Chey Loveday (@LovedayChey) header
Chey Loveday (@LovedayChey) profile picture

Agentic risk researcher. Stats engineering for governed AI (conformal prediction + sequential testing + behavioural sigs). ex-ICR, PhD in Cancer Genetics.

Wivenhoe, England
Joined October 2017
Photo by Chey Loveday (@LovedayChey)Photo by Chey Loveday (@LovedayChey)Photo by Chey Loveday (@LovedayChey)Photo by Chey Loveday (@LovedayChey)Photo by Chey Loveday (@LovedayChey)Photo by Chey Loveday (@LovedayChey)Photo by Chey Loveday (@LovedayChey)Photo by Chey Loveday (@LovedayChey)Photo by Chey Loveday (@LovedayChey)Photo by Chey Loveday (@LovedayChey)
The main proof in my first paper (currently a preprint) in number theory . Shifted-Square Obstructions and Excluded Fibres in Chebyshev–Pell Recurrences Main result: Assume a solution exists. The first move is to turn the shifted-square equation into a norm equation in a degree-17 or degree-41 number field. The second is to choose a detector unit so that a hypothetical square forces the odd-prime contributions to become trivial, while the contribution above 2 keeps the opposite sign. The third is what makes the argument global rather than computational: a Hensel argument shows that one exact dyadic calculation controls the whole admissible parameter family. Then the receipt is: - All odd finite places: +1 - The real place: +1 - Every complex place: +1 - Product over the places above 2: −1 - Global product: −1 - Hilbert reciprocity requires: +1 Thus.....No solution. Nothing is individually obstructive on any single line. The arrangement is what cannot exist. That is the bit I find super satisfying. Instead of showing directly that the equation has no solution, we force any supposed solution to have several perfectly valid local descriptions which cannot all belong to one global object. And, as far as I can establish, I haven’t found another AI-assisted number-theory proof whose decisive obstruction is this kind of Hilbert-reciprocity sign contradiction. AI-assisted mathematics is moving quickly, so if there is one, tell me. I’d much rather correct the claim than repeat it!!!
1
3
274
Chey Loveday retweeted
🚨 codex users!! GPT-6.2 Sol is BETTER than Opus 5.5 and its ONLY 1/10th of the price!! Litrally 10x savings!!
Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…
38
13
5
184
27,408
Chey Loveday retweeted
𝐋𝐞𝐚𝐧 𝟒.𝟑𝟔.𝟎 𝐢𝐬 𝐥𝐢𝐯𝐞! This release brings 222 changes, including 30 performance improvements, new 𝚐𝚛𝚒𝚗𝚍 and 𝚋𝚟_𝚍𝚎𝚌𝚒𝚍𝚎 capabilities, and experimental exception support in contracts. Notable improvements include: • Performance work including a type class cache kept between commands and the removal of several quadratic slowdowns; Mathlib now builds roughly 12% faster • 𝚐𝚛𝚒𝚗𝚍's bit-vector and fixed-width integer homomorphisms now cover masks, concatenation, rotations, signed comparisons, shifts, and exponentiation • 𝚋𝚟_𝚍𝚎𝚌𝚒𝚍𝚎 supports uninterpreted functions via +𝚞𝚏, adding congruence facts as needed (for functions returning 𝙱𝚘𝚘𝚕 or 𝙱𝚒𝚝𝚅𝚎𝚌) • Experimental 𝚝𝚑𝚛𝚘𝚠𝚜 clauses let a 𝚍𝚎𝚏 contract state what holds when the function throws 🔗 𝐅𝐮𝐥𝐥 𝐫𝐞𝐥𝐞𝐚𝐬𝐞 𝐧𝐨𝐭𝐞𝐬: lean-lang.org/doc/reference/… #LeanLang #LeanProver #FormalVerification
4
28
2
175
5,449
GPT-6.2 Astra and GPT-6.2 Sol a re now in codex. Happy codexing!!!
32
Chey Loveday retweeted
Great to see Lean being used for Rust verification!
Aristotle now formally verifies Rust, joining C and Verilog 🦀 Prove your Rust code has no undefined behavior and does exactly what its spec says, even if you use unsafe. And if you’re porting C to Rust, prove the new code behaves the same as the original.
7
29
3
320
15,340
Chey Loveday retweeted
Using Codex, we proved a conjecture Euler proposed in 1747—279 years ago! Can every positive integer that is 2 mod 4 be split into two positive integers whose prime factors are all 1 mod 4? Yes, allowing 1 as a summand. Our eight-page paper gives the proof.
23
32
2
364
18,479
Chey Loveday retweeted
Codex team might have actually fixed the "model at capacity" errors that have been pestering you.
2,511
248
386
9,189
1,337,572
Chey Loveday retweeted
There is a belief that AI-generated formalization is not reusable. This belief is outdated. We show 30 theorems in finite group theory that fall out easily from our CFSG formalization.
Recently, we formalized the Classification of Finite Simple Groups, the longest proof in math. Also the largest open-source software repo. Beyond certifying the result, we ask: Is it reusable? What else can we formalize? The answer is YES. 30 new results, built on top of it!🍇
5
9
49
2,466
"It's getting surreal." @ValerioCapraro co-authored a paper defining "Epistemia": linguistic plausibility substituting for epistemic evaluation. Then he applied it to @ElliotGlazer. Glazer was explicit about his limits: he clearly stated he didnt fully understand the proof himself, but had subjected it to AI-assisted auditing, spending roughly 50x more compute on verification than discovery. Capraro quoted the caveat as a confession, declared "nobody has actually verified," but didnt actually examine the reported audits. Independent verification is still needed (lots of progress has been made on this front). So it would have been a legitimate criticism if properly scoped. But dismissing an audit without examining what it tested, what it established and what remains open isn't epistemic evaluation either. A paper warning that plausibility replaces judgment, deployed in an argument that bypassed the available evidence. That's epistemic laundering. And thats what is surreal to me. ----- Btw, science has been vulnerable to this for decades. A Nature or Science logo & a prestigious lab can stand in for scrutiny. Case in point: Bell Labs physicist Jan Hendrik Schon published fabricated results in both journals, paper after paper, until other researchers spotted duplicated graphs. GenAI didn't invent this failure mode. It made it cheaper, faster & available to en masse.
2
39
The most exciting thing about a release is revisiting those problems intractable to the previous models.
JUST IN: GPT-6.2 Astra and GPT-6.2 Sol are releasing today.
1
38
Chey Loveday retweeted
JUST IN: GPT-6.2 Astra and GPT-6.2 Sol are releasing today.
29
2
1
134
23,419
ASTRA 6.2 incoming!!!!
This quoted post is unavailable.
1
1
121
Get urself out the basement @thsottiaux ..... im at 5%, with 2d20h to go and a very important conjecture to solve. ;)
It’s been a while without resets. Maybe I should come back out of the basement
51
Chey Loveday retweeted
最近发现一个专门给学术写作“去 AI 味”的开源 Skill:Academic Humanizer。 它不是简单删除“此外、值得注意的是”这些高频词,而是从六层检查论文: 常见AI句式、学术写作中的空泛表达、必要的学术规范、论点与证据是否匹配、个人与目标期刊的语言风格,以及基金申请书的特殊写法。 它比较有价值的一点,是不会为了“像人写的”就把论文改成小红书。 数字、实验结果、引用和必要的谨慎措辞都会保留,还能参考你以前的论文,以及目标期刊或会议的行文习惯。 但别把它当成AI检测器作弊器。最稳的用法是: 先让AI完成初稿,再把你过去写得最好的一篇论文、目标期刊和待修改章节一起交给它,要求逐项说明修改理由。 否则换一个 Humanizer,最后只是从一种AI味变成另一种AI味 GitHub:github.com/AIScientists-Dev/…
17
184
6
1,271
76,108
Chey Loveday retweeted
People working at AI companies are humans too. We have feelings. Many of us were in academia not long ago. Please remember this before writing nasty things on the Internet about us. Thank you.
87
7
8
124
25,128
Chey Loveday retweeted
DataColada[141]: Evidence of fraud in 5th paper co-authored by Francesca Gino datacolada.org/141 (Grant, Gino, & Hofmann, AOM, 2011)
9
61
27
342
73,912
I should s thanks @SAIRfoundation for Open Math OS platform sccess. I used it for a six-checker Lean --paranoid replay of 41 conditional results now in PR #27 to Elliot Glazer’s Z-in-Q project.
1
10
Chey Loveday retweeted
Replying to @ValerioCapraro
This critique is nonsensical on so many levels. 🤦‍♀️ 1. @ElliotGlazer was so humble about his contribution he didn't even put his name as an author on his own paper, yet you spitefully refuse to even mention his name, to avoid accidentally yielding him an inch of recognition for anything. 2. You ask "So, what is the problem?" but then fail to say anything that actually answers the question. 3. "The proof doesn’t even have the formal Lean verification": wait, "doesn't even"? Now this is suddenly a requirement, when all proofs a few months ago, and 99.9% of published proofs today, still don't satisfy it? (And when proofs do include a Lean certificate this same practice is mocked by critics of AI "button-pushers" who call it "leanslop" that "doesn't advance understanding"?) 4. "This is literally an AI generated proof that looks plausible, but nobody has actually verified." What's your point? That's how science works - someone does something, then shares it with the community so people can check it; if it's wrong this gets pointed out and we learn what there is to learn and move on. What exactly is the "literally" here meant to convey, other than that your logic is weak so you need to resort to rhetorical gimmicks? 5. "This is exactly what we call Epistemia". Oh, you gave a practice a fancy name so you get to say it's bad without having to explain why and nobody is allowed to disagree? 6. "Until recently, a mathematician would be embarrassed to announce a proof he couldn’t understand." WHAT??? First, "until recently" you couldn't actually discover a proof if you didn't "understand" (at a rather deep level) what you were doing, so this hypothetical way of being embarrassed didn't even exist. Second, there are plenty of proofs the discoverers themselves feel are mysterious and defy proper understanding. (Hence Von Neumann's famous quip: "Young man, in mathematics you don't understand things. You just get used to them.") 7. "Now it risks becoming the norm." If it "risks" becoming the norm for people to do work that tries to advance their science to the maximal extent they are capable of, then share the work with others to try to critique and improve on, this sounds like something we should welcome. And I would hope it's already the norm. 8. "Announcements like this get you thousands of followers and interviews with journalists. [... and then several lines down:] In the first comment, you’ll find our epistemological fault lines paper". Hmm. Did someone say something about short-term incentives...? 🤔
6
21
3
231
10,548
On #SocialistSunday, this..... "We must learn to live together as brothers or perish together as fools"
4
5
7
120