An interesting adversarial attack at KataGo, a professional level Go AI. Apparently funded by Fund for Alignment Research (FAR). Seems to be a good use of fund.
AI 2027 Compute Forecast basically completely ignores China for Compute Production section and I don't think it can be justified. This paper from Huawei is a timely reminder.
I used to think while OpenAI is pretty deceitful (eg for-profit conversion) it generally won't lie about its research. This is a pretty definitive case of lying, so I updated accordingly. I am posting here because it doesn't seem to be widely known.
Recently automated reasoning system was developed to solve IMO problems. It is a very impressive and exciting advance, but it must be noted IMO problems are to be solved in two days. Recently it was also proposed to make use of this advance to greatly automate formal verification of practical...
Comment deadline is June 10, 2023.
The announcement is of obvious importance to global AI governance. As I understand, you can email your comments to wajscy@cac.gov.cn, and I recommend everyone to do so.
Finbarr Timbers makes a point, obvious in retrospect, but which many people, including people forecasting AI timeline, seem to miss: since training cost is amortized over inference, optimal training depends on expected amount of inference. Both scaling laws from OpenAI and DeepMind assume zero (or negligible) inference, which is obviously...
> We performed a blind pairwise comparison between text-davinci-003 and Alpaca 7B, and we found that these two models have very similar performance: Alpaca wins 90 versus 89 comparisons against text-davinci-003. Interestingly, Alpaca is trained using supervised finetuning, not RLHF. (text-davinci-003 is trained using RLHF.) This seems to confirm my...
OpenSSL is extremely widely used and it is hard to argue with OpenSSL CVEs. On the other hand, I am starting to suspect OpenSSL is a somewhat special case. My understanding is unfortunately for such a widely used codebase, OpenSSL codebase is not in a good state. Someone on Hacker News noted 0/12 of CVEs apply to BoringSSL (Google's OpenSSL fork).
I have much less problems with curl CVEs and I think they are impressive.
The other lab leaders have not commented on the topic in public in 2025.
I don't think this is true. Amodei on AI: "There's a 25% chance that things go really, really badly".
2025-08 update. Anthropic now defaults to (you can opt out) using your chats for AI training, see for example https://techcrunch.com/2025/08/28/anthropic-users-face-a-new-choice-opt-out-or-share-your-data-for-ai-training/
I think IMO results were driven by general purpose advances, but I agree I can't conclusively prove it because we don't know details. Hopefully we will learn more as time goes by.
An informal argument: I think currently agentic software engineering is blocked on context rot, among other things. I expect IMO systems to have improved on this, since IMO time control is 1.5 hours per problem.
I think non-formal IMO gold was unexpected and we heard explicitly that it won't be in GPT-5. So I would wait to see how it would pan out. It may not matter in 2025 but I think it can in 2026.
I think it is important to note that Gemini 2.5 Pro Capable of Winning Gold at IMO 2025, with good enough scaffolding and prompt engineering.
Do you have any Solomonoff inductor you know? I don't, and I would like an introduction.
Ethan Mollick's Using AI Right Now: A Quick Guide from 2025-06 is in the same genre and pretty much says the same thing, but the presentation is a bit different and it may suit you better, so check it out. Naturally it doesn't discuss Grok 4, but it also does discuss some things missing here.
Anthropic does have a data program, although it is only for Claude Code, and it is opt in. See About the Development Partner Program. It gives you 30% discount in exchange.
AI 2027 Compute Forecast basically completely ignores China for Compute Production section and I don't think it can be justified. This paper from Huawei is a timely reminder.
I used to think while OpenAI is pretty deceitful (eg for-profit conversion) it generally won't lie about its research. This is a pretty definitive case of lying, so I updated accordingly. I am posting here because it doesn't seem to be widely known.
Recently automated reasoning system was developed to solve IMO problems. It is a very impressive and exciting advance, but it must be noted IMO problems are to be solved in two days.
Recently it was also proposed to make use of this advance to greatly automate formal verification of practical systems, including advanced AI systems. That's a very tall goal, so some kind of intermediate benchmark would be useful.
I suggest CompCert as a such benchmark. It is a formally verified compiler with manually written proof that compiles a large and practical subset of C programming language to (among other things) x86 assembly language, with optimizations such that performance is competitive with GCC -O1.... (read 213 more words →)
Comment deadline is June 10, 2023.
The announcement is of obvious importance to global AI governance. As I understand, you can email your comments to wajscy@cac.gov.cn, and I recommend everyone to do so.
Finbarr Timbers makes a point, obvious in retrospect, but which many people, including people forecasting AI timeline, seem to miss: since training cost is amortized over inference, optimal training depends on expected amount of inference. Both scaling laws from OpenAI and DeepMind assume zero (or negligible) inference, which is obviously incorrect. Any forecasting using scaling laws similarly is suspect and should be revised.
We performed a blind pairwise comparison between text-davinci-003 and Alpaca 7B, and we found that these two models have very similar performance: Alpaca wins 90 versus 89 comparisons against text-davinci-003.
Interestingly, Alpaca is trained using supervised finetuning, not RLHF. (text-davinci-003 is trained using RLHF.) This seems to confirm my suspicion that while RLHF improves performance it is not essential.
An interesting adversarial attack at KataGo, a professional level Go AI. Apparently funded by Fund for Alignment Research (FAR). Seems to be a good use of fund.
Playing an imperfect information game well has been a challenge for AI. DeepMind reports their (IMHO impressive) work on Stratego, an imperfect information game.
A thought occurred to me, and it's so logical, I concluded that it must be true.
Russia will not detonate a nuclear weapon in Ukraine, fearing repercussion.
Russia will detonate a nuclear weapon in Russia. In other words, Russia will do a nuclear test. Like North Korea did.
As far as I know, there is no international law against Russia doing a nuclear test. Russia is a recognized nuclear weapon state of NPT, unlike North Korea. The test will give weight to Putin's words that he is not bluffing.
I am quite uninformed, but when I read about compute multipliers I considered it to obviously include data-related improvements. To quip, FineWeb-Edu was algorithmically filtered, it obviously wasn't manually curated. As an evidence that it is not just my misunderstanding, I quote Dean W. Ball (my point is that it may well be my misunderstanding, but then such misunderstanding is common):