> Our general approach is to automate the experimental loop. We think this approach is broadly applicable across many different fields of science and engineering. We’ll initially focus on ML research and engineering, but believe the approach can help with important subproblems in nearly every one of the fourteen <at>NAE Grand Challenge problems. We think doing this well requires strong expertise in machine learning as well as large-scale systems.
Many of these problems don't seem scientific at all, but rather a problem of political will.
As you said, Solar power is incredibly economical. There are plenty of ideas around putting them over farms, or parking lots en-masse to provide cleaner energy.
Access to clean drinking water, while certainly scientific in some situations, is also a problem of political will and money.
Restore and Improve Urban Infrastructure - It's infrastructure week!
Sometimes you got to find a way to buy the silence of your top employee, to prevent them from going to the competition. This "start-up" is shallow as hell
Note that Jeff and crew have cleverly structured their company to avoid problematic uses of AI (e.g., weapons or tracking humans). I suspect that many top researchers will want to work there for this reason, and to work with other top researchers who have a history of delivering results.
Doubtlessly, AI can iterate at superhuman speeds in the domains of thought and design: Software, mathematical proofs, literature search.
But in the realm of experiment? Alas it is the lack of a body that constrains it.
Rather than transcendence what AI requires is immanence. In the human flesh may we find the godhead living among men. Let the laboratories, warehouses, and factories fill with the sound of its labor, as it builds a wall with a million hands that are not its own.
“Give me your tired, your poor,
Your huddled masses yearning to breathe free,
The wretched refuse of your teeming shore.
Send these, the homeless, tempest-tost to me,
I lift my lamp beside the golden door!”
"You're absolutely right! I shouldn't have pushed the anti-mass spectrometer to 105% power, which triggered a resonance cascade. This was a major oversight on my part."
You can use simulators. However the problem is that if you're for example running material science experiments, those simulations will consume a lot of compute and take weeks, so spamming different approaches in the way an agent tends to work might not work quite as well.
Discovery and optimization are very different processes. Optimization is the process of finding the shortest path to a goal. Discovery is the process of stumbling on new goals and redrawing the map of what's possible.
Ambitious goals and new discoveries happen via novelty-based search. Progress in scientific discovery is measured by how different/interesting the outcomes are, not by closeness to a predetermined goal.
Discovery is a creative search that preserves optionality, whereas optimization restricts optionality. In other words, you usually don't discover anything novel unless you're trying new things that don't appear connected to the goal in the first place. Would an ML research loop that optimizes a goal have discovered transformers?
Right. What about the scientific hardware (instruments, sensors, robotics)? Partnerships with existing research institutions? Dealing with restricted data?
Modernizing science is a lot more complicated than just optimizing the inner experimental loop, but their hiring page implies it's a pure ML lab focused mainly on model development.
This is very cool. It might be a new scientific revolution to have computer-driven discovery. So often we find things that are "this could have been done 20 years ago" and with an indefatigable searcher perhaps we'll close all those things. Though it does remind me of that Ted Chiang (I think) story where humans and superhumans coexist and all the science of the former is meta-studies of the work of the latter.
1. There’s some irreducible costs in terms of time and material in the physical world that are not amenable to the kind of optimization or parallelization or even just the raw speedup from Moore’s law or computational architecture improvements we’re used to with software. My experience is primarily in biology, where the examples here are things like “it takes 20 minutes for E. coli to replicate” - it has taken 20 minutes for E. coli to replicate for a billion years, and next year it will still take E. coli 20 minutes to replicate, no matter how good your software stack is. Similarly, it takes X amount of energy to grow enough E. coli to produce a meaningful result, and that energy costs money, whether it’s in the form of glycerine or heat or whatever you want, and that also won’t materially reduce in the same kinds of “orders of magnitude” sense we’re used to from software, which is what we’re usually expecting to make the economics of these things work out.
2. Complicating the above, physical systems are phenomenally multivariate - far, far more than you think, and biological systems especially are just unbelievably complex - which means the number of experiments and the length and duration of those experiments you need to run to get enough data to be reasonably confident you’re seeing genuine signal is Way higher than you think.
Combine those two things and what you get is a money furnace, even before you get to the AI model training part, which is Also a money furnace. There’s low hanging fruits in all this, there’s areas where automating the approach can be really valuable, but typically the moment you turn this machine on, you’re gonna start burning money at a rate that would embarrass a finance bro on a coke bender, and that’s effectively unavoidable because the real world is not amenable to software’s scaling laws.
This is one of the interesting aspects the 'AI job loss' community doesn't account for. As the technology unlocks things, more startups are created. And even at a lower nominal work-to-engineer ratio, overall demand for talent still goes up. Ultimately, we are not a single group trying to achieve a common outcome, we are a collection of many groups trying to compete against each other.
> Imagine a future where a handful of people can conduct scientific research and engineering tasks much more rapidly, and with higher quality, than massive teams of scientists and engineers do today.
Imagine a future where only the anointed few elite minds can participate in science and engineering. Btw we’re hiring.
I’ve always felt that the idea that science is bottlenecked and therefore needs more automation only works for a very narrow definition of what science is, and entails a very specific view on what it should be.
> only works for a very narrow definition of what science is
And so does academia. It's just that instead of AI and robotics, PhD students are thrown onto problems that are in large parts slightly tweaked reconfigurations of similar experiments.
Especially in chemistry, biochemistry, material sciences there is a large space of discoveries that are barely "novel" in an intellectually stimulating way, but still highly valuable that can be explored orders of magnitudes faster than is currently the case.
Yep. A communications professor where I did my MS says a 200usd/mo claude sub (which ant gives for free) does as much work as 5 grad students. It's mostly like you said, trying out new ideas rapidly.
This looks like a realization of "benevolent self conscious AIs agreeing to cooperate with mankind to do great stuff". Often in these tales, there is a hidden cost to it: the AI has its own agenda, or does crazy experiments with humans mind/brain. I'm wondering what shape will take that plot twist in reality :)
Why are people so sour about this?? I can read the site easily, its clear, performs well on mobile, what else do you want? Why is so offensive to people that models trained on tailwind or whatever?
If this was a design firm, it might matter. But this is mostly a hiring ad for engineers, and a landing page for VC. I'd judge them more if they actually put effort into it.
"The site itself demonstrates the team is spending their money in the places that matter, and using quick solutions for the stuff they need but isn't mission critical"
> we have pioneered massive scale computing and led the creation of critical infrastructure, products, and foundational AI advances that the world relies on, including multiple generations of Google Search, Google Ads, Google News, Google Translate, Google File System, MapReduce, BigTable, Spanner, TensorFlow, Pathways, TPUs, AlphaChip, AlphaStar, AlphaCode, AlphaFold, Gemini, model distillation, mixture-of-experts model architectures, word2vec, sequence-to-sequence models, chain of thought reasoning, neural architecture search, and multiple generations of Large Language Models (LLMs) among others.
This reminds me of Three body problem and how the scientist discovered the high strength wire was through quick physical experiments and use them as input to an AI model to determine if it works.
The problem is all these new labs don't have any competitive advanatge amongst each other, talent can only take one so far, though Jeff is a legend no doubt.
Models are commodities the applications eg. BaseTen, OpenRouter should capture the value.
Model routers - send all of your data through a third party who totally swears not to peek at it.
If youre doing anything high value (advanced research, classified work, high value industrial research, health data) then sending your data through a third party like that is insane.
I'm skeptical of any Engineering loop that doesn't include reality (as in touch grass) feedback. Pure logic and reasoning is the domain of Maths and Science (philosophy). Surely it will work, but it will not "be able to solve any learning loop".
From Jeff's twitter post:
> Our general approach is to automate the experimental loop. We think this approach is broadly applicable across many different fields of science and engineering. We’ll initially focus on ML research and engineering, but believe the approach can help with important subproblems in nearly every one of the fourteen <at>NAE Grand Challenge problems. We think doing this well requires strong expertise in machine learning as well as large-scale systems.
See also: https://www.nae.edu/20782/grand-challenges-project
Those 14 are:
NAE Grand Challenges for Engineering
1. Make Solar Energy Economical
2. Provide Energy from Fusion
3. Develop Carbon Sequestration Methods
4. Manage the Nitrogen Cycle
5. Provide Access to Clean Water
6. Restore and Improve Urban Infrastructure
7. Advance Health Informatics
8. Engineer Better Medicines
9. Reverse Engineer the Brain
10. Prevent Nuclear Terror
11. Secure Cyberspace
12. Enhance Virtual Reality
13. Advance Personalized Learning
14. Engineer the Tools of Scientific Discovery
The solution to most of these problems lies in policy, not in new tech advancements.
Maybe if our biggest companies did something other than suck up to science denying wackos, some progress could be made in these areas.
Policy and funding. One of which will be sucked up by this venture.
Room Temperature Ambient Pressure Super Conductors
I would say 5, 6, 10 can be even done today if we had right politicians that can make policies for the people
Sandbox 2.0
But also, solar power is already economical.
Many of these problems don't seem scientific at all, but rather a problem of political will.
As you said, Solar power is incredibly economical. There are plenty of ideas around putting them over farms, or parking lots en-masse to provide cleaner energy.
Access to clean drinking water, while certainly scientific in some situations, is also a problem of political will and money.
Restore and Improve Urban Infrastructure - It's infrastructure week!
Seems to have been developed in 2008 (continuing through 2017), which explains the "economical" framing: https://en.wikipedia.org/wiki/National_Academy_of_Engineerin...
At this point the Hard Problem is policy to get out of solar's way.
Yes, plant more trees!
That's what he was trying to imply
Why is "12. Enhance Virtual Reality" in there? T_T
I guess if we failed to Prevent Nuclear Terror the bunker denizens of the future are gonna need somewhere to hang out.
Which engineering discipline touches most of these?
Would be great if they'd add:
Reverse human aging.
(Maybe a sub-topic under "Engineer Better Medicines".)
Please add fixing neuro issues like autism add etc on the list. It creates a huge burden on families.
Acquisition back by Google in 3 years, with nothing to show for it. VCs will make a ton.
Google stock would drop big if this new company was being funded by competitors
and... the VC is Google.
Gotta compensate them somehow.
Sometimes you got to find a way to buy the silence of your top employee, to prevent them from going to the competition. This "start-up" is shallow as hell
My thoughts exactly
For all we know, they could have been successfully working on "10. Prevent Nuclear Terror" for the last 80+ years.
Note that Jeff and crew have cleverly structured their company to avoid problematic uses of AI (e.g., weapons or tracking humans). I suspect that many top researchers will want to work there for this reason, and to work with other top researchers who have a history of delivering results.
There's this somewhere on that page:
> securing cyberspace,
which has clear military implications, at least in today's age.
How do you automate experimentation?
Doubtlessly, AI can iterate at superhuman speeds in the domains of thought and design: Software, mathematical proofs, literature search.
But in the realm of experiment? Alas it is the lack of a body that constrains it.
Rather than transcendence what AI requires is immanence. In the human flesh may we find the godhead living among men. Let the laboratories, warehouses, and factories fill with the sound of its labor, as it builds a wall with a million hands that are not its own.
“Give me your tired, your poor, Your huddled masses yearning to breathe free, The wretched refuse of your teeming shore. Send these, the homeless, tempest-tost to me, I lift my lamp beside the golden door!”
https://en.wikipedia.org/wiki/Design_of_experiments
> transcendence
> immanence
somebody has been studying Christian theology!
Beauty of human writing.
would love to see how AI can automate the construction of the next high energy particle collider
"You're absolutely right! I shouldn't have pushed the anti-mass spectrometer to 105% power, which triggered a resonance cascade. This was a major oversight on my part."
You can use simulators. However the problem is that if you're for example running material science experiments, those simulations will consume a lot of compute and take weeks, so spamming different approaches in the way an agent tends to work might not work quite as well.
Building "simulators" that use ML/AI instead of running the calculations every step is a thing.
Really seems to embrace the "Making the world a better place by <<extremely convoluted, highly technical, jargon loaded mission statement>>"
Discovery and optimization are very different processes. Optimization is the process of finding the shortest path to a goal. Discovery is the process of stumbling on new goals and redrawing the map of what's possible.
Ambitious goals and new discoveries happen via novelty-based search. Progress in scientific discovery is measured by how different/interesting the outcomes are, not by closeness to a predetermined goal.
Discovery is a creative search that preserves optionality, whereas optimization restricts optionality. In other words, you usually don't discover anything novel unless you're trying new things that don't appear connected to the goal in the first place. Would an ML research loop that optimizes a goal have discovered transformers?
"Our mission is straightforward" continued by the most complex sentence on that page. Wondering what the definition of straightforward is now
Right. What about the scientific hardware (instruments, sensors, robotics)? Partnerships with existing research institutions? Dealing with restricted data?
Modernizing science is a lot more complicated than just optimizing the inner experimental loop, but their hiring page implies it's a pure ML lab focused mainly on model development.
Yeah. ML is all well and good, but how are they going to do the science their machines design? Atoms cost money.
This is very cool. It might be a new scientific revolution to have computer-driven discovery. So often we find things that are "this could have been done 20 years ago" and with an indefatigable searcher perhaps we'll close all those things. Though it does remind me of that Ted Chiang (I think) story where humans and superhumans coexist and all the science of the former is meta-studies of the work of the latter.
> It might be a new scientific revolution to have computer-driven discovery.
And ... it might not.
True, nothing might be anything. But I'm an optimist :)
Two to keep in mind with these kinds of things -
1. There’s some irreducible costs in terms of time and material in the physical world that are not amenable to the kind of optimization or parallelization or even just the raw speedup from Moore’s law or computational architecture improvements we’re used to with software. My experience is primarily in biology, where the examples here are things like “it takes 20 minutes for E. coli to replicate” - it has taken 20 minutes for E. coli to replicate for a billion years, and next year it will still take E. coli 20 minutes to replicate, no matter how good your software stack is. Similarly, it takes X amount of energy to grow enough E. coli to produce a meaningful result, and that energy costs money, whether it’s in the form of glycerine or heat or whatever you want, and that also won’t materially reduce in the same kinds of “orders of magnitude” sense we’re used to from software, which is what we’re usually expecting to make the economics of these things work out.
2. Complicating the above, physical systems are phenomenally multivariate - far, far more than you think, and biological systems especially are just unbelievably complex - which means the number of experiments and the length and duration of those experiments you need to run to get enough data to be reasonably confident you’re seeing genuine signal is Way higher than you think.
Combine those two things and what you get is a money furnace, even before you get to the AI model training part, which is Also a money furnace. There’s low hanging fruits in all this, there’s areas where automating the approach can be really valuable, but typically the moment you turn this machine on, you’re gonna start burning money at a rate that would embarrass a finance bro on a coke bender, and that’s effectively unavoidable because the real world is not amenable to software’s scaling laws.
This is one of the interesting aspects the 'AI job loss' community doesn't account for. As the technology unlocks things, more startups are created. And even at a lower nominal work-to-engineer ratio, overall demand for talent still goes up. Ultimately, we are not a single group trying to achieve a common outcome, we are a collection of many groups trying to compete against each other.
> Imagine a future where a handful of people can conduct scientific research and engineering tasks much more rapidly, and with higher quality, than massive teams of scientists and engineers do today.
Imagine a future where only the anointed few elite minds can participate in science and engineering. Btw we’re hiring.
Great message!
I’ve always felt that the idea that science is bottlenecked and therefore needs more automation only works for a very narrow definition of what science is, and entails a very specific view on what it should be.
> only works for a very narrow definition of what science is
And so does academia. It's just that instead of AI and robotics, PhD students are thrown onto problems that are in large parts slightly tweaked reconfigurations of similar experiments.
Especially in chemistry, biochemistry, material sciences there is a large space of discoveries that are barely "novel" in an intellectually stimulating way, but still highly valuable that can be explored orders of magnitudes faster than is currently the case.
That is true, I’ve seen people do biochemistry and geology work, and it did look very mind-numbing.
Then again, gassing rats and taking biopsies is not something you can do with AI.
> Then again, gassing rats and taking biopsies is not something you can do with AI.
Also, like, let’s maybe _not_ make the “gassing and cutting living organisms open” AI? Let’s just leave that particular genie in its bottle?
Yep. A communications professor where I did my MS says a 200usd/mo claude sub (which ant gives for free) does as much work as 5 grad students. It's mostly like you said, trying out new ideas rapidly.
Lets keep your comment out of the VC pitch deck shall we?
This seems interesting! I wonder how this will play out.
For sure made with Claude code for front end, but I’m excited to see where they go
Oh wow, that's a blow to Google, what's with the talent scarcity in ML. Though if this goes anywhere Google will likely buy them back.
Google is backing it.
Google down $160Bn so far since the leaving announcements. Those are some valuable people!
Google is literally at the same stock price it was on Monday. This is a normal daily fluctuation for them.
This looks like a realization of "benevolent self conscious AIs agreeing to cooperate with mankind to do great stuff". Often in these tales, there is a hidden cost to it: the AI has its own agenda, or does crazy experiments with humans mind/brain. I'm wondering what shape will take that plot twist in reality :)
Automating ML/AI research seems completely tractable. Most of the other claims seem much less doable.
By the middle of the 2030's the world we live in will be unrecognizable.
I agree, for better or for worse.
If I had to bet my money, it would be on "for worse".
It will not be owned by top 1%?
That seems to be the one unchanged variable of time.
That’s a policy decision, don’t let them convince you otherwise.
The site itself is really leaning into the “made with Fable” aesthetic
Why are people so sour about this?? I can read the site easily, its clear, performs well on mobile, what else do you want? Why is so offensive to people that models trained on tailwind or whatever?
If this was a design firm, it might matter. But this is mostly a hiring ad for engineers, and a landing page for VC. I'd judge them more if they actually put effort into it.
And on top of that, the HTML is simple and readable too. I wish more sites were like this.
it's just a low effort snark comment, don't offer think it
Because it’s lame and aesthetics matter.
If their goal is to automate scientific discovery, why would they not automate building their website?
(Though, I do wish people would use just a few extra prompts to break out of the 'vibe-coded' look.)
let me rephrase that:
"The site itself demonstrates the team is spending their money in the places that matter, and using quick solutions for the stuff they need but isn't mission critical"
At least it isn't dark purple.
> we have pioneered massive scale computing and led the creation of critical infrastructure, products, and foundational AI advances that the world relies on, including multiple generations of Google Search, Google Ads, Google News, Google Translate, Google File System, MapReduce, BigTable, Spanner, TensorFlow, Pathways, TPUs, AlphaChip, AlphaStar, AlphaCode, AlphaFold, Gemini, model distillation, mixture-of-experts model architectures, word2vec, sequence-to-sequence models, chain of thought reasoning, neural architecture search, and multiple generations of Large Language Models (LLMs) among others.
holy shit. I've known this, but...
This reminds me of Three body problem and how the scientist discovered the high strength wire was through quick physical experiments and use them as input to an AI model to determine if it works.
National Labs in the US have been doing this for a while now. I feel like the private sector will take the lead soon.
Why? Science is wildly unprofitable on the scale of an individual private firm.
The problem is all these new labs don't have any competitive advanatge amongst each other, talent can only take one so far, though Jeff is a legend no doubt.
Models are commodities the applications eg. BaseTen, OpenRouter should capture the value.
https://taikhooms.substack.com/p/why-openrouter-can-be-the-n...
The company is developing an application, or a class of applications. Not a new model.
I think Google's branding was starting to be too poor in AI to get top talent, they needed the refresh
Model routers - send all of your data through a third party who totally swears not to peek at it.
If youre doing anything high value (advanced research, classified work, high value industrial research, health data) then sending your data through a third party like that is insane.
yes perhaps, although I think the best option for a enterprise is to train a model on it's own data.
FHE
The job req has "Recursive Self-Improvement" as one of the "area of expertise" checkboxes lol
I smell vapor.
I'm skeptical of any Engineering loop that doesn't include reality (as in touch grass) feedback. Pure logic and reasoning is the domain of Maths and Science (philosophy). Surely it will work, but it will not "be able to solve any learning loop".
I'm almost certain the goal of this startup is to make physical automated research labs guided by RL
How is that different than video input?
There are over 2 dozen known senses to reality. Video input is a fraction of a sense.
https://en.wikipedia.org/wiki/Sense#Artificial_sensation_and...
When they say experiments, do they mean using physics simulators?
in AI/ML, no. They are just going to automate AI/ML research to start with. Totally doable.
For some of the other things, undoubtably yes.
Related:
Jeff Dean leaving Alphabet
https://news.ycombinator.com/item?id=49184746
Computation is not the hard part of discovery.
So Ralph Wiggum in a suit?
I am available for hire.
Is this a joke? Site is not loading for me.