I suspect current approaches probably significantly or even drastically under-elicit automated ML research capabilities.

I'd guess the average cost of producing a decent ML paper is at least 10k$ (in the West, at least) and probably closer to 100k's $.

In contrast, Sakana's AI scientist cost on average 15$/paper and .50$/review. PaperQA2, which claims superhuman performance at some scientific Q&A and lit review tasks, costs something like 4$/query. Other papers with claims of human-range performance on ideation or reviewing also probably have costs of <10$/idea or review.

Even the auto ML R&D benchmarks from METR or UK AISI don't give me at all the vibes of coming anywhere near close enough to e.g. what a 100-person team at OpenAI could accomplish in 1 year, if they tried really hard to automate ML.

A fairer comparison would probably be to actually try hard at building the kind of scaffold which could use ~10k$ in inference costs productively. I suspect the resulting agent would probably not do much better than with 100$ of inference, but it seems hard to be confident. And it seems harder still to be confident about what will happen even in just 3 years' time, given that pretraining compute seems like it will probably grow about 10x/year and that there might be stronger pushes towards automated ML.
 

This seems pretty bad both w.r.t. underestimating the probability of shorter timelines and faster takeoffs, and in more specific ways too. E.g. we could be underestimating by a lot the risks of open-weights Llama-3 (or 4 soon) given all the potential under-elicitation.

NVIDIA might be better positioned to first get to/first scale up access to AGIs than any of the AI labs that typically come to mind.
They're already the world's highest-market-cap company, have huge and increasing quarterly income (and profit) streams, and can get access to the world's best AI hardware at literally the best price (the production cost they pay). Given that access to hardware seems far more constraining of an input than e.g. algorithms or data, when AI becomes much more valuable because it can replace larger portions of human workers, they should be highly motivated to use large numbers of GPUs themselves and train their own AGIs, rather than e.g. sell their GPUs and buy AGI access from competitors. Especially since poaching talented AGI researchers would probably (still) be much cheaper than building up the hardware required for the training runs (e.g. see Meta's recent hiring spree); and since access to compute is already an important factor in algorithmic progress and AIs will likely increasingly be able to substitute top human researchers for algorithmic progress. Similarly, since the AI software is a complementary good to the hardware they sell, they should be highly motivated to be able to produce their own in-house, and sell it as a package with their hardware (rather than have to rely on AGI labs to build the software that makes the hardware useful). 

This possibility seems to me wildly underconsidered/underdiscussed, at least in public.

I don't have a strong opinion about how good or bad this is.

But it seems like potentially additional evidence over how difficult it is to predict/understand people's motivations/intentions/susceptibility to value drift, even with decades of track record, and thus how counterfactually-low the bar is for AIs to be more transparent to their overseers than human employees/colleagues.

I think I agree directionally with the post.

But I've been increasingly starting to wonder if software engineering might not be surprisingly easy to automate when the right data/environments are used at much larger scale, e.g. Github issues (see e.g. D3: A Large Dataset for Training Code Language Models to Act Diff-by-Diff) or semi-automated pipelines to build SWE RL environments (see e.g. Skywork-SWE: Unveiling Data Scaling Laws for Software Engineering in LLMs), which seem potentially surprisingly easy to automatically scale up. It now seems much more plausible to me that this could be a scaling data problem than a scaling compute problem, and that progress might be fast. Also, it seems likely that there might be some flywheel effect of better AIs -> better automated collection + filtering of SWE environments/data -> better AIs, etc. And 'Skywork-SWE: Unveiling Data Scaling Laws for Software Engineering in LLMs' has already shown data scaling laws:

Also, my impression is that SWE is probably the biggest bottleneck in automating AI R&D, based on results like those in Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers
and especially based on the length of the time horizons involved in the SWE part vs. other parts of the AI R&D cycle. 

It's probably also very synergystic with various d/acc approaches and measures more broadly.

Intuitively, the higher the (including global) barriers to AI takeover, human extinction, etc., the more likely the AI will have to do a lot of reasoning and planning to have a decent probability of success; so the more difficult it will be to successfully achieve that opaquely and not be caught by CoT (and inputs, and outputs, and tools, etc.) monitors.

And this suggests 100x acceleration in research cycles if ideation + implementation were automated, and humans were relegated to doing peer reviewing of AI-published papers:

https://x.com/BogdanIonutCir2/status/1940100507932197217 

I think it might really be wise to use a canary string, or some other mechanism to hide this kind of knowledge from future (pre)training runs, e.g. https://turntrout.com/dataset-protection 

Based on current trends fron https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/, this could already have happened by sometime between 2027 and 2030:

Excited about this research agenda, including its potential high synergy with unlearning from weights:

https://www.lesswrong.com/posts/9AbYkAy8s9LvB7dT5/the-case-for-unlearning-that-removes-information-from-llm.

Intuitively, if one (almost-verifiably) removes the most dangerous kinds of information from the weights, the model could then only access it in-context. If that's in the (current) form of text, it should fit perfectly with CoT monitorability.

My guess is that on the current 'default' ML progress trajectory, this combo would probably buy enough safety all the way to automation of ~all prosaic AI safety research. Related, this other comment of mine on this post:

https://www.lesswrong.com/posts/7xneDbsgj6yJDJMjK/chain-of-thought-monitorability-a-new-and-fragile?commentId=eqidvTEaom4bmmSkv