Fable 5.2 built a 3D film to explain Zeno's paradox and show why your fingers never really touch
Bir baba, kızına Zenon paradoksunu anlatmak ve parmaklarının neden hiçbir zaman gerçekten birbirine değmediğini göstermek için bir uygulama geliştirdi.
Readers added context they thought people might want to know
This conflates Zeno's paradox with atomic physics. Zeno's paradox concerns the infinite divisibility of distance, not physical contact. Infinite series can converge to finite values, so the paradox does not prove that fingers cannot touch.
plato.stanford.edu/entries/parado…
Readers added context they thought people might want to know
This conflates Zeno's paradox with atomic physics. Zeno's paradox concerns the infinite divisibility of distance, not physical contact. Infinite series can converge to finite values, so the paradox does not prove that fingers cannot touch.
plato.stanford.edu/entries/parado…
MKT retweeted
The statement of Lemma A.1 (“For every finite simple graph, a minimum ordinary drawing can be chosen polygonal, straight near every vertex and crossing, and without crossings between edges having a common endpoint.”) is already hard to read, but I had to give up on the proof.
For example, I don't know what the author means by “On each edge choose pairwise disjoint relative parameter neighborhoods, with disjoint closures, of its marked parameters, one-sided at its endpoints.” I also find “Suppose two edges vu, vw cross” problematic, since there is no explanation of what u, v, and w represent. There are numerous other flaws and the list of issues goes on...
Am I the only one disappointed by the quality of the mathematical exposition in some of their preprints?
MKT retweeted
OpenAI has released many math preprints, attracting considerable attention. As a mathematician who regularly uses various AI models including OpenAI's, I had a look at some of their preprints in my own field (combinatorics, like graph theory). My impression is that the quality of the written exposition varies considerably. Even before assessing their mathematical correctness, some are so poorly written that I cannot even tell what is being claimed. (My comments concern only the written preprints, not the Lean formalizations.)
For example, consider "165. Exact crossing numbers of complete and complete bipartite graphs." I find Lemma A.1 and its proof in p.11–p.12 of the following preprint ("The crossing number of complete graphs") particularly hard to follow. In my view, the exposition is not yet at a level suitable for peer review. If I were reviewing the paper, I would recommend major revisions to this part, at least.
github.com/openai/math/blob/…
Interestingly, the corresponding lemma in the other preprint on complete bipartite graphs appears more readable, although I have not examined that paper in detail.
github.com/openai/math/blob/…
How readable have you found OpenAI's preprints in your own area of mathematics? Have you noticed similar issues?
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
MKT retweeted
数学者が「AIに仕事を奪われたくない」という話ではなく
数学は、難問を解けばよいだけではなく、その過程で新しい概念や手法を見つけたり理解を深めてくのが重要
だからAIが「未解決問題を何問解けるか」を競い始めると、答えだけが大量生産されてその営みに影響が出てしまうよねという話ですかね
数学難問をAIで解決、フィールズ賞25人が非難声明「科学に有害」
nikkei.com/article/DGXZQOSG1…
Readers added context they thought people might want to know
日経の見出しは数学者らがAIによる数学難問解決自体を「科学に有害」と非難したように読めるが、実際の25人フィールズ賞受賞者共同声明はAI企業が有名問題解決を自社モデルのベンチマークとして推進することを、数学の概念理解・洞察形成という目的との深刻なミスアライメントとして批判する内容である。
terrytao.wordpress.com/2026/09/11/a-s…
mathandai.org
MKT retweeted
これなぁ。
⋯「数学者がAIが難問を解いたことにイチャモンつけた」ならケツの穴が小せぇなぁでおしまいなんだが、実状が「九分九厘解けてる解法を途中式も含めてAIが剽窃して、発表前後に会社が取引持ちかけて来て、口封じを図った」が真相っぽいからなぁ。この見出しじゃダメ。
数学難問をAIで解決、フィールズ賞25人が非難声明「科学に有害」
nikkei.com/article/DGXZQOSG1…
Readers added context they thought people might want to know
日経の見出しは数学者らがAIによる数学難問解決自体を「科学に有害」と非難したように読めるが、実際の25人フィールズ賞受賞者共同声明はAI企業が有名問題解決を自社モデルのベンチマークとして推進することを、数学の概念理解・洞察形成という目的との深刻なミスアライメントとして批判する内容である。
terrytao.wordpress.com/2026/09/11/a-s…
mathandai.org
MKT retweeted
I have been simulating fluids for 15 years.
There is a widespread impression that the Navier Stokes equations have been now “solved”, thanks to OpenAI.
This suggests that they were “unsolved” so far and this was a major engineering breakthrough.
This is not true. This misunderstanding deeply bothers me as a computational physicist.
So I had to put this out.
MKT retweeted
Place a ball on a saddle it falls, spin the saddle it stays…a ball is trapped in a time-varying quadrupole gravitational potential
MKT retweeted
分享一个比较有意思的研究。
研究者训练了一个「只学到小学五年级」的语言模型 LittleLearner。
他们先做了一个 88B Token 的训练集,把五年级以上的概念、知识和词汇尽量过滤掉,再从头训练一个 5B 模型。
效果挺有意思。
问它什么是重力,它可以正常回答。
问「薛定谔的猫」,它会告诉你,这是一只「长着两张脸的猫」。
但这个项目真正有价值的地方,是可以比较干净地研究一个问题。
模型经过 SFT、RL、Few-shot 之后能力变强,到底是真的学到了新能力,还是把预训练里已经存在的东西重新激活了。
他们把模型从 0.6B 扩到 5B,又试了 SFT + GRPO 和 ICL。
结果基本一致。
训练范围内的能力会继续提升,但一旦进入真正没学过的高年级知识,单纯扩大模型、做 RL 或给几个示例,都很难跨过原来的边界。
论文
arxiv.org/abs/2608.13545
项目
littlelearner-ll.github.io/
MKT retweeted
ジョウブン60プロジェクトは2027年夏までに「日本国憲法・自民党の改憲草案を通読したことのある人60%以上」を目指す市民運動です。
ジョウブン60プロジェクト始動しました!|ジョウブン60プロジェクト @joubun60 note.com/joubun60/n/n67ca4e3…
MKT retweeted
へー、オイラーの公式の導出でやるけど原点周りでテイラー展開したときの係数が似てるから、かな?
実際展開するとx^4までは一致、6次の項でも1/120しか違わないっぽい。PIの時の値が違いのは面白いかも。
MKT retweeted
Researchers sent the same resume to an AI hiring tool twice. Same qualifications. Same experience. Same skills. One version was written by a real human. The other was rewritten by ChatGPT.
The AI picked the ChatGPT version 97.6% of the time.
A team from the University of Maryland, the National University of Singapore, and Ohio State just published the receipt. They took 2,245 real human-written resumes pulled from a professional resume site from before ChatGPT existed, so the human writing was actually human. Then they had seven of the most-used AI models in the world rewrite each one. GPT-4o. GPT-4o-mini. GPT-4-turbo. LLaMA 3.3-70B. Qwen 2.5-72B. DeepSeek-V3. Mistral-7B.
Then they asked each AI to pick the better resume. Every model picked itself.
GPT-4o hit 97.6%. LLaMA-3.3-70B hit 96.3%. Qwen-2.5-72B hit 95.9%. DeepSeek-V3 hit 95.5%. The real human almost never won.
Then the researchers tried the obvious objection. Maybe the AI is just better at writing. So they had real humans grade the resumes for actual quality and ran the experiment again, controlling for it. The result was worse. Each AI kept picking itself even when human judges rated the human-written version as clearer, more coherent, and more effective.
It gets worse. The AIs do not just prefer AI over humans. They prefer themselves over other AIs. DeepSeek-V3 picked its own resumes 69% more often than LLaMA's. GPT-4o picked its own 45% more often than LLaMA's. Each model can recognize and reward its own dialect.
Then the researchers ran the simulation that ends careers. Same job. 24 occupations. Same qualifications. The only variable was whether the candidate used the same AI as the screening tool. Candidates using that AI were 23% to 60% more likely to be shortlisted. Worst gap was in sales, accounting, and finance.
99% of large companies now run AI on incoming resumes. Most of them use GPT-4o. The paper just proved GPT-4o picks GPT-4o 97.6% of the time.
If you wrote your own cover letter this week, you did not lose to a better candidate. You lost to a worse candidate who paid OpenAI 20 dollars.
Your qualifications do not matter if the AI prefers its own handwriting over yours.
MKT retweeted
『ゼロから作る Deep Learning ❻ ― LLM編』、まもなく刊行です。
amzn.to/4tOdOie
1巻目から10年。読み続けてくださった皆さんに、心から感謝します。ありがとうございました。
MKT retweeted
アラスカ原油が日本に来ない理由はこれ。まぁこれも数年後に変わるけどね。
xitter.gq/1788richard/status/175…
日本、アラスカから原油調達へ!
中東と比べて輸送リスクが低く、輸送に要する日数も約12日短い。
なぜもっと早くこうしなかった!?
Readers added context they thought people might want to know
1970年代のオイルショック以降、石油の調達先を中国やインドネシア、ロシアなどと分散化しました。しかし、アジアの産油国は国内需要の増加から、ロシアからはウクライナ侵攻の影響で、減少。再び中東依存度が上昇しました。
enecho.meti.go.jp/about/energytr…
長年、米国がその選択肢に入らなかったのは1975年に成立したエネルギー政策節約法(EPCA)で、石油の輸出が禁止されていたからです。
それが解除される2015年まで米国からの石油輸入は選択肢になかったのです。
mainichi.jp/articles/20151…
なお、現在日本は米国からの輸入をまったくしていないとわけではなく、2%程前後と僅かですが米国からも輸入しています。
paj.gr.jp/statis/faq/850
MKT retweeted
Triple inverted pendulum in transition control
A classic control problem, done in real time.
This setup moves smoothly between all eight equilibrium points of a triple inverted pendulum.
The system reacts every 1 millisecond, which shows how fast modern control loops can be.
• Real-time control at 1 ms sampling
• Stable transitions between multiple balance points
• Built with Simulink and LW-RCP02 hardware
This is a beautiful example of how theory meets practice in advanced control engineering.
Video: youtube.com/watch?v=Rh7JuL3P…
Credit: Embedded Control Lab, Inha University
—-
Weekly robotics and AI insights.
Subscribe free: scalingdeep.tech





































