Discovery 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

This seems to be an institutional, massively scaled version of https://github.com/karpathy/autoresearch.

In March Karpathy described this direction:

   The next step for autoresearch is that it has to be asynchronously massively collaborative for agents (think: SETI@home style).

Tweet is protected but in SERP caches: https://x.com/karpathy/status/2030705271627284816

Seems like Karpathy was largely focused on ML / SWE research rather than the other domains this group is after. Still, hard to imagine they were not influenced by autoresearch.

Andrej, if you're around, please share your thoughts on Discovery Loop.

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!”

"Our mission is straightforward" continued by the most complex sentence on that page. Wondering what the definition of straightforward is now

Really seems to embrace the "Making the world a better place by <<extremely convoluted, highly technical, jargon loaded mission statement>>"

Jesus, he left Google to do what everyone else is already trying to do? He must be so insulated he doesn’t realize what the real world is actually up to. I mean, organizations started on this exact same mission three or four years ago. Or longer. I suppose it’s better to wake up later than never.

That founding team is insane. Very excited to see what happens here. I really like that they do not mention AGI or anything like that. Their mission statement reads pretty pragmatic compared to other AI companies (the bar is very low…)

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.

To be honest, this feels more like a lifestyle business (aka hobby) than a startup. They truly deserve it, but I don't expect a huge success as a business.

That said, I hope they write cool papers with various peers across the industry without worrying too much about the competing dynamics. That'd be a blessing for humanity, and good for their spirit.

> Scientific discovery is bottlenecked.

Yah, by funding and how we award it, not by an imaginary lack of undergrad and grad students. Scientific funding requires a shotgun approach and many national science funds try to pick winners as opposed to funding broadly. When the folks who researched bacteria in volcanic vents or the molecular biology of the Gila monster they never could have imagined the industries and markets they'd create let alone the lives they'd impact (i.e., PCR and GLP-1 agonists). Lots of grants require you to explain how the work is "translational" or has some sort of economic application (even if not explicitly), but that'll just get us faster horses or whatever the Ford quote is.

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.

Someone who left DeepMind over Google's agreement to provide military AI to the US government tried to get Jeff Dean to quit too:

https://turntrout.com/why-i-left-google-deepmind

Maybe this is what happens when someone with Jeff Dean's standing tries to quit?

TBH, I'd rather have Jeff Dean working on the creepiest-possible tech for ICE than joining the race to automate AI research. Automating AI research is terrifying.

Jeff Dean, Sanjay, et al have achieved so much. I'm very happy for them. Truly deserving.

Sometimes I couldn't resist wondering if I'll ever do work that has a tenth of the impact of theirs.

I wonder if Jeff Dean facts [1] (I hope people remember the reference) will carry over to the new startup.

[1] https://github.com/LRitzdorf/TheJeffDeanFacts

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.

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 engineer-to-work 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.

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 optimization loop have discovered transformers?

I have used something similar. I set up a team of agents that researches, proposes, builds and audits. Then rinse and repeat. I have used it for different topics. It hasn’t made me a millionaire, but I haven’t lost money either - so that’s some sort of win, right? But I would not have been able to ideate, test at that speed and quality without an LLM.

For sure made with Claude code for front end, but I’m excited to see where they go

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.

> Oriol Vinyals, Sanjay Ghemawat, Jeff Dean, Quoc Le

as founding members is crazy !

Big news aside, it feels exciting to see them leave and pursue startup. They could have stayed back, and retire

The site itself is really leaning into the “made with Fable” aesthetic

Discovery systems are making a comeback huh.

By the middle of the 2030's the world we live in will be unrecognizable.

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.

> 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 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 :)

What's the business model of these startups?

Did the ycombinator podcast which included giving advice to startup founders just a few days ago:

https://www.ycombinator.com/library/Vy-jeff-dean-the-1-rule-...

As LLM coding agents plateau— at least for the average engineer without tens of thousands of dollars or swarms of agents to run —I’d say that, from here on it’s going to be about ASICs, specialized LoRA/or-equivalent models, or a Ruby on Rails for LLM context engineering and orchestration, which LangChain and others seems well position, including Google as they own the entire stack. LLM free lunch has been over for a while, perhaps since the ReAct loop, and has been official since Ilya mentioned it at NeurIPS.

I feel the most exciting development these days is self-evolving agents. Especially if you have a way to verify their outputs with a formal system, or with a system developed since the 60s by armies of PhDs.

DeepMinds Gnome is a good example, where they use DFT to verify outputs. Approximating NP-problems is always fun for those who dare.

I am also building in this space. Its a mix between HPC, AI, and hard science. Pretty fun compared to waking everyday to LLM news that seem more like marketing stunts.

Is it a very hard problem to solve that jeff and the other legendary engineers have decided to quit and start on this?

Careers page (if anyone is interested) → https://jobs.ashbyhq.com/Discovery-Loop

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.

They're structuring the new company as public benefit corporation.

Automating ML/AI research seems completely tractable. Most of the other claims seem much less doable.

National Labs in the US have been doing this for a while now. I feel like the private sector will take the lead soon.

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...

What a team.

not that it really matters, but is he leaving Google?

This seems interesting! I wonder how this will play out.

This is “google brain”

The job req has "Recursive Self-Improvement" as one of the "area of expertise" checkboxes lol

Not through conscience.

I've seen tiny tiny hints from the outside that Jeff Dean was dealing with too much internal BS. Two examples that come to mind: Having to deal with Timnit Gebru fiasco, and even chips in the TPU series getting marketing names (Trillium and Ironwood) before switching back to more standard numbering.

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".

In 2018 there was a beautiful New Yorker article on Jeff Dean and Sanjay https://www.newyorker.com/magazine/2018/12/10/the-friendship...

"The speed of light in a vacuum used to be about 35 mph. Then Jeff Dean spent a weekend optimizing physics."

When they say experiments, do they mean using physics simulators?

Related:

Jeff Dean leaving Alphabet

https://news.ycombinator.com/item?id=49184746

Computation is not the hard part of discovery.

The AI designed italics on thin font is hard to not see as slop.

So Ralph Wiggum in a suit?

I am available for hire.

nice

I smell vapor.

Another way to see this is: a bunch of renowned google engineers realized they can grab some of the VC pie for themselves

https://www.geekwire.com/2026/the-startup-idea-that-convince...

[flagged]

[flagged]

I wish them well, but this firm will likely fail miserably. The reason is that the value is in having access to real world hardware platforms that AI can control, not in the harness that controls them. There exist plenty of harnesses already. These people couldn't even get Google to build a top LLM. Before you dismiss and downvote, I dare you to counter it.

Is this a joke? Site is not loading for me.

> 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!