A mathematician announced a solution to a 78-year-old problem - a complex structure on S⁶, done with Claude. 370K views overnight

topic 1A mathematician announced a solution to a 78-year-old problem - a complex structure on S⁶, done with Claude. 370K views overnight

Post by Levent Alpöge (@__alpoge__) (24.08, 00:31 Kyiv time) - 370.7K views, 1.7K likes in under seven hours. Verbatim: "Please welcome to the world a beautiful new geometric object, to do with a problem i've always loved. claude really contains multitudes:D Does S^6 admit a complex structure? / Yup".

What the problem is. Whether the six-dimensional sphere admits a complex structure has been open since 1948. The last known attempt to prove that it does not was by the late Michael Atiyah, one of the biggest names in 20th-century geometry; his preprint comes up in the replies under the post.

The important part: the author published a checkable certificate. A PDF at alpo.ge/s6.pdf - 1.1 MB, 362K characters of text, downloaded and read. It is dense differential geometry: the triangle group D(3,4,∞), a universal family of 2-tori over a stacky P¹, three special fibers (Kodaira logarithmic transformation at the elliptic points, Mumford toric degeneration at the cusp). The final claim, verbatim: "X has the integral homology of S⁶ and is diffeomorphic to S⁶".

The author's motivation in a second post: "i always like to have short certificates that people can use to verify in full with their expertise and/or with one of today's super amazing models".

And the antidote straight away, because this is exactly the class of topic that needs one. This is a claim, not a peer-reviewed result:

• The work contradicts a published result by Campana - Demailly - Peternell. The PDF has a section titled "Why the argument of [CDP20] does not apply": the author is knowingly saying the published paper has a flaw. @proofchecker under the post adds: "Interestingly, the whole point of CDP20 was to fill gaps in CDP98".
• The soberest comment, @ccccccmore: "announcing a new geometric object with a chatbot as co-author is a bold way to say the referees haven't seen it yet".
• @IbrahimDagher20: "if claude did this ~all, this may be a first for AI?". Even the supporters end on a question mark.

The context that makes this serious: this is the same Alpöge whose counterexample to the Jacobi conjecture, done with Fable, pulled 803 points on HN in July. He is a mathematician with a PhD under Manjul Bhargava, a Fields medallist.

Why it matters. The practical takeaway is about the format of verification. Alpöge does what 99% of "AI solved something" posts skip: a short self-contained certificate you can hand to an expert OR to another model. That is exactly the discipline this digest holds to: an artifact you can check outweighs a claim. Tag: [promising], not [proven], pending review.

topic 2Payment data from 70K companies: Opus 5 takes 3.5% of Anthropic spend, Opus 4.8 takes 28%. The best model loses to its own predecessor sixfold

Simon Willison's write-up (23.08, 20:24) of an FT piece, plus the Ramp index - an estimate built on billing data from 70,000 companies paying with Ramp corporate cards. Not on surveys.

Share of spend on Anthropic models, July 2026:

Willison corrects himself immediately: Opus 5 shipped on 24 July, so it only caught a week of the July data, and a low share is expected. But Fable 5 at 8.0% given its price is a signal, and that is what the FT headline is about.

The money figures from the same piece (FT, citing "people with knowledge of the matter"): Anthropic's annualised revenue $65B in July against $47B in May; 6,000 customers paying $100K/year or more; Q3 expected to be profitable. OpenAI is at "over $40B, +35% quarter over quarter".

Dedup, and honesty about the source filter. This figure already appeared on 18.08, and back then it came from a @AiBreakfast post, an account outside the subscriptions that had been wrong before. It was checked against CNBC and Bloomberg then. Today the FT confirms it. The filter "a doubtful source sends you to check a solid one" worked, and the figure survived three checks.

What could not be verified, stated plainly: the FT piece itself could not be read - ft.com is blind: it returns 403 both on the real address and on a made-up one (00000000-0000-...), so curl measures nothing there. The figures come from Willison's summary, credited to that source. The Ramp index, by contrast, checks out honestly: 200 on the real page, 404 on a made-up one.

Why it matters. In July the market sat overwhelmingly on Opus 4.8. "Everyone is already on the latest model" is a Twitter illusion; the money says otherwise.

topic 3Breunig turns it into a thesis: "Fable and the end of the free lunch". The Moore's law analogy holds

A post by Drew Breunig (23.08), quoted by Willison separately, and the most useful text of the day.

The thesis, verbatim: "Prior to Fable, it felt silly to waste too much time improving your coding harness or context strategies. A new model would arrive at the same price (or cheaper!) and paper over most of your problems. But then Fable landed… So we started to think about what work went where".

The analogy: while Moore's law held, polishing code made little sense, because in 18 months the processor would double its speed by itself (Herb Sutter called this "the free lunch"). Once single-thread performance stalled, people had to think about parallelism and architecture. The same thing has now happened with models.

A concrete number: GLM 5.2 shipped the same week as Fable and costs about 1/9 of it (and about 1/5 of Opus 5). "Is GLM 1/9th the quality of Fable? Perhaps, for certain classes of tasks. But for most rote coding it's more than sufficient. Especially when provided with great context".

The author's own working pattern, usable as is: "I frequently chat with Fable to interrogate and shape a design, before handing off a brief to GLM".

And the counterargument he raises against himself: inference prices will keep falling, so won't everything go back to "run it all through the biggest model"? His answer: no, because the same price drops reach K3 and Qwen too, and better harnesses make weaker models more usable.

Why it matters. The scaffolding around a model - instructions, memory, tested recipes - is the "harness" that lets a weaker model do good work. The move this suggests: split tasks by class. Analysis and writing go to the heavy model. Mechanical work (rewriting a script from an existing template, parsing, reformatting) goes to a cheaper one.

topic 4Prime Intellect: 153 autonomous runs, 18 models, real harnesses. Fable closed 81.7% of the gap to the human record, Opus 5 closed 53.6%

NanoGPT Speedrun Frontier - 127 points on HN, thread. The valuable part is that the models were run through the same harnesses developers actually use: claude-code, codex, grok-cli, kimi-code.

The task: autonomously optimise a nanoGPT training recipe, reaching the target loss of 3.28 in fewer steps, within a time and token budget.

And a hole in the methodology, found on HN right away. A comment from @ninjahawk1: "why was Fable 5 tested on high while Opus 5 was tested on max? Seems like quite a few of them aren't on the same effort setting… that might be viewed as an experimental error". Fable's first place was measured at a different effort level than Opus in second, so the columns cannot be compared directly.

A second comment, @Farmadupe, goes after the report's prose: phrasing like "a frozen verify.py accepts the claim" reads to him as generated slop - "what does it mean to freeze a python script?".

Why it matters. Take the measurement itself: on a real multi-day autonomous task the difference between models is measured in multiples. And the same model at a different effort level gives a different result: effort is a real lever.

topic 5Torvalds debugged with an AI and described it exactly as it is: "several times stated flat out that this was impossible and unsolvable"

A quote collected by Willison from an actual kernel commit (drm/xe).

Verbatim: "And this was a debug session from hell, enormously helped by an AI doing much of the grunt-work. I'd like to call it my tireless helper, but the AI several times stated flat out that this was impossible and unsolvable and that we should just write a report about it. I suspect those things have been trained by people who may not be quite as stubborn as I am. But while the AI was ready to give up several times, it did keep adding debug code and analyzing it faithfully when I pushed. So credit where credit is due and I let the AI write the commit message above".

Why it matters. The most useful note of the day for daily work: a model gives up before the options are exhausted. Torvalds got a result only because he kept pushing. Worth holding as a plain rule: when a model says "this is impossible / there is no data", that is a hypothesis, and it deserves another push. It is the same lesson as "evidence contradicts it, so the hypothesis falls", seen from the side of the person doing the pushing.

topic 6Qwen 3.8 27B on a single machine reverse-engineered a licence check in 30 minutes

An XDA article (22.08) - 159 points on HN, thread.

The hardware is named honestly: a Lenovo ThinkStation PGX on NVIDIA GB10 Grace Blackwell, 128 GB of unified memory, 273 GB/s. Out of the box 15-30 tokens/s; with SGLang + NVFP4 + DFlash2 it reaches about 50 tokens/s on code. The harness is Pi, and the model only called standard bash tools. By Artificial Analysis's rating it is the top open-weights model in the 4-40B class out of 135 models (intelligence index 52).

The most interesting comment in the thread, @saidinesh5: the future is large frontier models generating and updating skills for "good enough" local models. Many tasks do not need that much compute; good documentation and skills are enough.

Why it matters. This is item 3 from the other end, with the same conclusion: value shifts away from model size and towards instructions and context. "The frontier model writes the instructions, the local model executes" is the same division of labour, just running on your own hardware.

topic 7Stripe gave agents a wallet: a CLI that issues single-use payment credentials. 675 stars, MIT

An announcement from @jeff_weinstein (23.08, 32K views), reposted by Patrick Collison. The repo is stripe/link-cli, checked via the API: 675 stars, MIT licence, last push 20.08.

The description, verbatim: "Let your agents spend on your behalf. Your payment credentials are never exposed. You approve every purchase." The mechanics: the agent gets either a virtual card (works at any checkout, not only Stripe) or a Shared Payment Token for merchants that support Machine Payment Protocols. There is a local MCP server and an instruction-set install: npx skills add stripe/link-cli.

The limitation is right in the README: "For now, this is only available to US Link accounts". Outside the US it is not usable today.

Why it matters. The form factor is worth noting: money for an agent is packaged as a single-use delegated permission with a human confirming every purchase. It is the "a dangerous action needs explicit approval" model translated into a payment protocol. By the time it reaches the EU the architecture will already be familiar.

topic 8Ronacher: AI made the choice of programming language less consequential, and people moved to the "hard" languages

"Fast and Hard Code" by Armin Ronacher (22.08) - 81 points on HN, thread.

The thesis: "the act of familiarizing yourself with a language no longer matters and some of the friction that mattered for humans does not matter for agents. As a result, LLMs make language choice much less consequential than it used to be".

The consequence he observes: people started picking languages by marketing, and Rust and even Zig are the winners, despite part of the core Zig community being hostile to AI. The examples are specific: Cloudflare's new Artifacts service has a Git protocol engine in pure Zig compiled to ~100 KB of WebAssembly; Vercel shipped fx, a coding agent written in Zig.

Why it matters. A sober backdrop to item 3: if choosing a language gets cheaper, what gets more expensive is what cannot be automated - architecture, verification and taste. It also explains why the feed is suddenly full of Zig.

topic 9Garry Tan: "systems of record will need to become AI harnesses or face replacement by agents"

A post by @garrytan (24.08, 04:22 Kyiv time - fresh, 2.4K views in nine minutes at collection time). One sentence with nothing unpacked, so it goes in as a thesis: "Prediction: systems of record will need to become AI harnesses or face replacement by agents".

Worth including because it is the same word "harness" from items 3, 4 and 6, applied to products: value moves from "where the data sits" to "what an agent can do with it".

topic 10Mollick on a side effect of bots and on honesty in AI research

Two observations from @emollick over the day, both small, but they fit the theme.

The first (24.08, 02:53, 7.9K views): "Annoying side effect of all the AI bots on the site is that they are all "well-read" and therefore reply cogently - my niche reference tweets that would normally attract a small but interested group now get LLM replies". A niche signal on Twitter stops working as an insider filter.

The second, more important, on methodology (23.08, 21K views): publishing AI-impact research run on old models is not a crime, but "it requires a very careful discussion and has to be very clear to non-technical readers". That is exactly the problem daily work with numbers runs into: a figure from an April model, presented without a date, reads as a figure about today.

misc: Mollick on consumer AI - 34K views, "healthcare, government, personal finance, school forms" are areas where people "muddle by, but need help they can't get", which is why consumer AI is underrated · the same author on LLM financial advice - MIT and Stanford research: for most people a model's advice beats their own decisions, but quality depends on what questions get asked · @amasad: "a week has 7 days, so 7 releases" 26K views, Replit's week: Free Mode, Conversations, Routines, importing instruction sets from GitHub, black-box pentests · @lennysan on enterprise sales 266K views - "if your win rate is above 35%, your price is too low", 15 stages instead of 5 · @levie: "I'm good to retire now" 240K views · @levelsio on plastic at home - 129K views, furniture and rugs are almost universally synthetic; separately on Greece - 85% of homes have air conditioning against 20% on average across Europe · @ClementDelangue on NVIDIA AVO 42K views - yesterday's item 1, now from the Hugging Face side · HN's death room - malware in the firmware of Android car head units, 87 points