sanxiyn

• sanxiyn 21 Feb 2026 5:32 UTC

  2 points

  0

  AF

  in reply to: Carl Feynman’s comment on: Why we should ex­pect ruth­less so­ciopath ASI

  Evolution also distinguishes between one and two progeny so it is not binary, but yeah, just a few bits per lifetime.

• sanxiyn 13 Feb 2026 4:16 UTC

  5 points

  0

  on: The na­ture of LLM al­gorith­mic progress

  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):

  … Amodei describes this as a “compute multiplier”: … These gains come from all sorts of places: … improvements to training datasets that allow the model to learn more quickly …

• sanxiyn 28 Jan 2026 6:25 UTC

  24 points

  8

  on: AI found 12 of 12 OpenSSL zero-days (while curl can­cel­led its bug bounty)

  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.

• sanxiyn 29 Oct 2025 0:26 UTC

  4 points

  2

  on: All the labs AI safety plans: 2025 edition

  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”.

• sanxiyn 5 Sep 2025 7:52 UTC

  10 points

  0

  in reply to: Zac Hatfield-Dodds’s comment on: The next wave of model im­prove­ments will be due to data quality

  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/​​

• sanxiyn 29 Jul 2025 2:17 UTC

  3 points

  0

  in reply to: ryan_greenblatt’s comment on: ryan_green­blatt’s Shortform

  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.

• sanxiyn 29 Jul 2025 0:08 UTC

  9 points

  4

  in reply to: ryan_greenblatt’s comment on: ryan_green­blatt’s Shortform

  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.

• sanxiyn 22 Jul 2025 3:35 UTC

  7 points

  0

  on: GDM also claims IMO gold medal

  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.

• sanxiyn 16 Jul 2025 0:56 UTC

  3 points

  0

  in reply to: Thane Ruthenis’s comment on: Grok 4 Var­i­ous Things

  Do you have any Solomonoff inductor you know? I don’t, and I would like an introduction.

• sanxiyn 13 Jul 2025 5:35 UTC

  1 point

  0

  on: OpenAI Model Differ­en­ti­a­tion 101

  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.

• sanxiyn 29 Jun 2025 1:27 UTC

  5 points

  0

  in reply to: Zac Hatfield-Dodds’s comment on: The next wave of model im­prove­ments will be due to data quality

  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.

• sanxiyn 18 Jun 2025 3:43 UTC

  2 points

  0

  in reply to: Vladimir_Nesov’s comment on: Serv­ing LLM on Huawei CloudMatrix

  CloudMatrix was not, but Huawei Ascend has been there for a long time, and was used to train LLM even back in 2022. I didn’t realize AI 2027 predated CloudMatrix but I still think ignoring China for Compute Production was unjustified.

• sanxiyn 17 Jun 2025 23:29 UTC

  1 point

  0

  in reply to: Vladimir_Nesov’s comment on: Serv­ing LLM on Huawei CloudMatrix

  This is a good argument and I think it is mostly true, but this absolutely should be in AI 2027 Compute Forecast page. Simply not saying a word about the topic makes it looks unserious and incompetent. In fact, that reaction happened repeatedly in my discussion with my friends in South Korea.

Serv­ing LLM on Huawei CloudMatrix

• sanxiyn 11 Jun 2025 4:53 UTC

  14 points

  0

  on: AI com­pa­nies’ eval re­ports mostly don’t sup­port their claims

  I know cyber eval results are underelicitation. Sonnet 4 can find zero day vulnerabilities, we are now in process of disclosing. If you can’t get it to do that it’s your skill issue.

• sanxiyn 20 May 2025 1:04 UTC

  5 points

  0

  on: Eliezer and I wrote a book: If Any­one Builds It, Every­one Dies

  Preordered ebook version on Amazon. I am also interested in doing Korean translation.

• sanxiyn 15 May 2025 4:43 UTC

  5 points

  4

  in reply to: Thane Ruthenis’s comment on: Fight­ing Ob­vi­ous Non­sense About AI Diffusion

  I disagree on DeekSeek and innovation. Yes R1 is obviously a reaction to o1, but its MoE model is pretty innovative, and it is Llama 4 that obviously copied DeepSeek. But yes I agree innovation is unpopular in China. But from interviews of DeepSeek founder Liang Wenfeng, we know DeepSeek was explicitly an attempt to overcome China’s unwillingness to innovate.

• sanxiyn 27 Mar 2025 5:30 UTC

  2 points

  1

  in reply to: solhando’s comment on: Re­cent AI model progress feels mostly like bullshit

  Maybe we are talking about different problems, but we found instructing models to give up (literally “give up”, I just checked the source) under certain conditions to be effective.

• sanxiyn 27 Mar 2025 5:23 UTC

  14 points

  0

  in reply to: ryan_greenblatt’s comment on: Re­cent AI model progress feels mostly like bullshit

  Our experience so far is while reasoning models don’t improve performance directly (3.7 is better than 3.6, but 3.7 extended thinking is NOT better than 3.7), they do so indirectly because thinking trace helps us debug prompts and tool output when models misunderstand them. This was not the result we expected but it is the case.

• sanxiyn 25 Mar 2025 22:49 UTC

  23 points

  3

  on: Re­cent AI model progress feels mostly like bullshit

  I happen to work on the exact sample problem (application security pentesting) and I confirm I observe the same. Sonnet 3.5/​3.6/​3.7 were big releases, others didn’t help, etc. As for OpenAI o-series models, we are debating whether it is model capability problem or model elicitation problem, because from interactive usage it seems clear it needs different prompting and we haven’t yet seriously optimized prompting for o-series. Evaluation is scarce, but we built something along the line of CWE-Bench-Java discussed in this paper, this was a major effort and we are reasonably sure we can evaluate. As for grounding, fighting false positives, and avoiding models to report “potential” problems to sound good, we found grounding on code coverage to be effective. Run JaCoCo, tell models PoC || GTFO, where PoC is structured as vulnerability description with source code file and line and triggering input. Write the oracle verifier of this PoC: at the very least you can confirm execution reaches the line in a way models can’t ever fake.