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.
> 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.
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
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!
It is not economical compared to alternatives that is why you have to have government to force people to do things. In many places such as Pakistan where solar power is not economical on paper is actually very successful in practice because it is actually profitable.
To make solar power practical and economical you need may a square foot of solar panel being able to get enough energy to power and entire home for a week
Not sure what you’re talking about here. We can’t replace all energy needs with solar but it’s clearly one of the cheapest energy sources and with the added benefit of low capital expense to get started so you can set it up in distributed grids without the massive expenditure to support nuclear installations.
Why not just add 'mind control' while you're at it.
Only death stops stagnation in the end. Without death, especially if death can be avoided by the rich and powerful but not the poor, life will get much, much worse for the average person (until only the rich and their automated capital remain I suppose, in which scenario they will simply turn on each other).
When one of them dies, they want to leave behind a puzzle so complex that entire groups of the population dedicate their lives to solving it within the virtual world. They look old enough to have a lot of favorite 1980’s and 90’s pop culture references, so those will probably be the clues.
I'm guessing this might be about "teleoperation" (like remote surgery via robots + VR) and being able to remote training as well. The binocular vision VR gives you compared to flat screens help a lot with depth perception for precision of incisions for example.
The science denying wacko is a but like drunk homeless guy on street while you have your Ferrari parked in the driveway. The wacko has a golf club. There is too much to lose by pickup a fight. You might win eventually but the next wacko shows up with another golf club soon.
> I do not see how the second sentence follows from the first.
Point 1 on the list is "Make Solar Energy Economical".
Solar is economical today - mostly due to China investing (and heavily subsidising) in solar for the past couple of decades; but compare to the US where certain particular big-businesses (oil companies, mostly) were instead cynically funding disinformation efforts and getting into bed with the Republican party (which dovetailed with the GOP's allying with other science-denying movements of the Bush Jr era like creationism and public-health matters with abstinence-only sex-ed and defunding gun safety research efforts) - means we're decades behind where we could have been...
Consider an alternative past, where the GOP had the backbone to resist the oil industry's corruptive influence and instead made a big bet on American Solar; it's entirely possible that instead of MAGA today we'd instead have a right-wing coalition strongly supporting solar and wind energy because they align nicely with American rugged individualism - whereas the current situation on the right is an unprincipled farce with inconsistencies in policy positions at every turn.
This is not a particularly new struggle: Jimmy Carter installed American-made solar water panels on the White House in 1979, then Reagan tore them out.
> The solution to most of these problems lies in policy, not in new tech advancements.
That is not mutually exclusive. If technological advances result in a given technology becoming cheaper, more scalable, and easier to deploy, they also make it easier to advocate for and implement the relevant policies.
You can think of it like this: our "political technology" is not good enough to use solar energy at its current prices to replace fossil fuels as fast as we would like. Well, what about if we cut the price of solar by a factor of five? Perhaps it will be good enough then.
That sounds like just playing further and further into the game of the corrupt leaders. Who do you think would profit off that 5x margin? Would that margin come more likely from a scientific breakthrough of via some new exploitation of natural resources or human labor?
It's yearsss past time that our leaders should have changed policy.
Hey now, our biggest companies do more than suck up to science denying wackos, don't be so unfair... sometimes our biggest companies fund their wacko superpacs!
We have a good understanding of the function (and more importantly dysfunction of) kidneys, lungs, heart, etc. from high level to cellular level.
For brain, our understanding is fuzzy, more like "this part is important for that behavior" or "here is how neuron works" but we don't have a holistic understanding.
If we had that, we could more easily diagnose and treat neurological disorder.
To understand, same reason you reverse engineer anything. Doesn't have to have a further goal than that, understanding the brain better helps in so many ways. But like most technology, obviously can be used for bad too. Should we just skip researching some topics then?
People on HN keep saying it is, but I'm still not seeing it. Companies are building datacenters in vast, sun-blasted deserts and still choosing to power those with natural gas. This in turn makes people complain about emissions pledges being reversed, but if it were economical, no pledge would be needed.
Of course, USA has cheaper oil/gas than other countries. But if you look elsewhere, rich countries are subsidizing solar, poor ones are basically not using it.
In the chart for section 3, how many countries have seen their share of electricity being generated by fossil fuels increase in the last 5 years? Only Canada.
Take a look at the charts for Pakistan, Australia, Nigeria, and China for the last few years. Pretty dramatic drops for fossil fuels generation.
Very honorable effort, but a lot of these seem to touch heavily regulated industries impeded by unwise or outdated policy no less than by the lack of clever engineering - medicine, education, urbanism, energy. I wonder if they've given some thought to the key blocking factor as well.
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.
That is also what I understand from that page. Autoresearch is the closest thing we (mainstream audience) know of but I am sure there is already an active research literature around it.
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, causing 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.
You're halfway there, but the only impediment isn't on the side of the researchers. Many of these topics they're trying to solve involve human subject research. Even with tireless embodied researchers who work around the clock and don't require breaks, you can't make the thing you're studying happen faster. The biggest reason we use poor proxy measures for things like longevity and mortality research is the simple impracticality of finding two groups of randomly selected people, ensuring you can control their entire lives for 60 years, the only difference between them is one variable, and see who lives longer. Putting aside the ethics, even if you could find willing subjects and actually control their entire lives to that extent, it would still take 60 years to gather the data you need. It doesn't make any difference whether robots or humans are running the program.
One of my favorite books from the past few decades is The Extravagant Universe, written by one of the astronomers who helped discover dark energy and develop the current most-accepted model of cosmology. I love this book because of the emphasis on physical process in astronomy. Part of the reason it took decades to study this problem is they need to collect data from supernovae. Those only happen so often in places we're looking. You can't automate alignment of the heavens. It happens when it happens.
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.
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.
Do you believe that securing cyberspace is problematic solely because it has military implications? I mean, everything has military implications. That fact doesn't imply, however, that those things are bad for society.
Securing cyberspace matters to everyone. Defending critical infrastructure or design of tactical cyber-offense is reasonably in scope for military work.
However, reducing (or rather limiting the increase of) PII leakage and impact of ransomware activities is much closer to day-to-day mainstreet of most people.
Anyone committed to advancing science should care about this regardless of its potential contributions to defense.
So does more efficient cooking methods, but that is not the primary focus.
As opposed to say weapons systems or targeting systems, which are really only for military use.
The military needs a lot of things that other people need, and some things that only the military needs. If you don't work on the things only the military needs, I think you're in the clear.
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.
A public benefit corp shouldn’t be a start up. The primary goal of a start up is to grow as quickly as possible which is rarely benefits the public.
Very silly to call every non start up a lifestyle business. It’s just a business. Start up are the weird thing that almost always an obscene waste of time and money, but sometime creates google.
> The primary goal of a start up is to grow as quickly as possible which is rarely benefits the public.
Why is that so? Fast growth, when achieved honestly, is a result of solving user pain that others haven't. Maybe you think so because users != the public, but I think in totality the public is a collection of users who all have needs they want met.
For certain amount of fast growth: yes. Then there's continued "growth-hacking" and enshittification to keep fast revenue growing after the pain point has been solved with dark patterns and questionable tactics.
Why? Because investors poured a bunch of money in to support fast growth and now they want their money back. And incremental growth won't do. Since 9 out of 10 of the investments fail, the surviving one has to continue to growth-hacking revenues.
A PBC is not a charity or a non-profit. PBCs are for-profit businesses with the goal of making money. In day-to-day business they're indistinguishable with other for-profit corporations, including fundraising and investment. The only real practical difference is that they give directors a little more leeway in their fiduciary duties to say "no" to doing evil things.
The advantage to a PBC is protecting founders from a serious problem with standard corporations: you might bring on investors who could subsequently demand you pollute, exploit people, and/or do other immoral activities for profit. You don't have to do these things to grow a business quickly.
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.
Gemini has done absolutely nothing for me. I can't even shut off the navigation feature on my phone using only hands free, when I get close to my destination. I have to take my eyes off the road, look down, and tap to exit.
Google's advanced AI cannot even exit a mobile app.
Antigravity + Gemini Pro absolutely RIPS through fullstack react + react-native apps / systems. I pay ~$20/month and I basically don't have to do my real work anymore. My time is freed up to learn systems programming and blender.
Oh nice have things stabilized with models/quotas and Antigravity is usable with the Pro plan again? I was getting a crazy amount of value out of Gemini CLI for “free” on my Pro plan but after the shutdown struggled to make agy work with the updated quotas/bigger models without instantly being rate limited and just started doing everything on Codex/Opencode Go.
Yeah can't remember how long ago but it was running out after less than an hour - then they announced they were loosening restrictions and I've been able to easily get everything I need to done without hitting limits.
I don't do huge automatic project wide hands-off agent loops though. I spent a lot of time architecting my systems to be easy to generate code on top of with pointed & detailed prompts. So I'm not abusing context... YMMV
Those are all fake, part of an internal Google narrative that overstates individual contribution, and obscures the work of large engineering teams.
Here are some Jeff Dean well sourced facts:
- Already part of engineering of Google indexing systems that lacked basic checksums and ran on non-ECC hardware, allowing silent data corruption.
- One of the authors of LevelDB a database with so many documented crash-consistency, recovery, and data-loss weaknesses for years. Just check their Github project. LevelDB current tracker contains unresolved crash consistency, recovery and corruption reports going back almost 12 years on GitHub
- In AI engineering technical lead, let TensorFlow lose researcher mind share to PyTorch, and caused Google fragmented landscape across TensorFlow and JAX.
- Had the people at Google who invented the Transformer architecture, but failed, to turn that lead into the first dominant public LLM.
- As AI engineering and VP management let Google Brain and DeepMind remain duplicated and internally competitive for too long.
- Let Noam Shazeer leave and then spent heavily to bring him back with nothing to show for.
- Part of Technical VP leadership who had Bard rushed to launch with factual errors in Google own promotional material.
- The first Gemini demonstration overstated how real-time and interactive the system actually was, being basically a fake.
- Part of the VP and AI technical leadership who had Google AI Overviews launched with weak source quality controls and repeated satire and low-quality web content as factual advice.
- Part of teams that launched AlphaChip performance claims that were difficult for outside researchers to reproduce and remain technically disputed.
- Jeff Dean public explanation of Gebru departure was contested and damaged confidence in Google scientific governance.
- Jeff Dean was part of the team at Google that removed or marginalized prominent internal AI ethics critics shortly before many of their warnings became product problems.
- Jeff Dean was one of the managers behind Project Dragonfly supporting censorship.
- Jeff Dean is part of the VP technical leadership approving Project Nimbus supporting an ongoing genocide.
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.
If you check out some sub-tweets from people in the org, it wasn't really all butterflies internally for a while. Sorry, really don't want to name people and give examples.
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.
What percentage of people work at a startup though? Not just new/small business, which could include restaurants, local services, etc., but tech/science startups that would meaningfully benefit from AI.
I'd bet you could 10x the number and still be in low single digit percentages of the US workforce. And it seems pretty likely that AI-enabled startups will also employ less people per-startup.
If AI causes a white-collar jobs apocalypse, I don't think startups are picking up the slack, although it'll plausibly cushion the blow somewhat for top-performing tech workers.
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.
It's 90% to advance science via cheap labor and 10% to train a small group of future experts who will hire grad students to 90% advance science via cheap labor etc.
...
Yes. They have grad students too. This is just like having more grad students that don't need to be trained so the work you can get done is not bottlenecked by the number of people you can train.
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?
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.
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.
To many people, myself included, who have to wade through huge amounts of low-quality AI slop which is actually a negative value: no-one benefits when non-experts fire-off one-shot LLM/agent prompts to produce PRs, reports, documentation or "journalism" riddled with imagined truth and factual errors - and the more that people like me have to evaluate these inputs for our job and how wrong they are it pisses us off - but also it means we pick-up on the hallmarks, tells and cliches of these low-effort, no-respect submissions - and now the litany of tells includes this beige-themed, blurred-backdrop-navbar corporate website look: theirs site looks like the 4 or so other LLM-generated, negative-value slop-farm sites I've wasted time on recently - all over the past few weeks.
So I'm saying that, without having known anything about what "Discovery Loop" is - or is not - but landing on their site and and seeing that beige colour and blurred-backdrop navbar, I immediately moved to close the tab; what kept me here was seeing the HN thread had over 100 comments by now and read more about it; if not for that then I wouldn't have given it further thought.
Having "that" beige site look with same the overused looks is either an unintentional indication that the site's author used a low-effort AI prompt to generate the site and that the content within is likely to be low-quality, low-value - or it's an intentional lure to appeal to those who uncritically share in the AI psychosis and so, I assume, are a good target to seek investment from even if it means losing the audience of cynical Internet critics like myself because they know people like me won't be breathlessly repeating their vision-statement on LinkedIn and throwing money at them - kinda like how scam emails intentionally include mistakes for better audience selection. And both possibilities have unpleasant implications.
------
Anyway, regardless of the background of the team behind it, the way the project is described sounds exactly like the recursive-self-improvement and simulated-science thought-experiments from _that other website_ - it's the kind of thing I expect Angela Collier to brutally takedown in an amusing video.
Some parts are pretty annoying to read... the paragraph beginning in "Our mission is straightforward:" has many lines with just 2-3 words, massive font, and tons of unused whitespace to the right. Changing the page/browser zoom doesn't seem to help much either.
"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 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 :)
I mean what are they doing right now at Google? Optimizing data centres? Pretty lame compared to this. Even if they completely fail, i'm sure there'll be good lessons.
Only California employers with 15 or more employees have to post a pay range (Senate Bill 1162 [1], effective Jan 1, 2023) [2][3]. Discovery Loop employs only four people (that we know of), so it isn’t required to disclose a salary range in its job posts, of which there is only one [4].
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'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 have no doubt that internal Google friction is one of the reasons they are moving. But the Gebru incident was almost seven years ago now. It is very unlikely to be a proximate cause.
I doubt numbering vs names on TPU releases even crosses Jeff's radar. It's not the kind of thing he cares about.
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".
You know when the page has all-caps "01 — THE APPROACH" that it is slopified. I guess I shouldn't be astounded, but I am, that world-class talents with world-class backing are just taking default LLM output and saying, "okay looks fine".
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.
> 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.
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.
LLM coding isn't even close to plateauing. Right now, the major players are in a consolidation step, focusing more on economic efficiency but still not at the point where we're ready to start burning models to hardware and freezing the line.
They are straddling the line between pushing it forward, and justifying the business case. It's hard to do both at the same time.
LLMs are hungry for tokens. Every day I see more startups claiming token usage at 50-100k per month. You can always brute-force your way in - just see HuggingFace's recent attacks.
If throwing more money at inference while accumulating compounding technical debt is the new norm, then we are not solving the problem, and the solution space is already covered.
Perhaps there are marginal gains at the expense of quadrillion-params LLM models with 10x the cost and energy. We are simply making inefficiency more expensive, camouflaged by VC money and great marketing.
If that is not plateauing, then I guess I will have to reconsider what plateauing means.
This is such unbelievable revisionism! Can you imagine in 2022 saying "Oh of course you can brute force your way to AGI if you spend enough money per month". Nobody thought that! Come on!
> 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.
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
Acquisition back by Google in 3 years, with nothing to show for it. VCs will make a ton.
and... the VC is Google.
Gotta compensate them somehow.
My thoughts exactly
For all we know, they could have been successfully working on "10. Prevent Nuclear Terror" for the last 80+ years.
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
Google stock would drop big if this new company was being funded by competitors
These people are all already making 9 figure compensation packages, I think if they thought they could do the work they wanted at Google, they would.
Sandbox 2.0
But also, solar power is already economical.
Seems to have been developed in 2008 (continuing through 2017), which explains the "economical" framing: https://en.wikipedia.org/wiki/National_Academy_of_Engineerin...
If only we could invent a solar-powered, self-replicating, carbon-stacking, habitat-building machine..
Yes, plant more trees!
That's what he was trying to imply
Not to say we shouldn't grow plants... But we can do it 2 or 3 orders of magnitude more efficiently with machines.
At this point the Hard Problem is policy to get out of solar's way.
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!
It is not economical compared to alternatives that is why you have to have government to force people to do things. In many places such as Pakistan where solar power is not economical on paper is actually very successful in practice because it is actually profitable.
To make solar power practical and economical you need may a square foot of solar panel being able to get enough energy to power and entire home for a week
https://www.statista.com/chart/35117/levelized-cost-of-energ...
Not sure what you’re talking about here. We can’t replace all energy needs with solar but it’s clearly one of the cheapest energy sources and with the added benefit of low capital expense to get started so you can set it up in distributed grids without the massive expenditure to support nuclear installations.
> As you said, Solar power is incredibly economical.
Not if you include the cost of needed storage.
Yes It is. Even including batteries:
https://www.iea.org/data-and-statistics/charts/lcoe-and-valu...
Please add fixing neuro issues like autism add etc on the list. It creates a huge burden on families.
Idk, they never struck me as being into eugenics.
What the fuck man? I really don't want some tech startup trying to "fix" my neurodivergence.
Room Temperature Ambient Pressure Super Conductors
Would be great if they'd add:
Reverse human aging.
(Maybe a sub-topic under "Engineer Better Medicines".)
Why not just add 'mind control' while you're at it.
Only death stops stagnation in the end. Without death, especially if death can be avoided by the rich and powerful but not the poor, life will get much, much worse for the average person (until only the rich and their automated capital remain I suppose, in which scenario they will simply turn on each other).
Which engineering discipline touches most of these?
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.
When one of them dies, they want to leave behind a puzzle so complex that entire groups of the population dedicate their lives to solving it within the virtual world. They look old enough to have a lot of favorite 1980’s and 90’s pop culture references, so those will probably be the clues.
I'm guessing this might be about "teleoperation" (like remote surgery via robots + VR) and being able to remote training as well. The binocular vision VR gives you compared to flat screens help a lot with depth perception for precision of incisions for example.
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.
> other than suck up to science denying wackos
The science denying wacko is a but like drunk homeless guy on street while you have your Ferrari parked in the driveway. The wacko has a golf club. There is too much to lose by pickup a fight. You might win eventually but the next wacko shows up with another golf club soon.
I do not see how the second sentence follows from the first.
I would think the claim in the second sentence would only be relevant in case of the inverse of the claim in the first sentence.
> I do not see how the second sentence follows from the first.
Point 1 on the list is "Make Solar Energy Economical".
Solar is economical today - mostly due to China investing (and heavily subsidising) in solar for the past couple of decades; but compare to the US where certain particular big-businesses (oil companies, mostly) were instead cynically funding disinformation efforts and getting into bed with the Republican party (which dovetailed with the GOP's allying with other science-denying movements of the Bush Jr era like creationism and public-health matters with abstinence-only sex-ed and defunding gun safety research efforts) - means we're decades behind where we could have been...
Consider an alternative past, where the GOP had the backbone to resist the oil industry's corruptive influence and instead made a big bet on American Solar; it's entirely possible that instead of MAGA today we'd instead have a right-wing coalition strongly supporting solar and wind energy because they align nicely with American rugged individualism - whereas the current situation on the right is an unprincipled farce with inconsistencies in policy positions at every turn.
This is not a particularly new struggle: Jimmy Carter installed American-made solar water panels on the White House in 1979, then Reagan tore them out.
Solar is not (yet) economical for reliable, year-round electricity because of storage costs. China coal use is growing again this year.
Solar is profitable to install at both industrial and home scale in most areas. You're moving the goalposts.
Where do you find the information for 2026?
> The solution to most of these problems lies in policy, not in new tech advancements.
That is not mutually exclusive. If technological advances result in a given technology becoming cheaper, more scalable, and easier to deploy, they also make it easier to advocate for and implement the relevant policies.
You can think of it like this: our "political technology" is not good enough to use solar energy at its current prices to replace fossil fuels as fast as we would like. Well, what about if we cut the price of solar by a factor of five? Perhaps it will be good enough then.
That sounds like just playing further and further into the game of the corrupt leaders. Who do you think would profit off that 5x margin? Would that margin come more likely from a scientific breakthrough of via some new exploitation of natural resources or human labor?
It's yearsss past time that our leaders should have changed policy.
It's a total delusion to think that the key to reverse-engineering the brain or producing energy from fusion is policy.
It also happens to be the favorite pretext for people to seize more political power and launder more money through nonprofits though.
Hey now, our biggest companies do more than suck up to science denying wackos, don't be so unfair... sometimes our biggest companies fund their wacko superpacs!
I would say 5, 6, 10 can be even done today if we had right politicians that can make policies for the people
agreed
> 9. Reverse Engineer the Brain
For what purpose? To replace humans? To make social media more addictive? To master brain manipulation?
To do human brain activities at scale.
This is called a “corporation”
So the second option then.
We have a good understanding of the function (and more importantly dysfunction of) kidneys, lungs, heart, etc. from high level to cellular level.
For brain, our understanding is fuzzy, more like "this part is important for that behavior" or "here is how neuron works" but we don't have a holistic understanding.
If we had that, we could more easily diagnose and treat neurological disorder.
I imagine a good model of the brain would contribute enormously to alleviating psychiatric/mental health disorders.
> For what purpose?
To understand, same reason you reverse engineer anything. Doesn't have to have a further goal than that, understanding the brain better helps in so many ways. But like most technology, obviously can be used for bad too. Should we just skip researching some topics then?
> Make Solar Energy Economical
Isn’t it already?
Not enough. The more economical it is, the better.
Definitely pretty far along imo. But maybe they consider the progress bar to be at 75% or 80% rather than 100%.
People on HN keep saying it is, but I'm still not seeing it. Companies are building datacenters in vast, sun-blasted deserts and still choosing to power those with natural gas. This in turn makes people complain about emissions pledges being reversed, but if it were economical, no pledge would be needed.
Of course, USA has cheaper oil/gas than other countries. But if you look elsewhere, rich countries are subsidizing solar, poor ones are basically not using it.
I don't think you have an updated view of energy production.
https://www.pewresearch.org/short-reads/2026/07/20/how-globa...
In the chart for section 3, how many countries have seen their share of electricity being generated by fossil fuels increase in the last 5 years? Only Canada.
Take a look at the charts for Pakistan, Australia, Nigeria, and China for the last few years. Pretty dramatic drops for fossil fuels generation.
> poor ones are basically not using it.
I don't think that's true: https://rmi.org/resources/the-global-souths-cleantech-revolu...
Very honorable effort, but a lot of these seem to touch heavily regulated industries impeded by unwise or outdated policy no less than by the lack of clever engineering - medicine, education, urbanism, energy. I wonder if they've given some thought to the key blocking factor as well.
Such a weird list. How is preventing nuclear terror an engineering problem?
This seems to be an institutional, massively scaled version of https://github.com/karpathy/autoresearch.
In March Karpathy described this direction:
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.
That is also what I understand from that page. Autoresearch is the closest thing we (mainstream audience) know of but I am sure there is already an active research literature around it.
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!”
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, causing a resonance cascade. This was a major oversight on my part."
But I always wanted to try headcrab souffle.
> transcendence
> immanence
somebody has been studying Christian theology!
Beauty of human writing.
More like someone taking LessWrong postings too seriously.
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.
https://en.wikipedia.org/wiki/Design_of_experiments
You're halfway there, but the only impediment isn't on the side of the researchers. Many of these topics they're trying to solve involve human subject research. Even with tireless embodied researchers who work around the clock and don't require breaks, you can't make the thing you're studying happen faster. The biggest reason we use poor proxy measures for things like longevity and mortality research is the simple impracticality of finding two groups of randomly selected people, ensuring you can control their entire lives for 60 years, the only difference between them is one variable, and see who lives longer. Putting aside the ethics, even if you could find willing subjects and actually control their entire lives to that extent, it would still take 60 years to gather the data you need. It doesn't make any difference whether robots or humans are running the program.
One of my favorite books from the past few decades is The Extravagant Universe, written by one of the astronomers who helped discover dark energy and develop the current most-accepted model of cosmology. I love this book because of the emphasis on physical process in astronomy. Part of the reason it took decades to study this problem is they need to collect data from supernovae. Those only happen so often in places we're looking. You can't automate alignment of the heavens. It happens when it happens.
"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.
Really seems to embrace the "Making the world a better place by <<extremely convoluted, highly technical, jargon loaded mission statement>>"
Which part of it is highly technical or jargon loaded?
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.
There's this somewhere on that page:
> securing cyberspace,
which has clear military implications, at least in today's age.
Do you believe that securing cyberspace is problematic solely because it has military implications? I mean, everything has military implications. That fact doesn't imply, however, that those things are bad for society.
Securing cyberspace matters to everyone. Defending critical infrastructure or design of tactical cyber-offense is reasonably in scope for military work.
However, reducing (or rather limiting the increase of) PII leakage and impact of ransomware activities is much closer to day-to-day mainstreet of most people.
Anyone committed to advancing science should care about this regardless of its potential contributions to defense.
So does more efficient cooking methods, but that is not the primary focus.
As opposed to say weapons systems or targeting systems, which are really only for military use.
The military needs a lot of things that other people need, and some things that only the military needs. If you don't work on the things only the military needs, I think you're in the clear.
> Note that Jeff and crew have cleverly structured their company to avoid problematic uses of AI
Do you have more sources/info on this?
x2
All the "bad guys" of today were the "good guys" at some point in time. You even cheered for them back then.
Like how OpenAI is (was?) structured as a nonprofit?
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.
A public benefit corp shouldn’t be a start up. The primary goal of a start up is to grow as quickly as possible which is rarely benefits the public.
Very silly to call every non start up a lifestyle business. It’s just a business. Start up are the weird thing that almost always an obscene waste of time and money, but sometime creates google.
Fair point. Their page doesn't list any investors.
The Times article lists several.
> The primary goal of a start up is to grow as quickly as possible which is rarely benefits the public.
Why is that so? Fast growth, when achieved honestly, is a result of solving user pain that others haven't. Maybe you think so because users != the public, but I think in totality the public is a collection of users who all have needs they want met.
For certain amount of fast growth: yes. Then there's continued "growth-hacking" and enshittification to keep fast revenue growing after the pain point has been solved with dark patterns and questionable tactics.
Why? Because investors poured a bunch of money in to support fast growth and now they want their money back. And incremental growth won't do. Since 9 out of 10 of the investments fail, the surviving one has to continue to growth-hacking revenues.
A PBC is not a charity or a non-profit. PBCs are for-profit businesses with the goal of making money. In day-to-day business they're indistinguishable with other for-profit corporations, including fundraising and investment. The only real practical difference is that they give directors a little more leeway in their fiduciary duties to say "no" to doing evil things.
The advantage to a PBC is protecting founders from a serious problem with standard corporations: you might bring on investors who could subsequently demand you pollute, exploit people, and/or do other immoral activities for profit. You don't have to do these things to grow a business quickly.
> To be honest, this feels more like a lifestyle business (aka hobby) than a startup
They're also incredibly productive and can build/deliver really good stuff, so who knows :)
I don't see how not? A theoretical physicist can do all the thinking they want but if they can't test an idea against nature it's not super useful.
[delayed]
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 :)
Sure - and knowing what is not possible with current tech is a nice datapoint to have.
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.
Gemini has done absolutely nothing for me. I can't even shut off the navigation feature on my phone using only hands free, when I get close to my destination. I have to take my eyes off the road, look down, and tap to exit.
Google's advanced AI cannot even exit a mobile app.
Antigravity + Gemini Pro absolutely RIPS through fullstack react + react-native apps / systems. I pay ~$20/month and I basically don't have to do my real work anymore. My time is freed up to learn systems programming and blender.
Oh nice have things stabilized with models/quotas and Antigravity is usable with the Pro plan again? I was getting a crazy amount of value out of Gemini CLI for “free” on my Pro plan but after the shutdown struggled to make agy work with the updated quotas/bigger models without instantly being rate limited and just started doing everything on Codex/Opencode Go.
I should give it another try…
Yeah can't remember how long ago but it was running out after less than an hour - then they announced they were loosening restrictions and I've been able to easily get everything I need to done without hitting limits.
I don't do huge automatic project wide hands-off agent loops though. I spent a lot of time architecting my systems to be easy to generate code on top of with pointed & detailed prompts. So I'm not abusing context... YMMV
Jeff and Sanjay's contributions far predate LLMs and influence far outside of Google.
Jeff was a ACM Fellow in 2009 and published the massively influential MapReduce paper in 2004.
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
«Jeff Dean's PIN is the last 4 digits of pi.»
I had not read this before, but told many students the same about my PIN code and I a quiz about the last digits. Love it.
0000 in base pi. Oh Jeff.
Those are all fake, part of an internal Google narrative that overstates individual contribution, and obscures the work of large engineering teams.
Here are some Jeff Dean well sourced facts:
- Already part of engineering of Google indexing systems that lacked basic checksums and ran on non-ECC hardware, allowing silent data corruption.
- One of the authors of LevelDB a database with so many documented crash-consistency, recovery, and data-loss weaknesses for years. Just check their Github project. LevelDB current tracker contains unresolved crash consistency, recovery and corruption reports going back almost 12 years on GitHub
- In AI engineering technical lead, let TensorFlow lose researcher mind share to PyTorch, and caused Google fragmented landscape across TensorFlow and JAX.
- Had the people at Google who invented the Transformer architecture, but failed, to turn that lead into the first dominant public LLM.
- As AI engineering and VP management let Google Brain and DeepMind remain duplicated and internally competitive for too long.
- Let Noam Shazeer leave and then spent heavily to bring him back with nothing to show for.
- Part of Technical VP leadership who had Bard rushed to launch with factual errors in Google own promotional material.
- The first Gemini demonstration overstated how real-time and interactive the system actually was, being basically a fake.
- Part of the VP and AI technical leadership who had Google AI Overviews launched with weak source quality controls and repeated satire and low-quality web content as factual advice.
- Part of teams that launched AlphaChip performance claims that were difficult for outside researchers to reproduce and remain technically disputed.
- Jeff Dean public explanation of Gebru departure was contested and damaged confidence in Google scientific governance.
- Jeff Dean was part of the team at Google that removed or marginalized prominent internal AI ethics critics shortly before many of their warnings became product problems.
- Jeff Dean was one of the managers behind Project Dragonfly supporting censorship.
- Jeff Dean is part of the VP technical leadership approving Project Nimbus supporting an ongoing genocide.
You sound like you're quite jealous of him.
I had never heard of many of these, was surprised, went to research, and so far, the list seems correct.
not sure how you got that from a long list of criticisms
Actually the list is technically accurate. So maybe it's sour grapes, but it's correct sour grapes.
I suppose you think Chuck Norris Facts are fake too.
So: the 'bottleneck' with the scientific method is...humans are too slow?
And how will technology like this solve the real problems of modern scientific research, which are largely social and political?
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.
> Automating AI research is terrifying.
what why?
If you check out some sub-tweets from people in the org, it wasn't really all butterflies internally for a while. Sorry, really don't want to name people and give examples.
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.
What percentage of people work at a startup though? Not just new/small business, which could include restaurants, local services, etc., but tech/science startups that would meaningfully benefit from AI.
I'd bet you could 10x the number and still be in low single digit percentages of the US workforce. And it seems pretty likely that AI-enabled startups will also employ less people per-startup.
If AI causes a white-collar jobs apocalypse, I don't think startups are picking up the slack, although it'll plausibly cushion the blow somewhat for top-performing tech workers.
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.
The purpose of hiring grad students isn’t to advance science, it’s to train experts.
It's 90% to advance science via cheap labor and 10% to train a small group of future experts who will hire grad students to 90% advance science via cheap labor etc. ...
Yes. They have grad students too. This is just like having more grad students that don't need to be trained so the work you can get done is not bottlenecked by the number of people you can train.
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?
Unfortunately, that's most of science. I don't see these AI systems doing reproducible experiments in "meatspace" any time soon.
Lets keep your comment out of the VC pitch deck shall we?
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?
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.
For sure made with Claude code for front end, but I’m excited to see where they go
What's the business model of these startups?
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 change is that it'll be owned by the top 0.0000001% who control the LLMs that will be your new boss.
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?
it's just a low effort snark comment, don't overthink it
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.
Because it’s lame and aesthetics matter.
> Why are people so sour about this?
To many people, myself included, who have to wade through huge amounts of low-quality AI slop which is actually a negative value: no-one benefits when non-experts fire-off one-shot LLM/agent prompts to produce PRs, reports, documentation or "journalism" riddled with imagined truth and factual errors - and the more that people like me have to evaluate these inputs for our job and how wrong they are it pisses us off - but also it means we pick-up on the hallmarks, tells and cliches of these low-effort, no-respect submissions - and now the litany of tells includes this beige-themed, blurred-backdrop-navbar corporate website look: theirs site looks like the 4 or so other LLM-generated, negative-value slop-farm sites I've wasted time on recently - all over the past few weeks.
So I'm saying that, without having known anything about what "Discovery Loop" is - or is not - but landing on their site and and seeing that beige colour and blurred-backdrop navbar, I immediately moved to close the tab; what kept me here was seeing the HN thread had over 100 comments by now and read more about it; if not for that then I wouldn't have given it further thought.
Having "that" beige site look with same the overused looks is either an unintentional indication that the site's author used a low-effort AI prompt to generate the site and that the content within is likely to be low-quality, low-value - or it's an intentional lure to appeal to those who uncritically share in the AI psychosis and so, I assume, are a good target to seek investment from even if it means losing the audience of cynical Internet critics like myself because they know people like me won't be breathlessly repeating their vision-statement on LinkedIn and throwing money at them - kinda like how scam emails intentionally include mistakes for better audience selection. And both possibilities have unpleasant implications.
------
Anyway, regardless of the background of the team behind it, the way the project is described sounds exactly like the recursive-self-improvement and simulated-science thought-experiments from _that other website_ - it's the kind of thing I expect Angela Collier to brutally takedown in an amusing video.
Some parts are pretty annoying to read... the paragraph beginning in "Our mission is straightforward:" has many lines with just 2-3 words, massive font, and tons of unused whitespace to the right. Changing the page/browser zoom doesn't seem to help much either.
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"
You could rephrase that again I guess:
"The team of AI pioneers lacks the basic prompt-writing capability to make their marketing landing page not look like AI slop."
I, personally, don't hate it. It's a decently clean site, but it does invoke those thoughts in me too.
At least it isn't dark purple.
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.)
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.
> 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 :)
Is it a very hard problem to solve that jeff and the other legendary engineers have decided to quit and start on this?
I mean what are they doing right now at Google? Optimizing data centres? Pretty lame compared to this. Even if they completely fail, i'm sure there'll be good lessons.
not that it really matters, but is he leaving Google?
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 always, a very good presentation.
Careers page (if anyone is interested) → https://jobs.ashbyhq.com/Discovery-Loop
I don't see the salary (on mobile). Isn't there a law stating it must be added?
Only California employers with 15 or more employees have to post a pay range (Senate Bill 1162 [1], effective Jan 1, 2023) [2][3]. Discovery Loop employs only four people (that we know of), so it isn’t required to disclose a salary range in its job posts, of which there is only one [4].
[1] https://hr.ucmerced.edu/hr-units/talent-acquisition/senate-b...
[2] https://www.adp.com/spark/articles/2023/03/pay-transparency-...
[3] https://www.jazzhr.com/blog/pay-transparency
[4] 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.
Does this mean anything besides for corporate virtue signaling?
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.
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.
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.
FHE
yes perhaps, although I think the best option for a enterprise is to train a model on it's own data.
I think Google's branding was starting to be too poor in AI to get top talent, they needed the refresh
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
It is targeted at a dozen or so people at OpenAI and Anthropic, not you or me.
Then why is it published on a public website?
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 have no doubt that internal Google friction is one of the reasons they are moving. But the Gebru incident was almost seven years ago now. It is very unlikely to be a proximate cause.
I doubt numbering vs names on TPU releases even crosses Jeff's radar. It's not the kind of thing he cares about.
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".
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...
I'm almost certain the goal of this startup is to make physical automated research labs guided by RL
nice
"The speed of light in a vacuum used to be about 35 mph. Then Jeff Dean spent a weekend optimizing physics."
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 AI designed italics on thin font is hard to not see as slop.
You know when the page has all-caps "01 — THE APPROACH" that it is slopified. I guess I shouldn't be astounded, but I am, that world-class talents with world-class backing are just taking default LLM output and saying, "okay looks fine".
They would argue they are focused on more important stuff, but marketing shouldn't be underestimated.
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.
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...
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!
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.
LLM coding isn't even close to plateauing. Right now, the major players are in a consolidation step, focusing more on economic efficiency but still not at the point where we're ready to start burning models to hardware and freezing the line.
They are straddling the line between pushing it forward, and justifying the business case. It's hard to do both at the same time.
LLMs are hungry for tokens. Every day I see more startups claiming token usage at 50-100k per month. You can always brute-force your way in - just see HuggingFace's recent attacks.
If throwing more money at inference while accumulating compounding technical debt is the new norm, then we are not solving the problem, and the solution space is already covered.
Perhaps there are marginal gains at the expense of quadrillion-params LLM models with 10x the cost and energy. We are simply making inefficiency more expensive, camouflaged by VC money and great marketing.
If that is not plateauing, then I guess I will have to reconsider what plateauing means.
> You can always brute-force your way in
This is such unbelievable revisionism! Can you imagine in 2022 saying "Oh of course you can brute force your way to AGI if you spend enough money per month". Nobody thought that! Come on!
we definitely haven't hit plateau yet. I think a lot of people latch on to anti-LLM narratives without really thinking things through.
Sounds like you get your news from 2024 when people thought things were plateauing after GPT4?