Everyone is fighting for their view of the world right now and for those of us who do not have any investment in the success of one of the models, it feels like pure speculation at this point.
I don't think there's a human on this planet who can even predict the state of the industry in 3 years. In my entire time in the industry, I have always felt like I had a good line of sight three years away. Even when the iphone came on the scene, it felt like a generational increase rather than a revolution.
We just have no idea. We don't know the extent of how it can improve. We don't know if we are still on exponential improvement or the end of the S curve. We don't know what investment is going to be like. We don't know if there is autonomy in the future. We don't know if it's going to look more like the advancement of autonomous vehicles where everyone thought we were just a year or two away from full autonomy - or at least people bought the hype cycle.
Any anyone who says they know has something to sell you.
It’s exciting, isn’t it? I’ve been programming for a quarter century now and this is the first time in 20 years that the future of tech is exciting again and the world is software’s oyster.
Ha this is so true. All the major LLMs have a surprisingly ad free experience right now. Likely because they’re all aggressively competing for customers. But as soon as one of them realises the money that can be made from ads and how that will pay for daily operation costs after the investors give up investing, then the game is up
They're not ad-free. Ads are integrated into the content of the responses. For example, ChatGPT biases towards itself as a provider of services even when asking it not to. This behavior is easy to see and probably will be replicated out for paying clients.
Exciting is one word for it. The dot com crash was also "exciting", I suppose.
I see it as a chance for the capital class to sell everyone shovels and build railroads that will further cement their power and influence, all the while insisting software and art are more democratic than ever. All the while using the same tools to build surveillance infrastructure that will make any dissent impossible.
Is it though ? I feel like what ML can achieve is amazing, but, call me a pessimist, I'm bracing for the influx of elaborate scams, propaganda, deep fakes etc who will manage to drive a wedge even further in our society than social networks did. As for programming, my skills are not good enough not to risk an atrophy if I use code completion. I do ask Claude for feedback though. I've always been a tech enthusiast but this is the first time I want to build a cabin in the woods and live off-grid. The tech itself is amazing but I'm not looking forward to all the slop we're going to be flooded with.
Oh absolutely. If the history of the internet, social media, and smartphones is anything to go by, those negative consequences are coming like a freight train and this will all end in tears in a decade or two. Or even sooner, because that cycle seems to be accelerating too - education probably being the prime example.
But like the OP said, we can’t even predict what’s going to happen even three years out so I’ve just resigned myself to “going with the flow” and enjoying the ride as much as I can. If the negative consequences are coming for us, might as well get as much benefit as we can now while we’re all still wide eyed and bushy tailed.
Yea i think a lot of wars have come about from a change in communication technology (eg, radio, film), and can imagine AI might cause a lot of chaos until we learn how to deal with AI-generated video, audio, text, and images. Imagine if Goebbels or the Rwandan genocide had access to AI?
This isn‘t your main point, but the iPhone absolutely was the biggest revolution since the Internet. The world before it is wholly different to the world after it. AI looks to have a similar impact, but just like the iPhone it‘ll be a few years before everyone realizes the world has changed.
The iPhone opened up a whole new world of opportunities which were very clear from the start.
No one, except Steve Ballmer, would describe it as a potential fad or question how good it can actually get before Apple goes bankrupt from all the investment into this new tech.
I like this new stuff we get now, but the iPhone felt like a clear win with no downsides of a potential societal collapse.
We can make some educated guesses and extrapolate some trends. But you are right that most of the people currently claiming to know what's going to happen in 3 years were fast asleep when chat gpt launched, which is almost 3 years ago. And I include myself in that group. Most in the industry did not see that coming. Not even a little bit.
At this point we have half the industry being overly pessimistic and the other half being unreasonably optimistic. The median truth would be in the middle. But I don't think that's a very sound position to take either.
The reason is that I think we're actually dealing with a severe imagination deficit in society. That always happens around big technological changes. And this definitely looks and feels like such a thing. Ten years from now it might all seem obvious in retrospect. But right now we have the optimist camp predicting what boils down to the automotive equivalent of "faster horses" (AGI, I robot, self driving cars, and all the rest). It's going to be this wonderful utopia where no-one works and everything runs by itself. I'm not a big believer in that and I don't think that's how economies work.
And we have a bunch of pessimists predicting that it's all going to end in tears. Dystopia, everybody is going to be unemployed, and a lot of other Luddite nonsense.
The optimists basically lack imagination so they just reach for what science fiction told them is going to happen (i.e. rely on other people's science fiction). And then the pessimists basically are stuck imagining the worst always happens and failing to imagine that there might be things that actually do work.
It's fairly easy to predict/bet that both sides are probably imagining things wrong. Just like people did three years ago. Including myself here. So, not making a prediction here. But, kind of curious to see how the next few years will unfold. Lots of amazing stuff in the past three. I'll have some more of that please.
I’m not sure it was so much folks “not seeing it coming” with ChatGPT, only that anyone with anything beyond surface-level understanding of ML workloads and data science wouldn’t have thought to attempt to package such an immense, expensive, and potentially powerful product as a “chat bot”, and certainly nobody would have thought it would be a good idea to then heavily market it as some sci-fi notion of “artificial intelligence”, and proceed to gaslight the entire planet with what it might eventually be capable of. Only because, with the exception of Sam Altman, such a person would have been laughed out of a career. The only real surprise with LLMs seems to be that unscrupulous people have managed to raise so much goddamn money lying about its potential, that it’s got a real chance of crashing the global economy once everyone realizes it’s never actually gonna replace any meaningful jobs.
Just to echo another comment, I remember seeing the iPhone for the first time (2 maybe?). It was amazing. Mind blowingly good. It was revolutionary in every sense of the word. I'd heard the hype but if anything it undersold the reality.
You saw this thing and knew the future had changed.
It's very easy to forget Android just shamelessly ripped it off wholesale. The early version of Android looked nothing like the complete iPhone clone they ended up with.
And I have an android phone right now. But it only exists because of the revolutionary iPhone.
Credit where credit's due.
And right now AI feels like just before the iPhone. Like then we had touch screens, and we had new battery tech and we had mini-apps on the internet. And we had iPods showing small form factor MP3 players. But they didn't work together coherently.
And then the iPhone came along.
Like we're waiting for the iPhone of AI, to bring the tech together to an actually usable state.
The most important thing to do is play and build small things with all of them as they take turns being in the lead or running the fastest sideways in this horse race.
What's new to someone might be old to most etc. What's critical to the individual is knowing what's new to them vs what they know is out there.
I am not sure after 2 years of this if there will be a single clearcut winner, as much as they will all contribute best practices to something more emergent.
Webapps when they first began also didn't have much structure, frameworks, etc, and it evolved from just building a lot, and trying to build better.
Nice to see some balanced skepticism rather than the constant barrage of prognostication and bold opinions stated as fact. This kind of take doesn’t drive engagement though.
The trick I've always used in these circumstances is the cynical approach. That is assume nothing changes. If it does change, adopt late, rather than burn all your time, money and energy on churn and experimentation.
In the last 35 years of doing this, I've seen perhaps 10% of technology actually stick around for more than a few years. I'll adopt at maturity and discard when it's thoroughly obsolete.
Being fintech, no technology so far has fundamentally changed what business we do or even how it's done for a long time even if we pretend it does. A lot of the changes have just been a cost naively written off through arbitrary justification or keeping up with trends. 99% of what we do is CRUD, shit reports and batch processing, just like it was when it was S/390.
Even fewer things have had an ROI or a real customer benefit. Then again we have actual customers not investors.
You clearly are selling something, too: doubt. (Note your appeal to your own authority.)
Don't kid yourself. Skepticism is not neutrality. These days throwing shade is a growth industry. There's money to be made shorting just like there is going long. Neither is the objective, disinterested position, although skepticism always enjoys the appearance of prudence, at least to the ignorant.
Anyone who says they're not trying to sell you something lacks self-awareness about what they themselves have been sold.
I don't know, it seems like I'm the minority but I like Cursor. I think it adds value beyond the terminal style editors. Yes it relies on the Claude model but I get a lot of value from the visual component, history, auto-complete, etc.
Couldn't you make the same argument around something like S3? How many companies are basically S3 wrappers? Or companies that use general AWS infra and make it slightly better. There could still be a market for add on products. Why would Claude or OpenAI want the headache of managing an IDE? They're okay giving up some margin there.
I agree there is a huge rush of "AI wrapper" companies, whose moat is basically prompt engineering. Like a "AI buddy" or whatever. Those are all going to zero IMO. But things like Cursor have a future. Maybe not at the hyped valuation but long term something like this will exist
I’d love for someone to try to define “AI wrapper”.
I’m trying to imagine a graph where at some point in time t
the status of a company changes from “wrapper” (not enough “original” engineering)
to “proper company” (they own the IP, and they fought for it!!!)
At what point did OpenAI cease being an NVIDIA wrapper and become the world’s leading AI lab? At what point did NVIDIA graduate from being a TSMC wrapper?
Clearly any company that gets TSMC N2 node allocation is going to win, the actual details of the chip don’t matter super much.
'Wrapper' in this instance is your primary source of value is a prompt.
I think you can think of it as how long would it take someone to come up with the product given enough information about the product.
Take for instance an app that is a "companion" app. It's simplest form is a prompt + LLM + interface. They don't own the LLM so they have the prompt and interface. The prompt is simple enough to figure out (often by asking the app in a clever way) so the interface is left. How easy is it to replicate? If it's like chatgpt, pretty easy.
Now there are a few complications. Suppose there are network effects (Instagram is a wrapper around a protocol), but the network effects are the value. And LLM wrapper can create network effects (maybe there is a way to share or something) but difficult.
OpenAI is not a wrapper on NVIDIA because it would take billions of dollars to train the LLM with the NVIDIA chips (in energy). It would take me a weekend to recreate a GPT wrapper or just fork an open source implementation. There is also institutional knowledge (which is why Meta is offering 1bn+ for a single eng). Or take something like Excel. People know how it works, people have dissected it endlessly. But the cost to recreate even with perfect knowledge is very high, plus there is network effects.
Taking a shot at this, one concrete definition might be that the business model is essentially white labeling, that is, the base LLM is rebranded, but task performance in the problem domain is not functionally improved in some measurable way. As a corollary, it means the user could receive the same value if they had gone straight to the base LLM provider.
I think this might be more narrow than most uses of the term “wrapper” though.
It's not really the same because the provider in this case isn't necessarily shipping a traditional service, they're shipping intelligence. We've confused APIs as the end-state for providers. Providers are going to eat every abstraction along the way in their delivery of intelligent capabilities. Claude Code is just the start. A true agentic intelligent capability that shifts a paradigm for ways of working. It will evolve into Claude Agent for general-purpose digital work.
There's a lot of talk around economics. What is going to be more economic than a provider building abstractions/margin-optimizations around the tokens, and shipping directly to consumer. Vs token arbitrage.
Lastly, there's a lot of industry hype and narrative around agents. In my opinion, Claude Code is really the only effective & actual agent; the first born. It shows that Anthropic is a signaling that the leading providers will no longer just train models. They are creating intelligent capabilities within the post training phases / in RL. They are shipping the brain and the mech suit for it. Hence, eat the stack. From terminal to desktop, eventual robotics.
> Providers are going to eat every abstraction along the way in their delivery of intelligent capabilities. [...] There's a lot of talk around economics. What is going to be more economic than a provider building abstractions/margin-optimizations around the tokens, and shipping directly to consumer. Vs token arbitrage.
The strongman counter-argument would be that specialized interfaces to AI will always require substantial amounts of work to create and maintain.
If true, then similar to Microsoft, it might make more financial sense for Anthropic et al. to cede those specialized markets to others, focus on their core platform product, take a cut from many different specialized products, and end up making more as the addressable market broadens.
The major AI model providers substantially investing in specialized interfaces would suggest they're pessimistic about revolutionary core model improvements and are thus looking to vertically integration to preserve margin / moat.
But relatively speaking, it doesn't seem like interfaces are being inordinately invested in, and coding seems such an obvious agentic target (and dogfoodable learning opportunity!) that it shouldn't prompt tea leaf reading.
> It shows that Anthropic is a signaling that the leading providers will no longer just train models.
I think it instead (or also?) shows a related but orthogonal signal: that the ability and resources to train models are a strong competitive advantage. This is most obvious with deep research and I haven’t seen any wrapper or open source project achieve anywhere near the same quality as Gemini/Claude deep research, but Claude Code is a close runner up.
I've used Claude Code but not the vs code plugin. I get enough value from the auto-complete that I'll use Cursor regardless, but I don't think it's worth $20 for that.
But now it's subsidized so I easily spend over $50 of Claude credits for my $20 in Cursor.
Also the ability to swap out models is a big value add and I don't have to worry about latest and greatest. I switch seamlessly. Something comes out, next day its on Claude. So now I'm using GPT which is less than half the price. I don't want to have to think about it or constantly consider other options. I want a standardized interface and plug in whatever intelligence I want. Kind of like a dropbox that can worry about whether they store in AWS, Azure or GCP depending which one is the best value prop.
The OP postulates two paths for the future: "1: LLM Labs go direct" and "2: LLMs become commodities, wrappers win". I just happened to have published a blogpost on llm code assistants used by GitHub repos[0][1]. Claude Code has just overtaken Cursor as the most popular. Gemini CLI and OpenAI Codex also has a steeper growth curve than Cursor. So on just this question, it looks like the drugs are beating the dealers.
I know this isn’t the point of the article, but that’s not how the drug trade works at all. The producers make very little. The real money is in the logistics. Which you could argue is the cloud providers but the analogy isn’t great anyway.
Yes. For cocaine, for example, they have large farmer networks that cultivate coca, harvest and process it lightly (drying, etc.). The cartels then finish the process and undertake the logistics of shipping it across borders, into the US, or across the Atlantic into Europe.
This logistical leg is where most of the work is done since you have to:
Maintain large slush funds for bribing law enforcement.
Run workshops and technicians that strip civilians cars, embed cocaine in the nooks and hand them over to American civilians to drive over the border.
Hire engineers from Pakistani universities to build narco-submarines in riverine deltas, which are then used to cross the Atlantic for European supplies.
Maintain contact with your African coastal syndicates who have another trans-Saharan route for getting drugs into Europe.
Run payroll for your workforce (this is a business after all).
Maintain a decently trained fighting force to slaughter enemies that encroach on your turf. Or informants, uncompromising cops, politicians, etc. This includes training, paying, initiating them, and hiring good experienced fighters. Right now, it's credibly reported that Mexican cartels are volunteering to fight in Ukraine to gain experience with drones and other UAVs to expand their war-making capabilities.
Hire chemistry undergrads from local STEM universities to turn synthetic precursors from Asia into fentanyl, etc.
So, just like African cocoa farmers and American growers see just a tiny slice of the profit the end-products produce, the cartels are in the logistics & firepower business; they've outsourced a huge chunk of the business to growers, just like their peers in the chocolate and grocery business.
Imagine you're studying, for example math, and you have to learn calculus, and you're living in a dormitory with a colleague, who's great at integration calculus, and every time you get another integration function puzzle to solve, this guy (or gal) pops up and says "it's 2*e^2/x+C dx", "it's e^2+C dx", "it's sin(x)/cos(x)+C dx" after 5 milliseconds, even before you fully comprehend the function...
And yes, I memorized a bunch of integrals for my calculus class and then promptly forgot them all after the final exam. It's not worth it to remember how to do integrals other than simple cases like polynomials or sin/cos on their own.
If calculators had a camera on the back that let you take a photo of a word problem on your homework or exam and then provided a complete answer including steps and "work shown" I expect that would have a detrimental effect on students' learning of the subject matter and ability to reproduce it without the calculator.
That's a silly argument, because arithmetic is so simple people pick it up in spite of their use of calculators. Hell, people learn simple arithmetic even if they've never had any schooling (they can tell you that 5 apples and 3 apples makes 8 apples, things like this). And arithmetic with big numbers is so tedious regardless of how well you know it that it's better to just use a calculator, and people have always done this (before digital calculators you had an abacus, multiplication tables, and many other similar instruments).
Calculus or programming or advanced algebra etc are nowhere near the same difficulty, and the same rules don't apply.
The article's headline is great and it delivers the message clearly. But the whole premise to support the message is very assumptive - "there is no moat". There is a lot moat in cursor, bolt, lovable. The same way there is moat in the chat apps of openai, anthropic, gemini...
They say there is no moat, but in fact, a feature in anthropic takes a good few months up to a year to appear on openai chatapp, and the same is true vice versa.
You could say some of those issues are solvable by allocating more money, and resources, which might be true, and it could be true that it would be beneficial for openai to develop their own cursor platform in the future, to get better margins. But in reality, who knows when that future would come? Maybe by then cursor will have much more moat and entering the market would be much more difficult. Maybe openai will continue developing their core product and entering other domains will not be worth the effort.
Currently, LLMs as a product have not been solved. All companies operate at a lose in order to rise the top, and we still don't know how it will be monetized in the future. But as it stands - there is already moat, moat in infrastructure - even though a few years ago they said that llms have no moat, now there is already a strong set of features and "agents" that deliver us the deep reasoning, online searching, and multimodal experience.
So, there is moat. But moat can accumulate over time. For the article to be true - it should prove the the current moat is low, and it can not accumulate.
I think this is up to the user. I actually found tab so annoying that it was a big reason I quit cursor and cancelled my sub. I couldn't think straight with it constantly suggesting things to put in after every key stroke and caused a few annoying bugs for me.
I find pure claude and neovim to be a great pair. I set up custom vim commands to make sharing file paths, line numbers, and code super easy. that way I can move quickly through code for manual developing as well as have claude right there with the context it needs quickly.
I know the submission parent of the discussion I'm gonna link is not for everyone and might be considered a "rant", but the subject immediately reminds me of this:
Specifically the product value compared to the operating cost.
Now, if the tool (Claude Code) really is very valuable, and Cursor is just a very good integration, and they manage to guard their moat (brand, subscription, glue code), maybe there's something to it.
I'm not a businessperson, like I said, just immediately reminded me of that post I read on the weekend.
super interesting breakdown. that being said, it's unclear to me if this is actually a problem outside of code gen. the labs have zeroed in on this use case since it's so obviously valuable but they're not going to launching products in every area.
also, yes, the labs control the supply but also there are many labs so there's lots of competition. they can't, for example, just jack up the prices on the dealers (apps) like a monopoly could. so again, not sure if being a dealer is actually bad here.
Can anyone tell me their experience with Cursor vs GitHub Copilot? I use GitHub Copilot Pro right now through Visual Studio Code, and tried Cursor, but Cursor just seemed like a more expensive GitHub Copilot Pro.
Like, I'm publishing https://github.com/andrewmcwattersandco/git-fetch-file right now with Claude Sonnet 4 (thank you for recently upvoting that to the front page). And the whole repository view that GitHub Copilot and Claude Sonnet 4 have on my projects seems like the same exact thing you get in Cursor, but Cursor for some reason took longer with the exact same models, and I'm not sure why.
Maybe they prompt the models differently? I haven't taken a look.
Also, Cursor seems to be literally a Visual Studio Code fork! But everyone's talking about it lately, and no one is mentioning this. I don't understand.
not exactly what you asked, but i've tried out Junie[0] because I've already got JetBrainz IDEs set up and love them.
It's terrible. For comparison, I've only used cursor on greenfield toy projects, but cursor is way better at the agentic stuff (the actual code generation AND the "review these changes" workflow) AND the tab/auto-complete stuff.
I hope Junie can make some leaps because I really like JetBrainz and dont want to see them fall behind
GitHub Copilot by itself is not directly comparable to Cursor. For example I use Zed + Copilot + Claude together at work for a similar workflow to Cursor.
This is where it starts to get subjective, but changing one part of the toolkit can have a huge effect on the quality of the assistant. For example I tried GPT 4.1 instead of Claude 4 recently and it took my setup from improving my productivity by 3-5x on coding tasks to, like, 0.5x. I can't point to a specific change other than tasks went from being done in 5 minutes if back and forth to being only partly done in 15 minutes.
I haven't used VSC in a year or Cursor at all, but I hear similar things from colleagues.
I do a ridiculous amount of dumb shit with cursor without paying extra for it and I can't be the only one. There is just no way they will ever make a profit with their current all-you-can-eat model. Like there have been times I have just copy and pasted code into a cursor window for the AI to add when I could have just pasted it myself just because it makes the context easier to deal with.
However, they will eventually get purchased by an AI company because the _product_ is great.
At least in Copilot, the revert changes is more like undo/redo of a series of changes. Often I want to keep all of the changes except for the most recent one
git checkout would destroy this (and "corrupts" the Copilot session state)
> However, they will eventually get purchased by an AI company because the _product_ is great.
'Great' is in the eye of the beholder. For me, Cursor was one of the least-effective solutions of the many options (from Cursor and other AIDEs, to repo-centric web-based options like Jules, to CLI-based options like Claude Code) I evaluated a few months ago.
Everyone is fighting for their view of the world right now and for those of us who do not have any investment in the success of one of the models, it feels like pure speculation at this point.
I don't think there's a human on this planet who can even predict the state of the industry in 3 years. In my entire time in the industry, I have always felt like I had a good line of sight three years away. Even when the iphone came on the scene, it felt like a generational increase rather than a revolution.
We just have no idea. We don't know the extent of how it can improve. We don't know if we are still on exponential improvement or the end of the S curve. We don't know what investment is going to be like. We don't know if there is autonomy in the future. We don't know if it's going to look more like the advancement of autonomous vehicles where everyone thought we were just a year or two away from full autonomy - or at least people bought the hype cycle.
Any anyone who says they know has something to sell you.
It’s exciting, isn’t it? I’ve been programming for a quarter century now and this is the first time in 20 years that the future of tech is exciting again and the world is software’s oyster.
Can’t wait for AI 2.0 and ads :(
> and ads
Ha this is so true. All the major LLMs have a surprisingly ad free experience right now. Likely because they’re all aggressively competing for customers. But as soon as one of them realises the money that can be made from ads and how that will pay for daily operation costs after the investors give up investing, then the game is up
They're not ad-free. Ads are integrated into the content of the responses. For example, ChatGPT biases towards itself as a provider of services even when asking it not to. This behavior is easy to see and probably will be replicated out for paying clients.
Yes for sure companies are already paying for having their products added to the training set of popular LLMs.
Exciting is one word for it. The dot com crash was also "exciting", I suppose.
I see it as a chance for the capital class to sell everyone shovels and build railroads that will further cement their power and influence, all the while insisting software and art are more democratic than ever. All the while using the same tools to build surveillance infrastructure that will make any dissent impossible.
So yeah, exciting is one word you could choose.
Is it though ? I feel like what ML can achieve is amazing, but, call me a pessimist, I'm bracing for the influx of elaborate scams, propaganda, deep fakes etc who will manage to drive a wedge even further in our society than social networks did. As for programming, my skills are not good enough not to risk an atrophy if I use code completion. I do ask Claude for feedback though. I've always been a tech enthusiast but this is the first time I want to build a cabin in the woods and live off-grid. The tech itself is amazing but I'm not looking forward to all the slop we're going to be flooded with.
Oh absolutely. If the history of the internet, social media, and smartphones is anything to go by, those negative consequences are coming like a freight train and this will all end in tears in a decade or two. Or even sooner, because that cycle seems to be accelerating too - education probably being the prime example.
But like the OP said, we can’t even predict what’s going to happen even three years out so I’ve just resigned myself to “going with the flow” and enjoying the ride as much as I can. If the negative consequences are coming for us, might as well get as much benefit as we can now while we’re all still wide eyed and bushy tailed.
Yea i think a lot of wars have come about from a change in communication technology (eg, radio, film), and can imagine AI might cause a lot of chaos until we learn how to deal with AI-generated video, audio, text, and images. Imagine if Goebbels or the Rwandan genocide had access to AI?
As exciting as nuclear armageddon, sure.
This isn‘t your main point, but the iPhone absolutely was the biggest revolution since the Internet. The world before it is wholly different to the world after it. AI looks to have a similar impact, but just like the iPhone it‘ll be a few years before everyone realizes the world has changed.
The iPhone opened up a whole new world of opportunities which were very clear from the start.
No one, except Steve Ballmer, would describe it as a potential fad or question how good it can actually get before Apple goes bankrupt from all the investment into this new tech.
I like this new stuff we get now, but the iPhone felt like a clear win with no downsides of a potential societal collapse.
Mobile phone addiction is our generation's smoking. We just don't realize it yet.
Good point.
Pretty sure it's known. But, just like smoking, it's tolerated. Can't make the line go down.
It wasn’t on day one though.
We can make some educated guesses and extrapolate some trends. But you are right that most of the people currently claiming to know what's going to happen in 3 years were fast asleep when chat gpt launched, which is almost 3 years ago. And I include myself in that group. Most in the industry did not see that coming. Not even a little bit.
At this point we have half the industry being overly pessimistic and the other half being unreasonably optimistic. The median truth would be in the middle. But I don't think that's a very sound position to take either.
The reason is that I think we're actually dealing with a severe imagination deficit in society. That always happens around big technological changes. And this definitely looks and feels like such a thing. Ten years from now it might all seem obvious in retrospect. But right now we have the optimist camp predicting what boils down to the automotive equivalent of "faster horses" (AGI, I robot, self driving cars, and all the rest). It's going to be this wonderful utopia where no-one works and everything runs by itself. I'm not a big believer in that and I don't think that's how economies work.
And we have a bunch of pessimists predicting that it's all going to end in tears. Dystopia, everybody is going to be unemployed, and a lot of other Luddite nonsense.
The optimists basically lack imagination so they just reach for what science fiction told them is going to happen (i.e. rely on other people's science fiction). And then the pessimists basically are stuck imagining the worst always happens and failing to imagine that there might be things that actually do work.
It's fairly easy to predict/bet that both sides are probably imagining things wrong. Just like people did three years ago. Including myself here. So, not making a prediction here. But, kind of curious to see how the next few years will unfold. Lots of amazing stuff in the past three. I'll have some more of that please.
I’m not sure it was so much folks “not seeing it coming” with ChatGPT, only that anyone with anything beyond surface-level understanding of ML workloads and data science wouldn’t have thought to attempt to package such an immense, expensive, and potentially powerful product as a “chat bot”, and certainly nobody would have thought it would be a good idea to then heavily market it as some sci-fi notion of “artificial intelligence”, and proceed to gaslight the entire planet with what it might eventually be capable of. Only because, with the exception of Sam Altman, such a person would have been laughed out of a career. The only real surprise with LLMs seems to be that unscrupulous people have managed to raise so much goddamn money lying about its potential, that it’s got a real chance of crashing the global economy once everyone realizes it’s never actually gonna replace any meaningful jobs.
I really like this. It feels like the only inevitability is change. Change, to some extent, not yet known.
Just to echo another comment, I remember seeing the iPhone for the first time (2 maybe?). It was amazing. Mind blowingly good. It was revolutionary in every sense of the word. I'd heard the hype but if anything it undersold the reality.
You saw this thing and knew the future had changed.
It's very easy to forget Android just shamelessly ripped it off wholesale. The early version of Android looked nothing like the complete iPhone clone they ended up with.
And I have an android phone right now. But it only exists because of the revolutionary iPhone.
Credit where credit's due.
And right now AI feels like just before the iPhone. Like then we had touch screens, and we had new battery tech and we had mini-apps on the internet. And we had iPods showing small form factor MP3 players. But they didn't work together coherently.
And then the iPhone came along.
Like we're waiting for the iPhone of AI, to bring the tech together to an actually usable state.
It's understandable how it feels this way.
The most important thing to do is play and build small things with all of them as they take turns being in the lead or running the fastest sideways in this horse race.
What's new to someone might be old to most etc. What's critical to the individual is knowing what's new to them vs what they know is out there.
I am not sure after 2 years of this if there will be a single clearcut winner, as much as they will all contribute best practices to something more emergent.
Webapps when they first began also didn't have much structure, frameworks, etc, and it evolved from just building a lot, and trying to build better.
Nice to see some balanced skepticism rather than the constant barrage of prognostication and bold opinions stated as fact. This kind of take doesn’t drive engagement though.
Agreed.
The trick I've always used in these circumstances is the cynical approach. That is assume nothing changes. If it does change, adopt late, rather than burn all your time, money and energy on churn and experimentation.
In the last 35 years of doing this, I've seen perhaps 10% of technology actually stick around for more than a few years. I'll adopt at maturity and discard when it's thoroughly obsolete.
Being fintech, no technology so far has fundamentally changed what business we do or even how it's done for a long time even if we pretend it does. A lot of the changes have just been a cost naively written off through arbitrary justification or keeping up with trends. 99% of what we do is CRUD, shit reports and batch processing, just like it was when it was S/390.
Even fewer things have had an ROI or a real customer benefit. Then again we have actual customers not investors.
You clearly are selling something, too: doubt. (Note your appeal to your own authority.)
Don't kid yourself. Skepticism is not neutrality. These days throwing shade is a growth industry. There's money to be made shorting just like there is going long. Neither is the objective, disinterested position, although skepticism always enjoys the appearance of prudence, at least to the ignorant.
Anyone who says they're not trying to sell you something lacks self-awareness about what they themselves have been sold.
I don't know, it seems like I'm the minority but I like Cursor. I think it adds value beyond the terminal style editors. Yes it relies on the Claude model but I get a lot of value from the visual component, history, auto-complete, etc.
Couldn't you make the same argument around something like S3? How many companies are basically S3 wrappers? Or companies that use general AWS infra and make it slightly better. There could still be a market for add on products. Why would Claude or OpenAI want the headache of managing an IDE? They're okay giving up some margin there.
I agree there is a huge rush of "AI wrapper" companies, whose moat is basically prompt engineering. Like a "AI buddy" or whatever. Those are all going to zero IMO. But things like Cursor have a future. Maybe not at the hyped valuation but long term something like this will exist
I’d love for someone to try to define “AI wrapper”.
I’m trying to imagine a graph where at some point in time t
the status of a company changes from “wrapper” (not enough “original” engineering)
to “proper company” (they own the IP, and they fought for it!!!)
At what point did OpenAI cease being an NVIDIA wrapper and become the world’s leading AI lab? At what point did NVIDIA graduate from being a TSMC wrapper?
Clearly any company that gets TSMC N2 node allocation is going to win, the actual details of the chip don’t matter super much.
'Wrapper' in this instance is your primary source of value is a prompt.
I think you can think of it as how long would it take someone to come up with the product given enough information about the product.
Take for instance an app that is a "companion" app. It's simplest form is a prompt + LLM + interface. They don't own the LLM so they have the prompt and interface. The prompt is simple enough to figure out (often by asking the app in a clever way) so the interface is left. How easy is it to replicate? If it's like chatgpt, pretty easy.
Now there are a few complications. Suppose there are network effects (Instagram is a wrapper around a protocol), but the network effects are the value. And LLM wrapper can create network effects (maybe there is a way to share or something) but difficult.
OpenAI is not a wrapper on NVIDIA because it would take billions of dollars to train the LLM with the NVIDIA chips (in energy). It would take me a weekend to recreate a GPT wrapper or just fork an open source implementation. There is also institutional knowledge (which is why Meta is offering 1bn+ for a single eng). Or take something like Excel. People know how it works, people have dissected it endlessly. But the cost to recreate even with perfect knowledge is very high, plus there is network effects.
Taking a shot at this, one concrete definition might be that the business model is essentially white labeling, that is, the base LLM is rebranded, but task performance in the problem domain is not functionally improved in some measurable way. As a corollary, it means the user could receive the same value if they had gone straight to the base LLM provider.
I think this might be more narrow than most uses of the term “wrapper” though.
Every company you mentioned is just a wrapper around elemental carbon and silicon.
Claude Code with current IDE integration is already very good. Only thing missing is completion that Cursor is pretty good at.
For me VScode with Github Copilot + Claude Code hits the sweet spot
I'm super excited to try the Cursor CLI https://cursor.com/cli
Claude Code has always been unparalleled. It's almost as if other AI CLI devs have no idea what they're doing.
It's not really the same because the provider in this case isn't necessarily shipping a traditional service, they're shipping intelligence. We've confused APIs as the end-state for providers. Providers are going to eat every abstraction along the way in their delivery of intelligent capabilities. Claude Code is just the start. A true agentic intelligent capability that shifts a paradigm for ways of working. It will evolve into Claude Agent for general-purpose digital work.
There's a lot of talk around economics. What is going to be more economic than a provider building abstractions/margin-optimizations around the tokens, and shipping directly to consumer. Vs token arbitrage.
Lastly, there's a lot of industry hype and narrative around agents. In my opinion, Claude Code is really the only effective & actual agent; the first born. It shows that Anthropic is a signaling that the leading providers will no longer just train models. They are creating intelligent capabilities within the post training phases / in RL. They are shipping the brain and the mech suit for it. Hence, eat the stack. From terminal to desktop, eventual robotics.
> Providers are going to eat every abstraction along the way in their delivery of intelligent capabilities. [...] There's a lot of talk around economics. What is going to be more economic than a provider building abstractions/margin-optimizations around the tokens, and shipping directly to consumer. Vs token arbitrage.
The strongman counter-argument would be that specialized interfaces to AI will always require substantial amounts of work to create and maintain.
If true, then similar to Microsoft, it might make more financial sense for Anthropic et al. to cede those specialized markets to others, focus on their core platform product, take a cut from many different specialized products, and end up making more as the addressable market broadens.
The major AI model providers substantially investing in specialized interfaces would suggest they're pessimistic about revolutionary core model improvements and are thus looking to vertically integration to preserve margin / moat.
But relatively speaking, it doesn't seem like interfaces are being inordinately invested in, and coding seems such an obvious agentic target (and dogfoodable learning opportunity!) that it shouldn't prompt tea leaf reading.
> It shows that Anthropic is a signaling that the leading providers will no longer just train models.
I think it instead (or also?) shows a related but orthogonal signal: that the ability and resources to train models are a strong competitive advantage. This is most obvious with deep research and I haven’t seen any wrapper or open source project achieve anywhere near the same quality as Gemini/Claude deep research, but Claude Code is a close runner up.
Have you tried Claude code vscode plugin? It's has almost everything cursor has to offer
I've used Claude Code but not the vs code plugin. I get enough value from the auto-complete that I'll use Cursor regardless, but I don't think it's worth $20 for that.
But now it's subsidized so I easily spend over $50 of Claude credits for my $20 in Cursor.
Also the ability to swap out models is a big value add and I don't have to worry about latest and greatest. I switch seamlessly. Something comes out, next day its on Claude. So now I'm using GPT which is less than half the price. I don't want to have to think about it or constantly consider other options. I want a standardized interface and plug in whatever intelligence I want. Kind of like a dropbox that can worry about whether they store in AWS, Azure or GCP depending which one is the best value prop.
The crazy thing is that yes Cursor are losing money reselling ChatGPT / Claude - but Claude and OpenAI are ALSO losing money!
The OP postulates two paths for the future: "1: LLM Labs go direct" and "2: LLMs become commodities, wrappers win". I just happened to have published a blogpost on llm code assistants used by GitHub repos[0][1]. Claude Code has just overtaken Cursor as the most popular. Gemini CLI and OpenAI Codex also has a steeper growth curve than Cursor. So on just this question, it looks like the drugs are beating the dealers.
[0] https://aleyan.com/blog/2025-llm-assistant-census/ [1] https://news.ycombinator.com/item?id=44863713
I can’t read things written in pure hyperbole like this.
“Netflix isn’t a real company, they are nothing more than shills for camera-workers who film all of the content upstream.”
Such a lame take. I’d be more interested if someone actually took a minute to think about the competitive landscape more seriously than this.
Hasn't the proliferation of streaming services shown that the content is the "drug"? To remain competitive, they had to start producing themselves.
I know this isn’t the point of the article, but that’s not how the drug trade works at all. The producers make very little. The real money is in the logistics. Which you could argue is the cloud providers but the analogy isn’t great anyway.
This is new to me. Are you saying the people producing the drugs make very little?
Yes, the average heroin farmer is making more money than a wheat farmer, but he sells it to someone who can get it to America.
Yes. For cocaine, for example, they have large farmer networks that cultivate coca, harvest and process it lightly (drying, etc.). The cartels then finish the process and undertake the logistics of shipping it across borders, into the US, or across the Atlantic into Europe.
This logistical leg is where most of the work is done since you have to:
Maintain large slush funds for bribing law enforcement.
Run workshops and technicians that strip civilians cars, embed cocaine in the nooks and hand them over to American civilians to drive over the border.
Hire engineers from Pakistani universities to build narco-submarines in riverine deltas, which are then used to cross the Atlantic for European supplies.
Maintain contact with your African coastal syndicates who have another trans-Saharan route for getting drugs into Europe.
Run payroll for your workforce (this is a business after all).
Maintain a decently trained fighting force to slaughter enemies that encroach on your turf. Or informants, uncompromising cops, politicians, etc. This includes training, paying, initiating them, and hiring good experienced fighters. Right now, it's credibly reported that Mexican cartels are volunteering to fight in Ukraine to gain experience with drones and other UAVs to expand their war-making capabilities.
Hire chemistry undergrads from local STEM universities to turn synthetic precursors from Asia into fentanyl, etc.
So, just like African cocoa farmers and American growers see just a tiny slice of the profit the end-products produce, the cartels are in the logistics & firepower business; they've outsourced a huge chunk of the business to growers, just like their peers in the chocolate and grocery business.
yeah, this makes no sense. Don't the cartels control all aspects from production to distribution?
Yes, but the distribution (of a highly addictive product) is what sells. Not the raw ingredients or manufacturing, per se.
Imagine you're studying, for example math, and you have to learn calculus, and you're living in a dormitory with a colleague, who's great at integration calculus, and every time you get another integration function puzzle to solve, this guy (or gal) pops up and says "it's 2*e^2/x+C dx", "it's e^2+C dx", "it's sin(x)/cos(x)+C dx" after 5 milliseconds, even before you fully comprehend the function...
I have a question for you
They can already do this just by downloading Mathematica.
Solving integrals is the original selling point of Mathematica: https://www.wolfram.com/mathematica/scrapbook/1988/03/04/198...
And yes, I memorized a bunch of integrals for my calculus class and then promptly forgot them all after the final exam. It's not worth it to remember how to do integrals other than simple cases like polynomials or sin/cos on their own.
Do you also oppose calculators for the same reason?
If calculators had a camera on the back that let you take a photo of a word problem on your homework or exam and then provided a complete answer including steps and "work shown" I expect that would have a detrimental effect on students' learning of the subject matter and ability to reproduce it without the calculator.
That's a silly argument, because arithmetic is so simple people pick it up in spite of their use of calculators. Hell, people learn simple arithmetic even if they've never had any schooling (they can tell you that 5 apples and 3 apples makes 8 apples, things like this). And arithmetic with big numbers is so tedious regardless of how well you know it that it's better to just use a calculator, and people have always done this (before digital calculators you had an abacus, multiplication tables, and many other similar instruments).
Calculus or programming or advanced algebra etc are nowhere near the same difficulty, and the same rules don't apply.
I'm a high school dropout dude...
The article's headline is great and it delivers the message clearly. But the whole premise to support the message is very assumptive - "there is no moat". There is a lot moat in cursor, bolt, lovable. The same way there is moat in the chat apps of openai, anthropic, gemini...
They say there is no moat, but in fact, a feature in anthropic takes a good few months up to a year to appear on openai chatapp, and the same is true vice versa.
You could say some of those issues are solvable by allocating more money, and resources, which might be true, and it could be true that it would be beneficial for openai to develop their own cursor platform in the future, to get better margins. But in reality, who knows when that future would come? Maybe by then cursor will have much more moat and entering the market would be much more difficult. Maybe openai will continue developing their core product and entering other domains will not be worth the effort.
Currently, LLMs as a product have not been solved. All companies operate at a lose in order to rise the top, and we still don't know how it will be monetized in the future. But as it stands - there is already moat, moat in infrastructure - even though a few years ago they said that llms have no moat, now there is already a strong set of features and "agents" that deliver us the deep reasoning, online searching, and multimodal experience.
So, there is moat. But moat can accumulate over time. For the article to be true - it should prove the the current moat is low, and it can not accumulate.
I think Cursor tab-completion is entirely in-house, right? That feature on its own is worth at least $5/month, it's super well done.
I think this is up to the user. I actually found tab so annoying that it was a big reason I quit cursor and cancelled my sub. I couldn't think straight with it constantly suggesting things to put in after every key stroke and caused a few annoying bugs for me.
I find pure claude and neovim to be a great pair. I set up custom vim commands to make sharing file paths, line numbers, and code super easy. that way I can move quickly through code for manual developing as well as have claude right there with the context it needs quickly.
I’m paying $20/m just for tab, and willing to pay $40/m just to have it in Rider so I can return back to using single IDE.
Doesn't Rider have JetBraims AI? It's basically the same thing as Cursor.
I agree, their tab completion is magical.
And dare I say their only remaining moat.
JetBrains IDEs also have that.
I know the submission parent of the discussion I'm gonna link is not for everyone and might be considered a "rant", but the subject immediately reminds me of this:
https://news.ycombinator.com/item?id=44424456
Make Fun of Them (42 days ago, 43 comments)
Specifically the product value compared to the operating cost.
Now, if the tool (Claude Code) really is very valuable, and Cursor is just a very good integration, and they manage to guard their moat (brand, subscription, glue code), maybe there's something to it.
I'm not a businessperson, like I said, just immediately reminded me of that post I read on the weekend.
super interesting breakdown. that being said, it's unclear to me if this is actually a problem outside of code gen. the labs have zeroed in on this use case since it's so obviously valuable but they're not going to launching products in every area.
also, yes, the labs control the supply but also there are many labs so there's lots of competition. they can't, for example, just jack up the prices on the dealers (apps) like a monopoly could. so again, not sure if being a dealer is actually bad here.
for me, I wasted the whole day fighting LLMs to in the end write a single test function for authz function...
LLMs are a huge waste of time
Don't worry they will offer you Rivermind Plus soon to solve all the problems
One could have argued the same for Stack Overflow being the drug and Google the dealer too...
Can anyone tell me their experience with Cursor vs GitHub Copilot? I use GitHub Copilot Pro right now through Visual Studio Code, and tried Cursor, but Cursor just seemed like a more expensive GitHub Copilot Pro.
Like, I'm publishing https://github.com/andrewmcwattersandco/git-fetch-file right now with Claude Sonnet 4 (thank you for recently upvoting that to the front page). And the whole repository view that GitHub Copilot and Claude Sonnet 4 have on my projects seems like the same exact thing you get in Cursor, but Cursor for some reason took longer with the exact same models, and I'm not sure why.
Maybe they prompt the models differently? I haven't taken a look.
Also, Cursor seems to be literally a Visual Studio Code fork! But everyone's talking about it lately, and no one is mentioning this. I don't understand.
not exactly what you asked, but i've tried out Junie[0] because I've already got JetBrainz IDEs set up and love them.
It's terrible. For comparison, I've only used cursor on greenfield toy projects, but cursor is way better at the agentic stuff (the actual code generation AND the "review these changes" workflow) AND the tab/auto-complete stuff.
I hope Junie can make some leaps because I really like JetBrainz and dont want to see them fall behind
[0] https://www.jetbrains.com/junie/
Ah, nice. So far I've now seen Visual Studio Code with GitHub Copilot, Cursor, Zed with Agentic Editing, and now Junie.
It looks like chat-based agentic editing like this is going to be table stakes for AI-assisted editing moving forward.
> So far I've now seen Visual Studio Code with GitHub Copilot, Cursor, Zed with Agentic Editing, and now Junie.
Kiro, Void, Windsurf, Cline, Kilo, ... many, many others.
GitHub Copilot by itself is not directly comparable to Cursor. For example I use Zed + Copilot + Claude together at work for a similar workflow to Cursor.
Zed Agentic mode, Cursor, and Visual Studio Code with GitHub Copilot all have the same developer experience. That's what I'm confused about.
Cursor seems like the weakest player of the three, because it's just a Visual Studio Code fork.
This is where it starts to get subjective, but changing one part of the toolkit can have a huge effect on the quality of the assistant. For example I tried GPT 4.1 instead of Claude 4 recently and it took my setup from improving my productivity by 3-5x on coding tasks to, like, 0.5x. I can't point to a specific change other than tasks went from being done in 5 minutes if back and forth to being only partly done in 15 minutes.
I haven't used VSC in a year or Cursor at all, but I hear similar things from colleagues.
Annoyingly I'm finding cursor's autocomplete to be better than others, even though it's agent editing is not as good as claude code.
So I'm using CC in cursor (the little integration is nice) to get the best of both. None of cursors other AI features are helping though.
I do a ridiculous amount of dumb shit with cursor without paying extra for it and I can't be the only one. There is just no way they will ever make a profit with their current all-you-can-eat model. Like there have been times I have just copy and pasted code into a cursor window for the AI to add when I could have just pasted it myself just because it makes the context easier to deal with.
However, they will eventually get purchased by an AI company because the _product_ is great.
You are not the only one.
"Revert the changes to <file>..." 4 zillion tokens... 10 seconds...
Instead of > git checkout <file>
just to keep Cursor in the loop.
I assume I have probably eaten up my $20/month in tokens just on stuff like that.
At least in Copilot, the revert changes is more like undo/redo of a series of changes. Often I want to keep all of the changes except for the most recent one
git checkout would destroy this (and "corrupts" the Copilot session state)
Believe it or not, you can use Cursor and type in git checkout yourself. It's pretty wild.
> However, they will eventually get purchased by an AI company because the _product_ is great.
'Great' is in the eye of the beholder. For me, Cursor was one of the least-effective solutions of the many options (from Cursor and other AIDEs, to repo-centric web-based options like Jules, to CLI-based options like Claude Code) I evaluated a few months ago.
Don’t have a team with one AI service on it ….. have a team with three.
You’re crazy to only use one AI service if you’re doing serious development.
Use the 3 big ones all at the same time.
Ask them all to solve the same problem. Ask them all to evaluate each others solutions. Do this over and over in multiple iterations.
Each model is good at different things.
When you’re not getting a great result with this one, switch to another.
Using one AI is crazy when three together are more powerful.