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.
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.
Yeah, the reason I don't see it how it would be successfull is that most public labs are usually struggling with money, so it would definitely cost them less to build their own automated pipeline.
I'm skeptical of any Engineering loop that doesn't include reality (as in touch grass) feedback. Pure logic and reasoning is the domain of Maths and Science (philosophy). Surely it will work, but it will not "be able to solve any learning loop".
I’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.
> 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.
> 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.
This reads as hype of the kind: "well, we can't reliably do these rather mundane things with AI, but we're going to run it extra hard and extra long in some novel way, and it will do amazing things".
The issues are with verification and with detecting drift from the goal. These are related, if not roughly the same issue. And, if they can solve this, then they will have essentially fixed AI. Maybe even AGI.
But, if this were the goal, then it seems more reasonable to solve the relatively more mundane verifiable challenges (e.g. generating solid, reliable code). Then, working up from there.
And, that's exactly what gives this the hype smell. No use for solving problems that don't get the oohs and aahs. Just straight to NAE Grand Challenge problems.
1. Make Solar Energy Economical - Disallow fossil fuels
2. Provide Energy from Fusion - See 1
3. Develop Carbon Sequestration Methods See 1
4. Manage the Nitrogen Cycle - See 1
5. Provide Access to Clean Water - See 1
6. Restore and Improve Urban Infrastructure - See 1
7. Advance Health Informatics - See 1
8. Engineer Better Medicines - See 1
9. Reverse Engineer the Brain - See 1
10. Prevent Nuclear Terror - See 1
11. Secure Cyberspace - See 1
12. Enhance Virtual Reality - See 1
13. Advance Personalized Learning - See 1
14. Engineer the Tools of Scientific Discovery - See 1
FF is the real threat in time, money, health. Can't sweep aside that it will destroy most life on Earth and we'll never get to the other things if we are at the mercy of FF
Many of those require a paradigm shift in our understanding of Universe, so I am not sure they are achievable with any convex combination of existing knowledge. We'd need some sharp mind connecting the dots but with heavy use of AI and incentives to use it, we might just never give a chance for such mind to arise.
We’ll face a bigger problem sooner rather than later, which is population collapse. Youth don’t procreate amy more. Birth numbers are at an all time low. We see the issue arise in rats and the experiment is all too relevant for the current age of social media and fearmongering (John B. Calhoun’s rodent “utopia” experiment). All these “Grand Challenge” problems seem trivial to that.
Furthermore:
Why is 1 here when 2 is present. Again, solar is usually not relevant when we want power when it’s dark… even theoretical it wouldn’t work. We’d need a high capacity dirt cheap storage, and even then we can’t keep it till winter when there’s no sun to go around and effectively supply
3. Irrelevant when there’s population collapse. And even then, why would we want this instead of reforestation and low depth water protection from fishing and environmental issues.
5. How is this even a problem. Unless we got corrupt(ed/able) governments (read: lobbies) that allow exemptions in every law designed to protect the environment (also, settlements are a twisted way to fill governments pockets instead of rooting out evil)
7/8/9 the inverse effects are even worse health, as everything is fixable. 9 would incur even more social isolation, more so than the internet did
10 seems to be the first that’s actually reasonable
Same for 11
For 12 see 9
13 yes but that’s something that a self learner would already be able to do. AI is at a level that we can manage
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.
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.
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.
143 comments
[ 0.25 ms ] story [ 11.5 ms ] thread[1] https://github.com/LRitzdorf/TheJeffDeanFacts
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.
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.
https://www.ycombinator.com/library/Vy-jeff-dean-the-1-rule-...
holy shit. I've known this, but...
> Our general approach is to automate the experimental loop. We think this approach is broadly applicable across many different fields of science and engineering. We’ll initially focus on ML research and engineering, but believe the approach can help with important subproblems in nearly every one of the fourteen <at>NAE Grand Challenge problems. We think doing this well requires strong expertise in machine learning as well as large-scale systems.
See also: https://www.nae.edu/20782/grand-challenges-project
Those 14 are:
NAE Grand Challenges for Engineering
1. Make Solar Energy Economical
2. Provide Energy from Fusion
3. Develop Carbon Sequestration Methods
4. Manage the Nitrogen Cycle
5. Provide Access to Clean Water
6. Restore and Improve Urban Infrastructure
7. Advance Health Informatics
8. Engineer Better Medicines
9. Reverse Engineer the Brain
10. Prevent Nuclear Terror
11. Secure Cyberspace
12. Enhance Virtual Reality
13. Advance Personalized Learning
14. Engineer the Tools of Scientific Discovery
The issues are with verification and with detecting drift from the goal. These are related, if not roughly the same issue. And, if they can solve this, then they will have essentially fixed AI. Maybe even AGI.
But, if this were the goal, then it seems more reasonable to solve the relatively more mundane verifiable challenges (e.g. generating solid, reliable code). Then, working up from there.
And, that's exactly what gives this the hype smell. No use for solving problems that don't get the oohs and aahs. Just straight to NAE Grand Challenge problems.
Obvious near term trillions dollar market to disrupt.
2. Provide Energy from Fusion - See 1
3. Develop Carbon Sequestration Methods See 1
4. Manage the Nitrogen Cycle - See 1
5. Provide Access to Clean Water - See 1
6. Restore and Improve Urban Infrastructure - See 1
7. Advance Health Informatics - See 1
8. Engineer Better Medicines - See 1
9. Reverse Engineer the Brain - See 1
10. Prevent Nuclear Terror - See 1
11. Secure Cyberspace - See 1
12. Enhance Virtual Reality - See 1
13. Advance Personalized Learning - See 1
14. Engineer the Tools of Scientific Discovery - See 1
FF is the real threat in time, money, health. Can't sweep aside that it will destroy most life on Earth and we'll never get to the other things if we are at the mercy of FF
2-14. ???
16. Make Everyone Nice
17. Finally Impress a Girl
This one is already solved, right? The price of panels and batteries is on trend to displace all other forms of power generation within our lifetime
- Eliminate racism
- Eliminate poverty
- Eradicate crime
- Eradicate corruption
- Reverse climate change 100%
- wake me up when you got an AI project capable of doing this one
I only see "co-created" "co-founded" "managed a team" - what did he actually do?
That is already solved.
> Develop Carbon Sequestration Methods
That is not necessary, because 1 is solved.
> Reverse engineer the brain
What for? There was already the european human brain project, which didn't do anything useful.
> Prevent nuclear terror
Easy one: Every country stops developing nuclear weapons and destroys existing ones.
It seems this list itself has many flaws. Maybe we need a bigger computer which figures out the questions we really need to ask.
1. Make Solar Energy Economical — https://github.com/orgs/HardisonCo/projects/194
2. Provide Energy from Fusion — https://github.com/orgs/HardisonCo/projects/206
3. Develop Carbon Sequestration Methods — https://github.com/orgs/HardisonCo/projects/196
4. Manage the Nitrogen Cycle — https://github.com/orgs/HardisonCo/projects/197
5. Provide Access to Clean Water — https://github.com/orgs/HardisonCo/projects/195
6. Restore and Improve Urban Infrastructure — https://github.com/orgs/HardisonCo/projects/193
7. Advance Health Informatics — https://github.com/orgs/HardisonCo/projects/203
8. Engineer Better Medicines — https://github.com/orgs/HardisonCo/projects/200
9. Reverse Engineer the Brain — https://github.com/orgs/HardisonCo/projects/205 10. Prevent Nuclear Terror — https://github.com/orgs/HardisonCo/projects/204
11. Secure Cyberspace — https://github.com/orgs/HardisonCo/projects/201
12. Enhance Virtual Reality — https://github.com/orgs/HardisonCo/projects/202
13. Advance Personalized Learning — https://github.com/orgs/HardisonCo/projects/199
14. Engineer the Tools of Scientific Discovery — https://github.com/orgs/HardisonCo/projects/198
Specs and the per-challenge process lists: https://github.com/HardisonCo/opendl
It should also be funded by the Gov. IMO and 100% for oublic benifit e.g.: nsf.dev
We’ll face a bigger problem sooner rather than later, which is population collapse. Youth don’t procreate amy more. Birth numbers are at an all time low. We see the issue arise in rats and the experiment is all too relevant for the current age of social media and fearmongering (John B. Calhoun’s rodent “utopia” experiment). All these “Grand Challenge” problems seem trivial to that.
Furthermore: Why is 1 here when 2 is present. Again, solar is usually not relevant when we want power when it’s dark… even theoretical it wouldn’t work. We’d need a high capacity dirt cheap storage, and even then we can’t keep it till winter when there’s no sun to go around and effectively supply 3. Irrelevant when there’s population collapse. And even then, why would we want this instead of reforestation and low depth water protection from fishing and environmental issues.
5. How is this even a problem. Unless we got corrupt(ed/able) governments (read: lobbies) that allow exemptions in every law designed to protect the environment (also, settlements are a twisted way to fill governments pockets instead of rooting out evil) 7/8/9 the inverse effects are even worse health, as everything is fixable. 9 would incur even more social isolation, more so than the internet did 10 seems to be the first that’s actually reasonable Same for 11 For 12 see 9 13 yes but that’s something that a self learner would already be able to do. AI is at a level that we can manage
/rant
Sometimes I couldn't resist wondering if I'll ever do work that has a tenth of the impact of theirs.
Models are commodities the applications eg. BaseTen, OpenRouter should capture the value.
https://taikhooms.substack.com/p/why-openrouter-can-be-the-n...