Maybe they remembered that a lot of people aren't actually good writers. My brother will send 1000 word emails that meander through subjects like what he ate for breakfast to eventually get to the point of scheduling a meeting about negotiating a time for help with moving a sofa. Mind you, I see him several times a week so he's not lonely, this is just the way he writes. Then he complains endlessly about his coworkers using AI to summarize his emails. When told that he needs to change how he writes to cut right to the point, he adopts the "why should I change, they're the ones who suck" mentality.
So while Apple's AI summaries may have been poorly executed, I can certainly understand the appeal and motivation behind such a feature.
I feel too many humanities teachers are like your brother.
Why use 10 words when you could do 1000. Why use headings or lists, when the whole story could be written in a single paragraph spanning 3 pages.
I mean...this depends very heavily on what the purpose of the writing is.
If it's to succinctly communicate key facts, then you write it quickly.
- Discovered that Bilbo's old ring is, in fact, the One Ring of Power.
- Took it on a journey southward to Mordor.
- Experienced a bunch of hardship along the way, and nearly failed at the end, but with Sméagol's contribution, successfully destroyed the Ring and defeated Sauron forever.
....And if it's to tell a story, then you write The Lord of the Rings.
There’s a thread here that could be pulled - something about using AI to turn everyone into exactly who you want to communicate with in the way you want.
Probably a sci-fi story about it, if not, it should be written.
I think people read texts because they want to read them, and when they don't want to read the texts they are also not even interested in reading the summaries.
Why do I think this? ...in the early 2000's my employer had a company wide license for a document summarizer tool that was rather accurate and easy to use, but nobody ever used it.
The obvious use case is “I don’t want to read this but I am required to read this (job)” - the fact that people don’t want to use it even there is telling, imo.
You sometimes need to want to quickly learn what's in an email that was written by someone less helpful.
Eg sometimes the writer is outright antagonistic, because they have some obligation to tell you something, but don't actually want you to know.
Even bending over that far backwards to find a useful example comes up empty.
Those kinds of emails are so uncommon they’re absolutely not worth wasting this level of effort on. And if you’re in a sorry enough situation where that’s not the case, what you really need is the outside context the model doesn’t know. The model doesn’t know your office politics.
And you trust AI to accurately read between the lines?
I think humans are quite well capable of skimming text and reading multiple lines at once.
This is a pretty damning example of backwards product thinking. How often, truly, does this happen?
Never heard of terms of service?
No one cares about the terms of service. And if they actually do, they will need to read every word very carefully to know if they are in legal trouble. A possibly wrong summary of a terms of service document is entirely and completely useless.
It's not even that they are useless, they are actively wrong. I could post pages upon pages of screenshots of the summaries being literally wrong about the content of the messages it summarised.
Maybe they used their AI to design Liquid Glass. Impressive at first sight, but unusable in practice.
All form and no function, or in other words, slop.
Now their strategy is to allow for Apple Events to work with the MCP.
https://9to5mac.com/2025/09/22/macos-tahoe-26-1-beta-1-mcp-i...
My wife was in China recently and was sending back pictures of interesting things - one came in while I was driving and my iPhone read out a description of the picture that had been sent - "How cool is that!" I thought.
However, when I stopped driving and looked at the picture the AI generated description was pretty poor - it wasn't completely wrong but it really wasn't what I was expecting given the description.
It’s been surprisingly accurate at times “a child holding an apple” in a crowded picture, and then sometimes somewhat wrong.
What really kills me is “a screenshot of a social media post” come on it’s simple OCR read the damn post to me you stupid robot! Don’t tell me you can’t, OCR was good enough in the 90s!
The description said "People standing in front of impressive scenery" (or something like that) - it got the scenery part correct but the people are barely visible and really small.
is this a complaint about the wife or the ai?
The Apple AI
Apple's whole brand is built around tight control, predictable behavior, and a super polished UX which is basically the opposite of how LLMs behave out of the box
> I think Apple execs genuinely underestimated how difficult it would be to get LLMs to perform up to Apple's typical standards of polish and control
Not only Apple, this is happening across the industry. Executives' expectations of what AI can deliver are massively inflated by Amodei et al. essentially promising human-level cognition with every release.
The reality is aside from coding assistants and chatbot interfaces (a la chatgpt) we've yet to see AI truly transform polished ecosystems like smartphones and OS' for a reason.
Standard hype cycle. We are probably creating the top of the peak of inflated expectations.
What I don't get is there's some fairly ... easy bits they could do, but have not.
Why not take the easy wins? Like let me change phone settings with Siri or something, but nope.
A lot of AI seems to be mismanaging it into doing things AI (LLMs) suck at... while leaving obvious quick wins on the table.
Apple’s typical standards of “polish and control” seem to be slipping drastically if MacOS Tahoe is anything to go by.
You need to reduce the standard to fit the Apple Intelligence (AI) in. This is also industry best practice.
Which is ironic, given all I really want from Siri is an advanced-voice-chat-level chat gpt experience - being able to carry on about 90% of a natural conversation with gpt, while Siri vacillates wildly between 1) simply not responding 2) misunderstanding and 3) understand but refusing to engage - feels awful.
That tells you AAPL didn't have the staff necessary to make this happen.
Probably the issue is it is free. If people paid for it they could scale infra to cope.
The thought that a company like Apple, which surely put hundreds of engineers to work on these tools and went through multiple iterations of their capabilities, would launch the capabilities...Only for its executives to realize after release that current AI is not mature enough to add significant commercial value to their products, is almost comical.
The reality is that if they hadn’t announced these tools and joined the make-believe AI bubble, their stock price would have crashed. It’s okay to spend $400 million on a project, as long as you don’t lose $50 billion in market value in an afternoon.
I'm happy they ate shit here because I like my mac not getting co-pilot bullshit forced into it, but apparently Apple had two separate teams competing against each other on this topic. Supposedly a lot of politics got in the way of delivering on a good product combined with the general difficulty of building LLM products.
I do prefer that Apple is opting to have everything run on device so you aren’t being exposed to privacy risks or subscriptions. Even if it means their models won’t be as good as ones running on $30,000 GPUs.
On device.
If you have say 16GB of GPU RAM and around 64GB of RAM and a reasonable CPU then you can make decent use of LLMs. I'm not a Apple jockey but I think you normally have something like that available and so you will have a good time, provided you curb your expectations.
I'm not an expert but it seems that the jump from 16 to 32GB of GPU RAM is large in terms of what you can run and the sheer cost of the GPU!
If you have 32GB of local GPU RAM and gobs of RAM you can rub some pretty large models locally or lots of small ones for differing tasks.
I'm not too sure about your privacy/risk model but owning a modern phone is a really bad starter for 10! You have to decide what that means for you and that's your thing and your's alone.
It also means that when the VC money runs dry, it's sustainable to run those models on-device vs. losing money running on those $$$$$ GPUs (or requiring consumers to opt for expensive subscriptions).
I’m kind of surprised to see people gloss over this aspect of it when so many folks here are in the “if I buy it, I should own it” camp.
> Apple had two separate teams competing against each other on this topic
That is a sign of very bad management. Overlapping responsibilities kill motivation as winning the infighting becomes more important than creating a good product. Low morale, and a blaming culture is the result of such "internal competition". Instead, leadership should do their work and align goals, set clear priorities and make sure that everybody rows in the same direction.
It’s how Apple (relatively famously?) developed the iPhone, so I’d assume they were using this as a model.
> In other words, should he shrink the Mac, which would be an epic feat of engineering, or enlarge the iPod? Jobs preferred the former option, since he would then have a mobile operating system he could customize for the many gizmos then on Apple’s drawing board. Rather than pick an approach right away, however, Jobs pitted the teams against each other in a bake-off.
https://www.nbcnews.com/news/amp/wbna44904886
But that's not the same thing right? That means having two teams competing for developing the next product. That's not two organisations handling the same responsibilities. You may still end up in problems with infighting. But if there is a clear end date for that competition and then no lasting effects for the "losers" this kind of "competition" will have very different effects than setting up two organisations that fight over some responsibility
Apparently? From what? Where did this information come from that they had two competing teams?
I feel like I hear people referring to Wayne Ma’s reporting for The Information to that effect.
https://www.theinformation.com/articles/apple-fumbled-siris-...
> Distrust between the two groups got so bad that earlier this year one of Giannandrea’s deputies asked engineers to extensively document the development of a joint project so that if it failed, Federighi’s group couldn’t scapegoat the AI team.
> It didn’t help the relations between the groups when Federighi began amassing his own team of hundreds of machine-learning engineers that goes by the name Intelligent Systems and is run by one of Federighi’s top deputies, Sebastien Marineau-Mes.
https://archive.is/Ncefp
This is a pretty good article, and worth reading if you aren't aware that Apple has seemingly mostly abandoned the vision of on-device AI (I wasn't aware of this)
Mac LLM vs Lisa LLM?
Apple ][ LLM Forever!
https://paleotronic.com/2025/08/03/connect-ai-to-microm8-app...
There are four main levers for improving an ML system:
1. You can change the training data.
2. You can change the objective function.
3. You can change the network topology.
4. You can change various hyperparameters (learning rate, etc.).
From there, I think it is better to look at the process as one of scientific discovery rather than a software debugging task. You form hypotheses and you try to work out how test them by mutating things in one of the four categories above. The experiments are expensive and the results are noisy, since the training process is highly randomized. A lot of times the effect sizes are so small it is hard to tell if they are real. The universe of potential hypotheses is large, and if you test a lot of them, you have to correct for the chance that some will look significant just by luck. But if you can add up enough small, incremental improvements, they can produce a total effect that is large.
The good news is that science has a pretty good track record of improving things over time. The bad news is that it can take a lot of time, and there is no guarantee of success in any one area.
I found this statement particularly relevant:
This possibility is compounded exponentially when MCP[0] is used.0 - https://github.com/modelcontextprotocol
I wonder if a safer approach to using MCP could involve isolating or sandboxing the AI. A similar context was discussed in Nick Bostrom's book Superintelligence. In the book, the AI is only allowed to communicate via a single light signal, comparable to Morse code.
Nevertheless, in the book, the AI managed to convince people, using the light signal, to free it. Furthermore, it seems difficult to sandbox any AI that is allowed to access dependencies or external resources (i.e. the internet). It would require (e.g.) dumping the whole Internet as data into the Sandbox. Taking away such external resources, on the other hand, reduces its usability.
Yeah in that regard we should always treat it like a junior something. Very much like you can't expect your own kids to never do something dangerous even if tell it for years to be careful. I got used to getting my kid from the Kindergarten with a new injury at least once a month.
I think it's very dangerous to use the term "junior" here because it implies growth potential, where in fact it's the opposite: you are using a finished product, it won't get any better. AI is an intern, not a junior. All the effort you're spending into correcting it will leave the company, either as soon as you close your browser or whenever the manufacturer releases next year's model -- and that model will be better regardless of how much time you waste on training this year's intern, so why even bother? Thinking of AI as a junior coworker is probably the least productive way of looking at it.
We should move well beyond human analogies. I have never met a human that would straight up lie about something, or build up so much deceptive tests that it might as well be lying.
Granted this is not super common in these tools, but it is essentially unheard of in junior devs.
> I have never met a human that would straight up lie about something
This doesn't match my experience. Consider high profile things like the VW emissions scandal, where the control system was intentionally programmed to only engage during the emissions test. Dictators. People are prone to lie when it's in their self interest, especially for self preservation. We have entire structures of government, courts, that try to resolve fact in the face of lying.
If we consider true-but-misleading, then politics, marketing, etc. come sharply into view.
I think the challenge is that we don't know when an LLM will generate untrue output, but we expect people to lie in certain circumstances. LLMs don't have clear self-interests, or self awareness to lie with intent. It's just useful noise.
There is an enormous amount of difference between planned deception as part of a product, and undermining your own product with deceptive reporting about its quality. The difference is collaboration and alignment. You might have evil goals, but if your developers are maliciously incompetent, no goal will be accomplished.
> it’s impossible for AI creators to definitively say their AI will never act maliciously or dangerously for any prompt it could be given
This is false, AI doesn't "act" at all unless you, the developer, use it for actions. In which case it is you, the developer, taking the action.
Anthropomorphizing AI with terms like "malicious" when they can literally be implemented with a spreadsheet—first-order functional programming—and the world's dumbest while-loop to append the next token and restart the computation—should be enough to tell you there's nothing going on here beyond next token prediction.
Saying an LLM can be "malicious" is not even wrong, it's just nonsense.
> bugs are usually caused by problems in the data used to train an AI
I think a fundamental problem is that many people assume that an LLM's failure to correctly perform a task is a bug that can be fixed somehow. Often times, the reason for that failure is simply a property of the AI systems we have at the moment.
When you accidentally drop a glass and it breaks, you don't say that it's a bug in gravity. Instead, you accept that it's a part of the system you're working with. The same applies to many categories of failures in AI systems: we can try to reduce them, but unless the nature of the system fundamentally changes (and we don't know if or when that will happen), we won't be able to get rid of them.
"Bug" carries an implication of "fixable" and that doesn't necessarily apply to AI systems.
But it's worse than that. Even if in theory the system could be fixed, we don't actually know how to fix it for real, the way we can fix a normal computer program.
The reason we can't fix them is because we have no idea how they work; and the reason we have no idea how they work is this:
1. The "normal" computer program, which we do understand, implement a neural network
2. This neural network is essentially a different kind of processor. The "actual" computer program for modern deep learning systems is the weights. That is, weights : neural net :: machine language : normal cpu
3. We don't program these weights; we literally summon them out of the mathematical aether by the magic of back-propagation and gradient descent.
This summoning is possible because the "processor" (the neural network architecture) has been designed to be differentiable: for every node we can calculate the slope of the curve with respect to the result we wanted, so we know "The final output for this particular bit was 0.7, but we wanted it to be 1. If this weight in the middle of the network were just a little bit lower, then that particular output would have been a little bit higher, so we'll bump it down a bit."
And that's fundamentally why we can't verify their properties or "fix" them the way we can fix normal computer programs: Because what we program is the neural network; the real program, which runs on top of that network, is summoned and not written.
That’s a very poetic description. Mine is simpler. It’s a generator. You train it to give it the parameters (weights) of the formula thag generate stuff (the formula is known). Then you give it some input data, and it will gives you an output.
Both the weights and the formula is known. But the weight are meaningless in a human fashion. This is unlike traditional software where everything from encoding (the meaning of the bits) to how the state machine (the cpu) was codified by humans.
The only ways to fix it (somewhat) is to come up with better training data (hopeless), a better formula, or tacking something on top to smooth the worst errors (kinda hopeless).
> The only ways to fix it
The correct way to fix it would be to build a decompiler to normal code, that would explain what it does, but this is akin to building the everything machine.
The part about AI being very sensitive to small perturbations of their input is actually a very active research topic (and coincidentally the subject of my PhD). Most vision AIs suffer from poor spatial robustness [1], you can drastically lower their accuracy simply by translating the inputs by well-chosen (adversarial) translations of a few pixels! I don't know much about text processing AIs but I can imagine their semantic robustness is also studied.
[1] https://arxiv.org/abs/1712.02779
Edit: typo
Is it fair to call this a robustness problem when you need access to the model to generate a failure case?
Many non-AI based systems lack robustness by the same standard (including humans)
> bugs are usually caused by problems in the data used to train an AI
This also is a misunderstanding.
The LLM can be fine, the training and data can be fine, but because the LLMs we use are non-deterministic (at least in regard to their being intentional attempts at entropy to avoid always failing certain scenarios) current algorithms are inherently by-design not going to always answer every question correctly that it potentially could have if the values that fall within a range had been specific values for that scenario. You roll the dice on every answer.
This is not necessarily a problem. Any programming or mathematical question has several correct answers. The problem with LLMs is that they don't have a process to guarantee that a solution is correct. They will give a solution that seems correct under their heuristic reasoning, but they arrived at that result in a non-logical way. That's why LLMs generate so many bugs in software and in anything related to logical thinking.
>> a solution that seems correct under their heuristic reasoning, but they arrived at that result in a non-logical way
Not quite ... LLMs are not HAL (unfortunately). They produce something that is associated with the same input, something that should look like an acceptable answer. A correct answer will be acceptable, and so will any answer that has been associated with similar input. And so will anything that fools some of the people, some of the time ;)
The unpredictability is a huge problem. Take the geoguess example - it has come up with a collection of "facts" about Paramaribo. These may or may-not be correct. But some are not shown in the image. Very likely the "answer" is derived from completely different factors, and the "explanation" in spurious (perhaps an explanation of how other people made a similar guess!)
The questioner has no way of telling if the "explanation" was actually the logic used. (It wasn't!) And when genuine experts follow the trail of token activation, the answer and the explanation are quite independent.
> Very likely the "answer" is derived from completely different factors, and the "explanation" in spurious (perhaps an explanation of how other people made a similar guess!)
This is very important and often overlooked idea. And it is 100% correct, even admitted by Anthropic themselves. When user asks LLM to explain how it arrived to a particular answer, it produces steps which are completely unrelated to the actual mechanism inside LLM programming. It will be yet another generated output, based on the training data.
Effortless lying, scary in humans, scarier in machines?
> Any programming or mathematical question has several correct answers.
Huh? If I need to sort the list of integer number of 3,1,2 in ascending order the only correct answer is 1,2,3. And there are multiple programming and mathematical questions with only one correct answer.
If you want to say "some programming and mathematical questions have several correct answers" that might hold.
I think what they meant is something along the lines of:
- In Math, there's often more than one logically distinct way of proving a theorem, and definitely many ways of writing the same proof, though the second applies more to handwritten/text proofs than say a proof in Lean.
- In programming, there's often multiple algorithms to solve a problem correctly (in the mathematical sense, optimality aside), and for the same algorithm there are many ways to implement it.
LLMs however are not performing any logical pass on their output, so they have no way of constraining correctness while being able to produce different outputs for the same question.
I find it quite ironical that while discussing the topic of logic and correct answers the OP talks rather "approximately" leaving the reader to imagine what he meant and others (like you) to spell it out.
Yes, I thought as well of your interpretation, but then I read the text again, and it really does not say that, so I choose to answer to the text...
No, but if you phrase it like "there are multiple correct answers to the question 'I have a list of integers, write me a computer program that sorts it'", that is obviously true. There's an enormous variety of different computer programs that you can write that sorts a list.
"1, 2, 3" is a correct answer
"1 2 3" is another
"After sorting, we get `1, 2, 3`" yet another
etc.
At least, that's how I understood GP's comment.
I think more charitably, they meant either that 1. There is often more than one way to arrive at any given answer, or 2. Many questions are ambiguous and so may have many different answers.
What about multiple notational variations?
1, 2, 3
1,2,3
[1,2,3]
1 2 3
etc.
What about them? It's possible for the question to unambiguously specify the required notational convention.
Is it? You have three wishes, which the maliciously compliant genie will grant you. Let’s hear your unambiguous request which definitely can’t be misinterpreted.
"Do my interpretation of the wish."
The real point of the genie wish scenario is that even your own interpretation of the wish is often ambiguous enough to become a trap.
"Do it so I am not surprised and don't change me."
> The problem with LLMs is that they don't have a process to guarantee that a solution is correct
Neither do we.
> They will give a solution that seems correct under their heuristic reasoning, but they arrived at that result in a non-logical way.
As do we, and so you can correctly reframe the issue as "there's a gap between the quality of AI heuristics and the quality of human heuristics". That the gap is still shrinking though.
I'll never doubt the ability of people like yourself to consistently mischaracterize human capabilities in order to make it seem like LLMs' flaws are just the same as (maybe even fewer than!) humans. There are still so many obvious errors (noticeable by just using Claude or ChatGPT to do some non-trivial task) that the average human would simply not make.
And no, just because you can imagine a human stupid enough to make the same mistake, doesn't mean that LLMs are somehow human in their flaws.
> the gap is still shrinking though
I can tell this human is fond of extrapolation. If the gap is getting smaller, surely soon it will be zero, right?
> doesn't mean that LLMs are somehow human in their flaws.
I don't believe anyone is suggesting that LLMs flaws are perfectly 1:1 aligned with human flaws, just that both do have flaws.
> If the gap is getting smaller, surely soon it will be zero, right?
The gap between y=x^2 and y=-x^2-1 gets closer for a bit, fails to ever become zero, then gets bigger.
The difference between any given human (or even all humans) and AI will never be zero: Some future AI that can only do what one or all of us can do, can be trivially glued to any of that other stuff where AI can already do better, like chess and go (and stuff simple computers can do better, like arithmetic).
> I'll never doubt the ability of people like yourself to consistently mischaracterize human capabilities
Ditto for your mischaracterizations of LLMs.
> There are still so many obvious errors (noticeable by just using Claude or ChatGPT to do some non-trivial task) that the average human would simply not make.
Firstly, so what? LLMs also do things no human could do.
Secondly, they've learned from unimodal data sets which don't have the rich semantic content that humans are exposed to (not to mention born with due to evolution). Questions that cross modal boundaries are expected to be wrong.
> If the gap is getting smaller, surely soon it will be zero, right?
Quantify "soon".
Humans learn. They don't recreate the world from scratch every time they start a new CLI session.
Human errors in judgement can also be discovered, explained, and reverted.
> That the gap is still shrinking though.
citation needed
Eh, proofs and logic have entered the room!
I think sometimes it gives a "wrong" answer not because it wasn't trained well, but because it could give multiple plausible answers and just happened to land on the unhelpful one
> With AI systems, almost all bad behaviour originates from the data that’s used to train them
Careful with this - even with perfect data (and training), models will still get stuff wrong.
> Because eventually we’ll iron out all the bugs so the AIs will get more reliable over time
Honestly this feels like a true statement to me. It's obviously a new technology, but so much of the "non-deterministic === unusable" HN sentiment seems to ignore the last two years where LLMs have become 10x as reliable as the initial models.
They have certainly gotten better, but it seems to me like the growth will be kind of logarithmic. I'd expect them to keep getting better quickly for a few more years and then kinda slow and eventually flatline as we reach the maximum for this sort of pattern matching kind of ML. And I expect that flat line will be well below the threshold needed for, say, a small software company to not require a programmer.
> kind of logarithmic
https://en.wikipedia.org/wiki/Sigmoid_function
Ironically, yes. :)
Ahaha, of course nothing will ever be able to do my job!
It will take at least a few more decades at least from the looks of it. I would be 6 feet under by then so yes "Nothing will ever be able to do my job".
Since "AGI has been achieved internally" tweet I've only seen incremental improvements that are guaranteed to never be able to do my job. Or most people's jobs.
Ironically I came to the comments to point out that all over hackernews you see this sentiment repeated, and that's by a group I would consider to be far more technically competent then your average person. And very helpfully there is one just a few comments down from the top.
Technical competence and an interest in sociological development do not always coincide. Technology often seeks simplicity, whereas sociology examines inherently complex human behavior.
Emergent misalignment and power seeking isn't a bug we can squash with a PR and a unit test
I read this as "10x better at generating code that looks correct but hides nasty bugs behind a facade of sane looking slop".
Right away my mind went to "well, are people more reliable than they used to be?" and I'm not sure they are.
Of course LLMs aren't people, but an AGI might behave like a person.
By the time a junior dev graduates to senior, I expect that they'll be more reliable. In fact, at the end of each project, I expect the junior dev to have grown more reliable.
LLMs don't learn from a project. At best, you learn how to better use the LLM.
They do have other benefits, of course, i.e. once you have trained one generation of Claude, you have as many instances as you need, something that isn't true with human beings. Whether that makes up for the lack of quality is an open question, which presumably depends on the projects.
> LLMs don't learn from a project.
How long do you think that will remain true? I've bootstrapped some workflows with Claude Code where it writes a markdown file at the end of each session for its own reference in later sessions. It worked pretty well. I assume other people are developing similar memory systems that will be more useful and robust than anything I could hack together.
For LLMs? Mostly permanently. This is a limitation of the architecture. Yes, there are workarounds, including ChatGPT's "memory" or your technique (which I believe are mostly equivalent), but they are limited, slow and expensive.
Many of the inventors of LLMs have moved on to (what they believe are) better models that would handle such learnings much better. I guess we'll see in 10-20 years if they have succeeded.
Permanently.
There’s an interplay between two different ideas of reliability here.
LLMs can only provide output which is somehow within training boundaries.
We can get better at expanding the area within these boundaries.
It will still not be reliable like code is.
Older people are generally more reliable than younger people.
I'm not sure that's generally true. However, older people have a track record, and a reliable older person is likely to be more reliable than a younger person without such a track record.
Reliable has different meanings. I think in this case the meaning is closer to "deterministic" and "follows instructions." An older worker will more reliably behave the same way twice, and more reliably follow the same set of instructions they've been following throughout their career.
> To make this more concrete, here are some example ideas that are perfectly true when applied to regular software but become harmfully false when applied to modern AIs: ...
Hmm, I don't think any of these were true with non-AI software. Commonly held beliefs, sure.
If anything, I am glad AI is helping us revisit these assumptions.
- Software vulnerabilities are caused by mistakes in the code
Setting aside social engineering, mistake implies these were knowable in advance. Was the lack of TLS in the initial HTTP spec a mistake?
- Bugs in the code can be found by carefully analysing the code
If this was the case, why do people reach for rewriting buggy code they don't understand?
- Once a bug is fixed, it won’t come back again
Too many counter examples to this one in my lived experience.
- Every time you run the code, the same thing happens
Setting aside seeding PRNGs, there's the issue of running the code on different hardware. Or failing hardware.
- If you give specifications beforehand, you can get software that meets those specifications
I have never seen this work without needing to revise the specification during implementation.
The "just find the bug and patch it" mindset is so deeply ingrained in engineering culture that it's easy to forget it doesn't apply here
And the public. Look at this article: https://www.bbc.co.uk/news/articles/cj07ley3jnpo
> "Oh my goodness, it worked, it's amazing it's finally been updated," she tells the BBC. "This is a great step forward."
She thinks someone noticed the bug about not being able to show one-armed people, figured out why it wasn't working and wrote a fix.
My current method for trying to break through this misconception is informing people that nobody knows how AI works. Literally. Nobody knows. (Note that knowing how to make something is not the same as knowing how it works. Take humans as an obvious example.)
I think you’re confusing knowing how a system works with being able to predict how that system will perform.
In a non-linear system the former is often easier than the latter. For example we know how planets “work” from the laws of motion. But planetary orbits involving > 2 bodies are non-linear, and predicting their motion far into the future is surprisingly difficult.
Neural networks are the same. They’re actually quite simple, it’s all undergraduate maths and statistics. But because they’re non-linear systems, predicting their behaviour is practically impossible.
I think there are people who know exactly how it works, they know how neural networks work, they know how the transformer architecture works, how attention works, embeddings, tokenization, etc. We just can’t define the weights of the connections between the neurons.
That's a bit like saying knowing how a pipe works is enough to explain a combustion engine. You're just listing part of how LLMs work.
Those mechanisms only explain next word prediction, not LLM reasoning.
That's an emergent property that no person, as far as I understand it, can explain past hand waving.
Happy to be corrected here.
There's no magic involved, the LLM creators can go anywhere and rebuild an LLM with pretty much the same outcome, if they have the same training data. With unlimited time you could even reproduce the output of an LLM manually as it is just a lot of mathematics. Including reasoning, as that is mostly adding words in the context that will steer the word predication to include reasoning. As this is a useful LLM behavior this context is now part of the training data, so it becomes part of the neural network weights.
Yes, but we can't inspect, reproduce or explain the emergent property independently. We can't pick out the "math reasoning" part or the "programming" part, or inspect how it's working, or selectively change any of it. You can't turn any dials or twiddle any knobs. You can't replace one part with another, or pick out components. You can't peek inside and say:
"hey it's got an irrational preference for naming its variables after famous viking warriors, lets change that!"
But worse, it's not that you can't change it, you just don't know! All you can do is test it and guess its biases.
Is it racist, is it homophobic, is it misogynistic? There was an article here the other day about AI in recruitment and the hidden biases. And there was a recruitment AI that only picked men for a role. The job spec was entirely gender neutral. And they hadn't noticed until a researcher looked at it.
It's a black box. So if it does something incorrectly, all they can do is retrain and hope.
Again, this is my present understanding of how it all works right now.
Well we are happy to work with other humans without knowing how they work.
How is it possible that nobody knows how it works - it’s running on hardware we have complete control over and perfect observability into, is it not? At any frame we can pause, examine the state, then step forward, examine the state, and observe what changes have occurred - we have perfect knowledge of the source code, the compiler, whatever components you prefer to break software down into -
What is it that we don’t understand?
The source code is not the LLM. The LLM is billioms of random floating point numbers that somehow encode everything the model knows and can do.
The ML field has a good understanding of the algorithms that produce these floating point numbers and lots of techniques that seem to produce “better” numbers in experiments. However, there is little to no understanding of what the numbers represent or how they do the things they do.
We know how each of the "parts" work, but there is a gazillion of parts (especially since you need to take the model weights into account, which are way larger in size than the code that generates them or uses them to generate stuff), and we found out that together they do something that we do not really understand why they do it.
And inspecting each part is not enough to understand how, together, they achieve what they achieve. We would need to understand the entire system in a much more abstract way, and currently we have nothing more than ideas of how it _might_ work.
Normally, with software, we do not have this problem, as we start on the abstract level with a fully understood design and construct the concrete parts thereafter. Obviously we have a much better understanding of how the entire system of concrete parts works together to perform some complex task.
With AI, we took the other way: concrete parts were assembled with vague ideas on the abstract level of how they might do some cool stuff when put together. From there it was basically trial-and-error, iteration to the current state, but always with nothing more than vague ideas of how all of the parts work together on the abstract level. And even if we just stopped the development now and tried to gain a full, thorough understanding of the abstract level of a current LLM, we would fail, as they already reached a complexity that no human can understand anymore, even when devoting their entire lifetime to it.
However, while this is a clear difference to most other software (though one has to get careful when it comes to the biggest projects like Chromium, Windows, Linux, ... since even though these were constructed abstract-first, they have been in development for such a long time and have gained so many moving parts in the meantime that someone trying to understand them fully on the abstract level will probably start to face the difficulty of limited lifetime as well), it is not an uncommon thing per se: we also do not "really" understand how economy works, how money works, how capitalism works. Very much like with LLMs, humanity has somehow developed these systems through interaction of billions of humans over a long time, there was never an architect designing them on an abstract level from scratch, and they have shown emergent capabilities and behaviors that we don't fully understand. Still, we obviously try to use them to our advantage every day, and nobody would say that modern economies are useless or should be abandoned because they're not fully understood.
Nobody knows (full scope and on every level) how human brains work. Still bosses rely on their employees' brains all the time.
If I need to manage an AI as I would manage an employee's brain, I'm going to need quite a few non-technical resources to actually achieve that: time, willpower to babysit it, ability to motivate it, leverage in the form of incentives (and reprimands), to name a few.
AI sits at a weird place where it can't be analyzed as software, and it can't be managed as a person.
My current mental model is that AGI can only be achieved when a machine experiences pleasure, pain, and "bodily functions". Otherwise there's no way to manage it.
> Nobody knows (full scope and on every level) how human brains work.
That is what the parent meant.
And my point is, that it does not matter. That we still rely on all kinds of things we do not fully grasp.
> One popular dataset, FineWeb, is about 11.25 trillion words long3, which, if you were reading at about 250 words per minute, would take you over 85 thousand years to read. It’s just not possible for any single human (or even a team of humans) to have read everything that an LLM has read during training.
Do you have to read everything in a dataset with your own eyes to make sense of it? This would make any attempt to address bias in the dataset impossible, and I think it's not, so there should be other ways to make sense of the dataset distribution without having to read it yourself.
It would help if this piece was clearer about the context in which "AI bugs" reveal themselves. As an argument for why you shouldn't have LLMs making unsupervised real-time critical decisions, these points are all well taken. AI shouldn't be controlling the traffic lights in your town. We may never reach a point where it can. But among technologists, the major front on which these kinds of bugs are discussed is coding agents, and almost none of these points apply directly to coding agents: agent coding is (or should be) a supervised process.
> nobody knows precisely what to do to ensure an AI writes formal emails correctly or summarises text accurately.
This is a bit of a hyperbole, a lot of the recent approaches rely on MoE, that are specialized. This makes it much more usuable for simple usecases.
Fortunately, we can have LLMs write code and keep all the benefits of normal software (determinism, reproducibility, permanent bug fixes etc.)
I don’t think anyone is advocating for web apps to take the form of an LLM prompt with the app getting created on the fly every time someone goes to the url.
The most likely danger with AI is concentrated power, not that sentient AI will develop a dislike for us and use us as "batteries" like in the Matrix.
The reality is that the CEO/executive class already has developed a dislike for us and is trying to use us as “batteries” like in the Matrix.
Do you know personally some CEO-s? I know a couple and they generally seem less empathic than the general population, so I don't think that like/dislike even applies.
On the other hand, trying to do something "new" is lots of headaches, so emotions are not always a plus. I could make a parallel to doctors: you don't want a doctor to start crying in a middle of an operation because he feels bad for you, but you can't let doctors doing everything that they want - there needs to be some checks on them.
I would say that the parallel is not at all accurate because the relationship between a doctor and a patient undergoing surgery is not the same as the one you and I have with CEOs. And a lot of good doctors have emotions and they use them to influence patient outcomes positively.
Even then, a psychopathic doctor at least has their desired outcomes mostly aligned with the patients.
CEOs (even most VCs) are labor too
Labor competes for compensation, CEOs compete for status (above a certain enterprise size, admittedly). Show me a CEO willingly stepping down to be replaced by generative AI. Jamie Dimon will be so bold to say AI will bring about a 3 day week (because it grabs headlines [1]) but he isn't going to give up the status of running JPMC; it's all he has besides the wealth, which does not appear to be enough. The feeling of importance and exceptionalism is baked into the identity.
[1] https://fortune.com/article/jamie-dimon-jpmorgan-chase-ceo-a...
Spoiler there’s no reason we couldn’t work three days a week now. And 100 might be pushing it, but having life expectancy to 90 as well within our grass today as well. We have just decided not to do that.
The reason we don't have 3 day weeks is because the system rewards revenue, not worker satisfaction.
That's the market's job. Once AI CEOs start outperforming human CEOs, investment will flow to the winners. Give it 5-10 years.
(Has anyone tried an LLM on an in-basket test? [1] That's a basic test for managers.)
[1] https://en.wikipedia.org/wiki/In-basket_test
Not if CEOs use their political power to make it illegal.
Almost everyone is "labor" to some extent. There is always a huge customer or major investor that you are beholden to. If you are independently wealthy then you are the exception.
Bingo
Do they know it?
Until shareholders treat them as such, they will remain in the ruling class
To me, the greatest threat is information pollution. Primary sources will be diluted so heavily in an ocean of generated trash that you might as well not even bother to look through any of it.
I see that as the death knell for general search engines built to indiscriminately index the entire web. But where that sort of search fails, opportunities open up for focused search and curated search.
Just as human navigators can find the smallest islands out in the open ocean, human curators can find the best information sources without getting overwhelmed by generated trash. Of course, fully manual curation is always going to struggle to deal with the volumes of information out there. However, I think there is a middle ground for assisted or augmented curation which exploits the idea that a high quality site tends to link to other high quality sites.
One thing I'd love is to be able to easily search all the sites in a folder full of bookmarks I've made. I've looked into it and it's a pretty dire situation. I'm not interested in uploading my bookmarks to a service. Why can't my own computer crawl those sites and index them for me? It's not exactly a huge list.
And it imitates all the unimportant bits perfectly (like spelling, grammar, word choice) while failing at the hard to verify important bits (truth, consistency, novelty)
It’s already been happening but now it’s accelerated beyond belief. I saw a video about how WW1 reenactment photos end up getting reposted away from their original context and confused with original photos to the point it’s impossible to tell unless you can track it back to the source.
Now most of the photos online are just AI generated.
Concentrated power is kinda a pre-requisite for anything bad happening, so yes, it's more likely in exactly the same way that given this:
"Linda is a bank teller" is strictly more likely than "Linda is a bank teller and is active in the feminist movement" — all you have is P(a)>P(a&b), not what the probability of either statement is.Why does an AI need the ability to "dislike" to calculate that its goals are best accomplished without any living humans around to interfere? Superintelligence doesn't need emotions or consciousness to be dangerous.
It needs to optimize for something. Like/dislike is an anthropomorphization of the concept.
It's an unhelpful one because it implies the danger is somehow the result of irrational or impulsive thought, and making the AI smarter will avoid it.
That's not how I read it.
Perhaps because most of the smartest people I know are regularly irrational or impulsive :)
I think most people don't get that; look at how often even Star Trek script writers write Straw Vulcans*.
* https://tvtropes.org/pmwiki/pmwiki.php/Main/StrawVulcan
The power concentration is already massive, and a huge problem indeed. The ai is just a cherry on top. The ai is not the problem.
"AI will take over the world".
I hear that. Then I try to use AI for simple code task, writing unit tests for a class, very similar to other unit tests. If fails miserably. Forgets to add an annotation and enters in a death loop of bullshit code generation. Generates test classes that tests failed test classes that test failed test classes and so on. Fascinating to watch. I wonder how much CO2 it generated while frying some Nvidia GPU in an overpriced data center.
AI singularity may happen, but the Mother Brain will be a complete moron anyway.
Regularly trying to use LLMs to debug coding issues has convinced me that we're _nowhere_ close to the kind of AGI some are imagining is right around the corner.
Sure, but also the METR study showed the rate of change is t doubles every 7 months where t ~= «duration of human time needed to complete a task, such that SOTA AI can complete same with 50% success»: https://arxiv.org/pdf/2503.14499
I don't know how long that exponential will continue for, and I have my suspicions that it stops before week-long tasks, but that's the trend-line we're on.
But will it actually get better or will it just get faster and more power efficient at failing to pair parentheses/braces/brackets/quotes?
Read the linked METR study please.
Or watch the Computerphile video summary/author interview, if you prefer: https://m.youtube.com/watch?v=evSFeqTZdqs
At least Mother Brain will praise your prompt to generate yet another image in the style of Studio Ghibli as proof that your mind is a tour de force in creativity, and only a borderline genius would ask for such a thing.
Most reasonable AI alarmists are not concerned with sentient AI but an AI attached to the nukes that gets into one of those repeating death loops and fires all the missiles.
In reality, this isn't a very serious threat. Rather, we're concerned about AI as a tool for strengthening totalitarian regimes.
"Just one more prompt, bro", and your problems will be solved.
You can say that, and I might even agree, but many smart people disagree. Could you explain why you believe that? Have you read in detail the arguments of people who disagree with you?
Seems like a self fulfilling prophecy
Definitely not ‘self’ fulfilling. There are plenty of people actively and vigorously working to fulfill that particular reality.
> power resides where men believe it resides
And also where people believe that others believe it resides. Etc...
If we can find new ways to collectively renegotiate where we think power should reside we can break the cycle.
But we only have time to do this until people aren't a significant power factor anymore. But that's still quite some time away.
I agree.
Our best technology at current require teams of people to operate and entire legions to maintain. This leads to a sort of balance, one single person can never go too far down any path on their own unless they convince others to join/follow them. That doesn't make this a perfect guard, we've seen it go horribly wrong in the past, but, at least in theory, this provides a dampening factor. It requires a relatively large group to go far along any path, towards good or evil.
AI reduces this. How greatly it reduces this, if it reduces it to only a handful, to a single person, or even to 0 people (putting itself in charge), seems to not change the danger of this reduction.
I'm not so sure it will be that either, it would be having multiple AIs essentially at war with each other over access to GPUs/energy or whatever the materials are needed to grow if/when that happens. We will end up as pawns in this conflict.
Given that even fairly mediocre human intelligences can run countries into the ground and avoid being thrown out in the process, it's certainly possible for an AI to be in the intelligence range where it's smart enough to win vs humans but also dumb enough to turn us into pawns rather just go to space and blot out the sun with a Dyson swarm made from the planet Mercury.
But don't count on it.
I mean, apart from anything else, that's still a bad outcome.
For one thing, we'd make shit batteries.
IIRC the original idea was that the machines used our brain capacity as a distributed array but then they decided batteries was easier to understand while been sillier, just burn the carbon they are feeding us, it’s more efficient.
If I could write the matrix reverted, Neo would discover that the last people put themselves in the pods because the world was so fucked up, and the machines had been caretakers that were trying to protect them from themselves. That revision would make the first movie perfect.
Given that the first Matrix was a paradise that's pretty much canon if you ignore the duracell.
Sounds about right, most of us already are. But why would the AI need our shit? Surely it wants electricity?
They farm you for attention, not electricity. Attention (engagement time) is how they quantify "quality" so that it can be gamed with an algorithm.
I mean, you can't really disprove either being an issue.
It's great to help people understand that AI can be both surprisingly good and disappointing, and that testing is the only way to know, but it's impossible to test everything. That sets expectations.
I think that means savvy customers will want details or control over testing, and savvy providers will focus on solutions they can validate, or where testing is included in the workflow (e.g., code), or where precision doesn't matter (text and meme generation). Knowing that in depth is gold for AI advocates.
Otherwise, I don't think people really know or care about bugs or specifications or how AI breaks prior programmer models.
But people will become very hostile and demand regulatory frenzies if AI screws things up (e.g., influencing elections or putting people out of work). Then no amount of sympathy or understanding will help the industry, which has steadily been growing its capability for evading regulation via liability disclaimers, statutory exceptions, arbitration clauses, pitting local/regional/national governments against each other, etc.
To me that's the biggest risk: we won't get the benefits and generational investments will be lost in cleaning up after a few (even accidental) bad actors at scale.
> bad behaviour isn’t caused by any single bad piece of data, but by the combined effects of significant fractions of the dataset
Related opposing data point to this statement: https://news.ycombinator.com/item?id=45529587
"Signficiant fraction" does not imply (to this data scientist) a large fraction.
The belief that it is useful.
I think your post is fundamentally wrong. See, you are comparing AI responses with the written code, which may not be the fair comparison. I see it as, better you could compare code generated by AI vs the code written by an engineer.
The original author seems to view the AI application as itself a software application which has desired or undesired, and predictable or unpredictable, behaviors. That doesn't seem like an invalid thing to talk about merely because there are other software-related conversations we can have about AIs (or other code-quality-related conversations).
AI responses and code generated by AI are literally the same thing
I'm confused by this comment. That's a completely different discussion.
I don't understand the "your boss" framing of this article, or more accurately, the title of this article. The article contents don't actually seem to have anything to do with management specifically. Is the reader is meant to believe that not being scared of AI is a characteristic of the managerial class? Is the unstated implication that there is some class warfare angle and anybody who isn't against AI is against laborers? Because what the article actually overtly argues, without any reading between the lines, is quite mundane.
> Is the unstated implication that there is some class warfare angle and anybody who isn't against AI is against laborers?
I didn't read it that way. I read "your boss" as basically meaning any non-technical person who may not understand the challenges of harnessing LLMs compared to traditional, (more) deterministic software development.
Indeed, from reading the article I could really see any discussion of "your boss", so I changed the title to something more representative, and a condensed version of a phrase from the article.
> In regular software, vulnerabilities are caused by mistakes in the lines of code that make up the software
> in modern AI systems, vulnerabilities or bugs are usually caused by problems in the data used to train an AI
In regular software, vulnerabilities are caused by lack of experience, therefor lack of proper training materials.
I think they're more caused by rushed deadlines, poor practices, and/or bad QA. Some folks just don't get it either and training doesn't help.
Apple’s underwhelming LLM rollout—like the pulled notification summaries and trivial emoji tools—proves even big tech struggles to turn AI hype into reliable, daily-useful features; I’d take a working email organizer over a glitchy "smart" summary any day.
> It’s entirely possible that some dangerous capability is hidden in ChatGPT, but nobody’s figured out the right prompt just yet.
This sounds a little dramatic. The capabilities of ChatGPT are known. It generates text and images. The qualities of the content of the generated text and images is not fully known.
Think of the news about the kid who got recommended to suicide by ChatGPT, or chatgpt providing the user information on how to do illegal activities, these capabilities are the ones that the author it's referring to
And that sounds a little reductive. There's a lot that can be done with text and images. Some of the most influential people and organizations in the world wield their power with text and images.
Yeah, and to riff off the headline, if something dangerous is connected to and taking commands from ChatGPT then you better make sure there’s a way to turn it off.
Also, there's a reason AI Red Teaming is now an ask that is getting line item funding from C-Suites.
Plus there is the 'monkeys with typewriters' problem with both danger and hypothetical good. In contrast, ChatGPT may technically reply to the right prompt with a universal cancer cure/vaccine. Psuedorandomly generating it wouldn't help as you wouldn't recognize it from all of the other queries of things we don't know of as true or false.
Likewise what to ask it for how to make some sort of horrific toxic chemical, nuclear bomb, or similar isn't much good if you cannot recognize it and dangerous capability depends heavily on what you have available to you. Any idiot can be dangerous with C4 and detonator or bleach and ammonia. Even if ChatGPT could give entirely accurate instructions on how to build an atomic bomb it wouldn't do much good because you wouldn't be able to source the tools and materials without setting off red flags.
think it's missing the biggest assumption, which is not necessarily true for regular software either but much more true than AI:
The same inputs should produce the same outputs.
And that assumption is important because dependability is the strength of an automated process.
It's not missing? https://boydkane.com/essays/boss#every-time-you-run-the-code...
that's what I get for trying to read with a screaming kid demanding music videos in my ear.
Aim bosses at this article in The Economist.[1] If your boss doesn't read The Economist, you need to escalate to a level that does.
[1] https://www.economist.com/leaders/2025/09/25/how-to-stop-ais...
https://archive.is/R0RJB
Management summary, from The Economist article:
"The worst effects of this flaw are reserved for those who create what is known as the “lethal trifecta”. If a company, eager to offer a powerful AI assistant to its employees, gives an LLM access to un-trusted data, the ability to read valuable secrets and the ability to communicate with the outside world at the same time, then trouble is sure to follow. And avoiding this is not just a matter for AI engineers. Ordinary users, too, need to learn how to use AI safely, because installing the wrong combination of apps can generate the trifecta accidentally."
Not the point, but I’m confused by the Geoguessr screenshot. Under the reasoning for its decision, it mentions “traffic keeps to the left” but that is not apparent from the photo.
Then it says the shop sign looks like a “Latin alphabet business name rather than Spanish or Portuguese”. Uhhh… what? Spanish and Portuguese use the Latin alphabet.
What is 12+12?
> The answer is 24! See the ASCII values of '1' is 49, '2' is 50, and '+' is 43. Adding all that together we get 3. Now since we are doing this on a computer with a 8-bit infrastructure we multiply by 3 and so the answer is 24.
Cool! I didn't understand any of that but it was correct and you sound smart. I will put this thing in charge of critical parts of my business.
It's an LLM.
It decided on the first line first (the place name), and then made the reasons on the rest of the text.
So the answer is more important on the justifications than the actual picture, and the reasoning that led it there doesn't enter the frame at all.
The said "don't use magic numbers" but LLMs are made almost entirely (by weight) of magic numbers...
Tremendous alpha right now in making scary posts about AI. Fear drives clicks. You don't even need to point to current problems, all you have to do is say we can't be sure they won't happen in the future.
How the tables have turned.
Maybe AI isn't ready to take over the world yet, it still can't write a simple unit test without getting stuck in a loop.
Even Apple found out that making AI work is harder than making emojis, maybe the hype train needs a reality check.
Warez is illegal vs warez is legal. In this case, warez is Anna's library and stuff.
The article doesn’t even mention prompting. Wha? Is it just talking about the ML foundations, not applications?
The main thing is that LLMs aren't software programs and as such should not be compared to them.
All the same criticisms are true about hiring humans. You don’t really know what they’re thinking, you don’t really know what their values and morals are, you can’t trust that they’ll never make a mistake, etc.
I think you're misreading the article; the point here is not "LLMs are bad and can't replace humans," the point is that many non-technical people have the expectation that LLMs can replace humans _but still behave like regular software_ with regard to reliability and operability.
When a CEO sees their customer chatbot call a customer a slur, they don't see "oh my chatbot runs on a stochastic model of human language and OpenAI can't guarantee that it will behave in an acceptable way 100% of the time", they see "ChatGPT called my customer a slur, why did you program it to do that?"
You can teach a human when they make a mistake. Can you do the same for an LLM?
Am I correct to assume "modern AI system" means "neural network"?
> It’s entirely possible that some dangerous capability is hidden in ChatGPT, but nobody’s figured out the right prompt just yet.
Or they have, but chose to exploit or stockpile it rather than expose it.
Boss: You can just turn it off, can't you?
Me: Ask me later.
But this is why using the AI in the production of (almost) deterministic systems makes so much sense, including saving on execution costs.
ISTR someone else round here observing how much more effective it is to ask these things to write short scripts that perform a task than doing the task themselves, and this is my experience as well.
If/when AI actually gets much better it will be the boss that has the problem. This is one of the things that baffles me about the managerial globalists - they don't seem to appreciate that a suitably advanced AI will point the finger at them for inefficiency much more so than at the plebs, for which it will have a use for quite a while.
A bunch of short scripts doesn't easily lead to a large-scale robust software platform.
I guess if managers get canned, it'll be just marketing types left?
>that baffles me about the managerial globalists
It's no different from those on HN that yell loudly that unions for programmers are the worst idea ever... "it will never be me" is all they can think, then they are protesting in the streets when it is them, but only after the hypocrisy of mocking those in the street protesting today.
Agreed. My dad was raised strongly fundamentalist, and in North America, that included (back then) strongly resisting unions. In hindsight, I've come to realize that my parent's weren't maybe even of average intelligence, and definitely of above-average gullibility.
Unionized software engineers would solve a lot of the "we always work 80 hour weeks for 2 months at the end of a release cycle" problems, the "you're too old, you're fired" issues, the "new hires seems to always make more than the 5/10+ year veterans", etc. Sure, you wouldn't have a few getting super rich, but it would also make it a lot easier for "unionized" action against companies like Meta, Google, Oracle, etc. Right now, the employers hold like 100x the power of the employees in tech. Just look at how much any kind of resistance to fascism has dwindled after FAANG had another round of layoffs..
Software "engineers" totally miss a key thing in other engineering professions as well, which is organizations to enforce some pretense of ethical standards to help push back against requests from product. Those orgs often look a lot like unions.
This article makes a solid case. The worst kinds of bugs in software are not the most obvious ones like syntax errors, they are the ones where the code appears to be working correctly, until some users do something slightly unusual after a few weeks of some code change being deployed and it breaks spectacularly but the bug only affects a small fraction of users so developers cannot reproduce the issue... And the cose change happened such time ago that the guilty code isn't even suspected.
Lovely blog, RSS please.
There's... something at https://boydkane.com/index.xml
I guessed the URL based on the Quartz docs. It seems to work but only has a few items from https://boydkane.com/essays/
the author (either of the blog or its software) would do well to consult https://www.petefreitag.com/blog/rss-autodiscovery/
> here are some example ideas that are perfectly true when applied to regular software
Hm, I'm listening, let's see.
> Software vulnerabilities are caused by mistakes in the code
That's not exactly true. In regular software, the code can be fine and you can still end up with vulnerabilities. The platform in which the code is deployed could be vulnerable, or the way it is installed make it vulnerable, and so on.
> Bugs in the code can be found by carefully analysing the code
Once again, not exactly true. Have you ever tried understanding concurrent code just by reading it? Some bugs in regular software hide in places that human minds cannot probe.
> Once a bug is fixed, it won’t come back again
Ok, I'm starting to feel this is a troll post. This guy can't be serious.
> If you give specifications beforehand, you can get software that meets those specifications
Have you read The Mythical Man-Month?
You should read the footnote marked [1] after "a note for technical folk" at the beginning of the article. He is very consciously making sweeping generalizations about how software works in order to make things intelligible to non-technical readers.
But are those sweeping generalizations true?
> I’m also going to be making some sweeping statements about “how software works”, these claims mostly hold, but they break down when applied to distributed systems, parallel code, or complex interactions between software systems and human processes.
I'd argue that this describes most software written since, uh, I hesitate to even commit to a decade here.
For the purposes of the article, which is to demonstrate how developing an LLM is completely different from developing traditional software, I'd say they are true enough. It's a CS 101 understanding of the software development lifecycle, which for non-technical readers is enough to get the point across. An accurate depiction of software development would only obscure the actual point for the lay reader.
At least the 1950’s. That’s when stuff like asynchrony and interrupts were worked out. Dijkstra wrote at length about this in reference to writing code that could drive a teletype (which had fundamentally non-deterministic timings).
If you include analog computers, then there are some WWII targeting computers that definitely qualify (e.g., on aircraft carriers).
Sure, but:
> these claims mostly hold, but they break down when applied to distributed systems, parallel code, or complex interactions between software systems and human processes
The claims the GP quoted DON’T mostly hold, they’re just plain wrong. At least the last two, anyway.
Does that really matter?
He is trying to lax the general public perception around AIs shortcomings. He's giving AI a break, at the expense of regular developers.
This is wrong on two fronts:
First, because many people foresaw the AI shortcomings and warned about them. This "we can't fix a bug like in regular software" theatre hides the fact that we can design better benchmarks, or accountability frameworks. Again, lots of people foresaw this, and they were ignored.
Second, because it puts the strain on non-AI developers. It blamishes all the industry, putting together AI with non-AI in the same bucket, as if AI companies stumbled on this new thing and were not prepared for its problems, when the reality is that many people were anxious about the AI companies practices not being up to standard.
I think it's a disgraceful take, that only serves to sweep things under a carpet.
I don't think he's doing that at all. The article is pointing out to non-technical people how AI is different than traditional software. I'm not sure how you think it's giving AI a break, as it's pointing out that it is essentially impossible to reason about. And it's not at the expense of regular developers because it's showing how regular software development is different than this. It makes two buckets, and puts AI in one and non-AI in the other.
He is. Maybe he's just running with the pack, but that doesn't matter either.
The fact is, we kind of know how to prevent problems in AI systems:
- Good benchmarks. People said several times that LLMs display erratic behavior that could be prevented. Instead of adjusting the benchmarks (which would slow down development), they ignored the issues.
- Accountability frameworks. Who is responsible when an AI fails? How the company responsible for the model is going to make up for it? That was a demand from the very beginning. There are no such accountability systems in place. It's a clown fiesta.
- Slowing down. If you have a buggy product, you don't scale it. First, you try to understand the problem. This was the opposite of what happened, and at the time, they lied that scaling would solve the issues (when in fact many people knew for a fact that scaling wouldn't solve shit).
Yes, it's kind of different. But it's a different we already know. Stop pushing this idea that this stuff is completely new.
>But it's a different we already know
'we' is the operative word here. 'We', meaning technical people who have followed this stuff for years. The target audience of this article are not part of this 'we' and this stuff IS completely new _for them_. The target audience are people who, when confronted with a problem with an LLM, think it is perfectly reasonable to just tell someone to 'look at the code' and 'fix the bug'. You are not the target audience and you are arguing something entirely different.
Let's pretend I'm the audience, and imagine that in the past I said those things ("fix the bug" and "look at the code").
What should I say now? "AI works in mysterious ways"? Doesn't sound very useful.
Also, should I start parroting innacurate outdated generalizations about regular software?
The post doesn't teach anything useful for a beginner audience. It's bamboozling them. I am amazed that you used the audience perspective as a defense of some kind. It only made it worse.
Please, please, take a moment to digest my critique properly. Think about what you just said and what that implies. Re-read the thread if needed.
I thought this blog post was a parody. And to my surprise both the author and the audience takes it seriously. Weird
Where did "can't you just turn it off?" in the title come from? It doesn't appear anywhere in the actual title or the article, and I don't think it really aligns with its main assertions.
The retitling now by HN appears more accurate
It shows up at https://boydkane.com under the link "Why your boss isn't worried about advanced AI". Must be some kind of sub-heading, but not part of the actual article / blog post.
Presumably it's a phrase you might hear from a boss who sees AI as similar to (and as benign/known/deterministic as) most other software, per TFA
Ah, thanks for that!
>Presumably it's a phrase you might hear from a boss who sees AI as similar to (and as benign/known/deterministic as) most other software, per TFA
Yeah I get that, but I think that given the content of the article, "can't you just fix the code?" or the like would have been a better fit.
In my experience it’s usually the engineers that aren’t worried about AI, because they see the limitations clearly every time they use it. It’s pretty obvious that whole thing is severely overhyped and unreliable.
Your boss (or more likely, your bosses’ bosses’s boss) is the one deeply worried about it. Though mostly worried about being left behind by their competitors and how their company’s use of AI (or lack thereof) looks to shareholders.
It depends on where you are in the chain, and what kind of engineering you’re doing. I think a lot of engineers are so focused on the logistics, capabilities, and flaws, and so used to being indispensable, that they don’t viscerally get that they’re standing on the wrong side of the tree branch they’re sawing through. AI does not need to replace a single engineer before increased productivity means we’ll have way too many engineers, which mean jobs are impossible to get, and the salaries are in the shitter. Middle managers are terrified because they know they’re not long for this (career) world. Upper managers are having 3 champagne lunches because they see big bonuses on the far side of skyrocketing profits and cratering payroll costs.
It's a sci-fi thing, think of it along the lines of "What do you mean Skynet has gone rogue? Can't you just turn it off?"
(I think something along these lines was actually in the Terminator 3 movie, the one where Skynet goes live for the first time).
Agreed though, no relation to the actual post.
This sci-fi thing goes as far back as the 1983 movie WarGames, where they wanted to pull the plug on a rogue computer, but there was a reason you couldn’t do that:
McKittrick: General, the machine has locked us out. It's sending random numbers to the silos.
Pat Healy: Codes. To launch the missiles.
General Beringer: Just unplug the goddamn thing! Jesus Christ!
McKittrick: That won't work, General. It would interpret a shutdown as the destruction of NORAD. The computers in the silos would carry out their last instructions. They'd launch.
Further than that, even - this trope appears in Colossus: The Forbin Project, released in 1970, where the rogue computer is buried underground with its own nuclear reactor, so it can't be powered off.
In real life it won’t be that the computer prevents you from turning it off. It’ll be that the computer is guarded by cultists who think its god, and unstoppable market forces that require it to keep running.
When AI ends up running everything essential to survival and society, it’ll be preposterous to even suggest pulling the plug just because it does something bad.
Can you imagine the chaos of completely turning off GPS or Gmail today? Now imagine pulling the plug on something in the near future that controls all electric power distribution, banking communications, and Internet routing.
This is the case with capitalism today. I don't like where he took the philosophy, but Nick Land did have an insight that all the worst things we believe about AI (e.g. paperclip optimizing etc) are capitalism in a nutshell.
Just listen to what these CEOs say on the topic and they basically admit something terrible is being built, but that the most important things is that they are the ones to do it first.
It's a poor choice of phrase if the purpose is to illustrate a false equivalence. It applies to AI both as much (you can kill a process or stop a machine just the same regardless of whether it's running an LLM) and as little (you can't "turn off" Facebook any more than you can "turn off" ChatGPT) as it does to any other kind of software.
Turning AI off comes up a lot in existential risk discussions so I was surprised the article isn't about that.
> AIs will get more reliable over time, like old software is more reliable than new software.
:)
Was that a humam Freudian slip, or artificial one?
Yes, old software is often more reliable than new.
Holy survivorship bias, Batman.
If you think modern software is unreliable, let me introduce you to our friend, Rational Rose.
Agreed.
Or debuggers that would take out the entire OS.
Or a bad driver crashing everything multiple times a week.
Or a misbehaving process not handing control back to the OS.
I grew up in the era of 8 and 16 bit micros and early PCs, they where hilariously less stable than modern machines while doing far less, there wasn’t some halcyon age of near perfect software, it’s always been a case of things been good enough to be good enough but at least operating systems did improve.
Remember BSODs? Used to be a regular occurrence, now they're so infrequent they're gone from windows 11
And the "cooperative multitasking" in old operating systems where one program locking up meant the whole system was locked up
Gone? I had two last year, lets not overstate things.
Daily+ occurrences to two in a year pretty much rounds to zero. Kind of like we said measles were eradicated because there was <X per year cases.
My anecdata is that my current PC is four years old, with the same OS install, and I can't even recall if I've seen one BSoD.
I guess that is because you run it on old hardware. When I've bought my Asus ROG expensive laptop I had bsod almost daily. A year later with all updates I had bsod once in a month on the same device and windows installation.
If you have faulty hardware no amount of software is going to solve your problems (other than software that just completely deactivates said faulty hardware).
The fact you continued to have BSOD issues after a full reinstall is pretty strong evidence you probably had some kind of hardware failure.
But there was no reinstall in my case. Years goes by and further in time there are less and less bsod.
My point is if you are using the same "old" modern hardware, bsod is very rare.
Still get them fairly regularly except now they come with a QR code.
Depends, if you install games with anti-cheat they can often conflict and cause BSODs.
It's why I don't play the new trackmania.
Mostly because Microsoft shut down kernel access, wrote it's own generic drivers for "simple" devices (USBs, printers, sound cards, ...) and made "heavy" drivers submit to their WHQL quality control to be signed to run.
They're definitely not gone.
I remember Linux being remarkable reliable throughout its entire life in spite of being rabidly worked on.
Windows is only stabilizing because it's basically dead. All the activity is in the higher layers, where they are racking their brains on how to enshittify the experience, and extract value out of the remaining users.
As a Windows driver developer: LOL
I grew up in the same era and I recall crashes being less frequent.
There were plenty of other issues, including the fact that you had to adjust the right IRQ and DMA for your Sound Blaster manually, both physically and in each game, or that you needed to "optimize" memory usage, enable XMS or EMS or whatever it was at the time, or that you spent hours looking at the nice defrag/diskopt playing with your files, etc.
More generally, as you hint to, desktop operating systems were crap, but the software on top of it was much more comprehensively debugged. This was presumably a combination of two factors: you couldn't ship patches, so you had a strong incentive to debug it if you wanted to sell it, and software had way fewer features.
Come to think about it, early browsers kept crashing and taking down the entire OS, so maybe I'm looking at it with rosy glasses.
You are looking back with rosy glasses indeed.
Last year I assembled a retro PC (Pentium 2, Riva TNT 2 Ultra, Sound Blaster AWE64 Gold) running Windows 98 to remember my childhood, and it is more stable than what I remembered, but still way worse than modern systems. There are plenty of games that will refuse to work for whatever reason, or that will crash the whole OS, specially when existing, and require a hard reboot.
Oh and at least in the '90s you could already ship patches, we used to get them with the floppies and later CDs provided by magazines.
FWIW, I was speaking of the 80s.
It truly depends on the quality of the software you were using at the time. Maybe the software you used didn't result in many issues. I know a lot of the games I played as a kid on my family's or friend's Win95 machines resulted in system lockups or blue screens practically every time we used them.
As I mess around with these old machines for fun in my free time, I encounter these kinds of crashes pretty dang often. Its hard to tell if its just the old hardware is broken in odd ways or not so I can't fully say its the old software, but things are definitely pretty unreliable on old desktop Windows running old desktop Windows apps.
As an OS/2 and Linux users, I mostly missed out on Win95 fun.
But I was thinking of the (not particularly) golden days of MS-DOS/DR-DOS/Amiga/Atari applications.
But at the time that software was "new" and unreliable.
Or just http without the s. We take it for granted now, but not even that long ago http was the standard.
You know, I had spent a good amount of years not having even a single thought about rational rose, and now that’s all over.
I do apologize. I couldn't bear this burden alone.
It could definitely be worse. I have the privilege of using it weekly :(
What? How? I thought we stamped it out in the Purge of 2007.
Some things are forged in hell and refuse to die.
Ah, yes, the Oracle product model.
How much of that do you think would be attributable to IBM or Rational Software?
I know very little about Rational Rose, other than it always sounded like the stage name of a Vulcan stripper.
I think old in this sense is "released" rather than "beta" - it takes time to make any software reliable. Many of the examples here further prove that young software is unreliable.
Remember when debuggers were young?
Remember when OSes were young?
Remember when multi-tasking CPUs were young?
Etc...
You misunderstand. They are explicitly referring to the survivors that have been iterated on and chosen for being good.
They're NOT saying all software in the past was better.
At least that project was wise enough to use Lisp for storing its project files.
Rational Rhapsody called and wants the crown back
Neither, you’re reading it wrong. Think of it as codebases getting more reliable over time as they accumulate fixes and tests. (As opposed to, say, writing code in NodeJS versus C++)
Age of Code does not automatically equal quality of code, ever. Good code is maintained by good developers. A lot of bad code is pushed out by management, and other situations, or just bad devs. This is a can of worms you're talking your way into.
You're using different words - the top comment only mentioned the reliability of the software, which is only tangentially related to the quality, goodness, or badness of the code used to write it.
Old software is typically more reliable, not because the developers were better or the software engineering targeted a higher reliability metric, but because it's been tested in the real world for years. Even more so if you consider a known bug to be "reliable" behavior: "Sure, it crashes when you enter an apostrophe in the name field, but everyone knows that, there's a sticky note taped to the receptionist's monitor so the new girl doesn't forget."
Maybe the new software has a more comprehensive automated testing framework - maybe it simply has tests, where the old software had none - but regardless of how accurate you make your mock objects, decades of end-to-end testing in the real world is hard to replace.
As an industrial controls engineer, when I walk up to a machine that's 30 years old but isn't working anymore, I'm looking for failed mechanical components. Some switch is worn out, a cable got crushed, a bearing is failing...it's not the code's fault. It's not even the CMOS battery failing and dropping memory this time, because we've had that problem 4 times already, we recognize it and have a procedure to prevent it happening again. The code didn't change spontaneously, it's solved the business problem for decades... Conversely, when I walk up to a newly commissioned machine that's only been on the floor for a month, the problem is probably something that hasn't ever been tried before and was missed in the test procedure.
Yup, I have worked on several legacy codebases, and a pretty common occurence is that a new team member will join and think they may have discovered a bug in the code. Sometimes they are even quite adamant that the code is complete garbage and could never have worked properly. Usually the conversation goes something like: "This code is heavily used in production, and hasn't been touched in 10 years. If it's broken, then why haven't we had any complaints from users?"
And more often than not the issue is a local configuration issue, bad test data, a misunderstanding of what the code is supposed to do, not being aware of some alternate execution path or other pre/post processing that is running, some known issue that we've decided not to fix for some reason, etc. (And of course sometimes we do actually discover a completely new bug, but it's rare).
To be clear, there are certainly code quality issues present that make modifications to the code costly and risky. But the code itself is quite reliable, as most bugs have been found and fixed over the years. And a lot of the messy bits in the code are actually important usability enhancements that get bolted on after the fact in response to real-world user feedback.
Old code that has been maintained (bugfixed), but not messed with too much (i.e. major rewrites or new features) is almost certain to be better than most other code though?
"Bugfixes" doesn't mean the code actually got better, it just means someone attempted to fix a bug. I've seen plenty of people make code worse and more buggy by trying to fix a bug, and also plenty of old "maintained" code that still has tons of bugs because it started from the wrong foundation and everyone kept bolting on fixes around the bad part.
One of frustrating truths about software is that it can be terrible and riddled with bugs but if you just keep patching enough bugs and use it the same way every time it eventually becomes reliable software ... as long as the user never does anything new and no-one pokes the source with a stick.
I much prefer the alternative where it's written in a manner where you can almost prove it's bug free by comprehensively unit testing the parts.
I’ve read parts of macOS’ open source code that surely has been around for a while, maintained and absolute rubbish.
I think we all agree that the quality of the code itself goes down over time. I think the point that is being made is that the quality of the final product goes up over time.
E.g. you might fix a bug by adding a hacky workaround in the code; better product, worse code.
It actually might. Older code running in production is almost automatically regression tested with each new fix. It might not be pretty, but it's definitely more reliable for solving real problems.
The list of bugs tagged regression at work certainly suggests it gets tested... But fixing those regressions...? That's a lot of dev time for things that don't really have time allocated for them.
The author didn't mean that an older commit date on a file makes code better.
The author is talking about the maturity of a project. Likewise, as AI technologies become more mature we will have more tools to use them in a safer and more reliable way.
I've seen too many old projects that are not by any means better no matter how much they get updates because management define priorities. I'm not alone in saying I've been in a few projects where the backlog is rather large. When your development is driven by marketing people trying to pump up sales, all the "non critical" bugs begin to stack up.
Absolutely. Which is why the author clearly meant "old code" as in mature. Not "old code" as in "created a long time ago".
Sounds more like survivorship bias. All the bad codebases were thrown out and only the good ones lasted a long time.
Survivorship bias is real, but is missing the important piece of the story when it comes to software, which doesn't just survive but is also maintained. Sure you may choose to discard/replace low quality software and keep high quality software in operation, which leads to survivorship bias, but the point here is that you also have a chance to find and fix issues in the one that survived, even if those issues weren't yet apparent in version 0.1. Author is not trying to say that version 0.1 of 30 year old software was of higher quality than version 0.1 of modern software -- they're saying that version 9 of 30 year old software is better than version 0.1 of modern software.
In my experience actively maintained but not heavily modified applications tend towards stability over time. It don't even matter if they are good or bad codebases -- even a bad code will become less buggy over time if someone is working on bug fixes.
New code is the source of new bugs. Whether that's an entirely new product, a new feature on an existing project, or refactoring.
Basically the Lindy Effect: https://en.wikipedia.org/wiki/Lindy_effect
I’ve always called this “Work Hardening”, as in, the software has been improved over time by real work being done with it.
Ok, but metal that has been hardened is more prone to snapping once it loses its ductility
You mean think of it as opposite to what is written in the remark, and then find it funny?
Yes, I did that.
Perhaps better rephrased as "software that's been running for a (long) while is more reliable than software that only started running recently."
70 years ago we were fascinated by the concept of converting analog to a perfect digital copy. In reality, that goal was a pipe drea!m and the closest we can ever get is a near identical facimile to which data fits... But it's still quite easy to determine digital from true analog with rudimentary means.
Human thought is analog. It is based on chemical reactions, time, and unpredictably (effectively) random physical characteristics. AI is an attempt to turn that which is purely digital into an rational analog thought equivalent.
No matter how much effort, money, power, and rare mineral eating TPUs will - ever - produce true analog data.
It's been closer to 100 years since we figured out information theory and discredited this idea (that continuous/analog processes have more, or different, information in them than discrete/digital ones)
In theory or in practice? Wouldn't the Nyquist frequency and Heisenberg's uncertainty principle put practical limits.
This is all true. But digital audio and video media has captured essentially all economic value outside of live performance. So it seems likely that we will find a "good enough" in this domain too.
Interesting point with economic value extraction. The economy sacrificed accuracy and warmth of analog storage for convenience and security of digital storage. With economic incentive I am sure society will sacrifice accuracy and precision for the convenience of AI