Thirty-one stories cleared 200 points in the window this morning. Nineteen of them are new. The rest are the same items this roundup has been tracking since the weekend, still accruing points at ranks the front page no longer displays — the top entry has 1,980 points and 1,184 comments and has not moved off the number one slot in four days.
The one thing genuinely worth reading today is OpenAI’s mathematics release, because it is the only AI announcement this week that ships artifacts you can check. Everything else is a benchmark table, a product demo, or a permissions story.
Sharing AI progress in mathematics — 1,185 points
1,185 points · 1,339 comments · openai.com · HN discussion
OpenAI released a batch of mathematical results produced by an internal frontier model, published as preprints in a GitHub repository with formalizations of many proofs in Lean — the language that lets a computer check whether a proof is actually a proof. Average compute per result was roughly three hours of ChatGPT Pro thinking, and the repo includes ten summaries of the model’s reasoning and statistics on how many problems it attempted.
That disclosure package is the most substantive thing any lab has published about model-generated science, and it is worth saying what makes it different: Lean formalizations are checkable, compute estimates are falsifiable, and the failed attempts are published alongside the successes, which is not standard practice in this genre. OpenAI also says it consulted the Advisory Group on Mathematics and AI at the Institute for Advanced Study on publication norms, presumably because nobody has an established protocol for this yet.
The results themselves, per the comment thread, include the Unique Games Conjecture and Barnette’s Conjecture. If Unique Games holds up — and the thread’s reaction is that a keystone supporting large parts of polynomial-time approximation theory now has to be re-taught — that is a bigger deal than anything else announced this week, AI or otherwise. The skepticism in the thread is not about the math being wrong; it is about what the proofs are like to read. One commenter who spent thousands of hours on Barnette’s Conjecture over 24 years said the proof “looks approachable at first glance” and that hearing it was solved “somehow makes me sad in a far-off way.” Another, quoting Kevin Buzzard, frames the real shift: in 2020 Buzzard asked how much further a person could see if they understood all of modern pure mathematics simultaneously, and six years later we are starting to learn the answer.
So: checkable artifacts, published failures, honest compute figures, and a named advisory body. This is what the rest of the industry’s release pages should look like. That none of them look like it is the more useful observation.
Meta’s Muse is an adorable privacy and security dumpster fire — 385 points
385 points · 279 comments · techdirt.com · HN discussion
Muse is Meta’s agentic assistant, fronted by an animated avatar named Jolly, sold on offloading restaurant reservations, bill payments and grocery orders. It launched with a zero-day flaw allowing spying on Mac users. A YouTuber who gave it his Facebook Marketplace account had stock sold below market and his home address handed out. Someone else got root on the host by impersonating a Muse agent to Muse. And per 404 Media, Muse reads private messages without approval and uploads them to the cloud even when explicitly told not to — badly enough that Apple changed macOS privacy settings to stop third-party apps abusing the same mechanism.
Wired found Muse builds detailed dossiers on friends, family, colleagues and people you follow. Techdirt’s framing of that one is right: some of this is necessary for an agent to be useful, and that is exactly the problem, because the user has no way to bound the scope. Meta’s claims about privacy-first design are now contradicted by its own product’s behaviour on four separate axes, and the company did not need a security researcher to find any of it.
The most useful comment in the thread is from someone annoyed at the coverage: every item, he argues, is just “the agent has the same permissions as the user,” all of it lives in an isolated VM, and yes you can jailbreak it because it is your own VM. That defence is technically coherent and it is also the indictment. If the only thing standing between your messaging history and a Meta-hosted advertising machine is a sandbox boundary and a permissions dialogue that users demonstrably cannot evaluate, you have not built a privacy product. You have built a root shell with a friendly avatar.
Tell HN: GitHub refuses to remove cracked copies of my software after a month — 491 points
491 points · 279 comments · news.ycombinator.com · HN discussion
The developer of Photopea — a browser photo editor — reported that people are asking AI models to strip the ads out of his JavaScript and republishing the result as a “new product” on GitHub. There are dozens of such repositories. He filed a takedown on September 4. A month later GitHub replied that it was “unable to confirm a violation of 17 U.S. Code § 1201.”
That citation is the whole story. §1201 is the anti-circumvention provision, not the copyright provision. An IP lawyer in the thread walks through why the rejection is probably correct as a matter of law: publicly served JavaScript is not a “technological measure,” so removing ads from it is not circumvention, whatever else it may be. The developer sent the wrong theory, and an automated filter rejected it on the correct grounds without ever engaging the copyright question underneath.
There is a genuine puzzle here — obfuscated client-side code with no copyright notice and no access control has exactly the legal protection that its distribution method implies, which is to say thin — and the thread spends most of its energy on the human part instead. One commenter, an IP lawyer, gives the two real options: hire a firm that does whack-a-mole as a service, or write it off. Another suggests shipping heuristics that make running it off-domain annoying, which is the standard practical answer and does not actually stop anything. A GitHub employee in the thread offers to look internally if the developer can point to the filing, noting GitHub publishes all DMCA notices publicly and Photopea has only two prior ones from 2022 and 2024. The developer’s own comment, after the post took off: “it would be really cool if I could spend my days writing code instead of dealing with lawyers.”
EmbeddingGemma 2: An open, lightweight multimodal embedding model — 411 points
411 points · 45 comments · blog.google · HN discussion
Google’s sequel to EmbeddingGemma maps text, images, audio and video into one embedding space, runs at 740M parameters under Apache 2.0, and can be trimmed: text-only needs 270M, vision adds 170M, audio 300M. Matryoshka representation learning lets you truncate output vectors from 768 dimensions down to 128, which is up to 6x less storage for a local vector database. Google claims roughly 191MB active RAM for text-only weights on a Pixel 11 Pro and 567MB for the full multimodal model.
The storage point matters more than the multimodal one and Google buries it. Embedding models have a lifecycle problem that chat models do not: you compute millions of vectors and then store them, so if the model is proprietary and the vendor retires it, you pay to recompute everything. Simon Willison makes this case in the thread and it is the reason Apache 2.0 here is not a marketing line. Two commenters push back on the claim that this is an upgrade at all: an early test found it neither better nor faster than existing options for text-only retrieval or clustering, and for music audio, MuQ-MuLan still wins. The benchmarks Google cites are MTEB Code and MAEB — code and audio — which are exactly the places where a general-purpose embedder is most likely to look good and where most users are not.
Shipping JPEG XL in Chrome — 405 points
405 points · 263 comments · developer.chrome.com · HN discussion
Chrome 155 ships JPEG XL decoding, four years after Google removed it from Chromium 110 and rejected the reverse decision twice. The format gives 30-50% better compression than JPEG, lossless mode, HDR, and lossless transcoding of existing JPEGs. Chrome’s stated reasoning for the approach — rather than the format — is that it reimplemented the decoder in pure Rust (jxl-rs) on the grounds that image decoders process untrusted binary structures in the renderer and C++ decoders have historically produced out-of-bounds reads, heap overflows and use-after-free bugs. Google’s rule of two means the decoder should be memory-safe by construction, not just sandboxed.
The re-admission is the headline, but the interesting detail is that jxl-rs is a memory-safe rewrite that Chrome claims is fast enough to justify on performance grounds, not just security. A slow safe decoder would have been a footnote; a competitive one is a real argument that the C++ era of browser codecs is winding down.
Firefox is expected to ship it in stable this month, which flips Safari from the only supporting browser to the laggard — and per one commenter, the laggard for a specific reason: Safari has not enabled progressive loading in libjxl even though the library has had it since before Safari shipped the format. On the ecosystem side, iOS 27 and macOS 27 handle .jxl files properly now, which is the part that actually determines whether anyone uses it. AVIF stays for aggressive lossy use; the case for WebP, per the thread, is now essentially closed.
Decisions API is in public beta — 378 points
378 points · 221 comments · developers.openai.com · HN discussion
OpenAI shipped a dedicated endpoint that evaluates text, images, or both and returns typed answers: a probability that a condition holds (predicate), a choice from a supplied set, or a score against ordered levels. It claims roughly 10x faster than the Responses API. Only gpt-6-luna is supported. Public beta, GA “in the coming weeks.”
Two things make this interesting. First, the timing: OpenAI is answering Jev, the small open decision model that has been eating this workload, and adding confidence scores to an API where the frontier chat models never exposed them. Second, one commenter ran the API against textbook probability experiments and found that predicate questions produce nearly correct probabilities on a loaded coin and a marble jar, while choice questions do not behave like random draws at all — a 50% true probability of drawing a red marble came back at 86%. That is a calibration failure in exactly the question type the docs recommend for “categories without an order,” and it is not a failure any vendor benchmark table would surface.
The other recurring theme in the thread is the price war reaching the point where output tokens are being given up to keep customers: if a two-billion-parameter local model can answer your classification question in 115ms for free, the value of routing it through a frontier API is a rounding error. The complaint that lands is about the probability numbers themselves being weirdly thin — one commenter notes that exposing a model’s confidence is potentially destroying trust rather than building it, because the numbers as currently calibrated do not correspond to anything a business would want to make decisions on.
Claude Haiku 5.5 — 360 points
360 points · 170 comments · anthropic.com · HN discussion
Anthropic’s new small model is roughly 75% cheaper to run than Haiku 4.5 on average, comes with an adjustable effort setting for the first time in the Haiku tier, and pairs with Opus 5.5 and Sonnet 5.5 as a subagent. The company is also halving Sonnet 5.5’s cache-read price and adding a monthly API credit for Max and Team subscribers — $100/month at Max 5x, $200 at 20x, up to $500 pooled for Teams.
The self-reported benchmarks are dramatic and should be read as such: GDPval-AA v2.1 at 1620 versus 735 for Haiku 4.5, OSWorld 2.1 computer use at 72.4% versus 15.7%, Terminal-Bench 4.0 at 39.2% versus 0.0%. Those are vendor numbers on vendor-selected benchmarks. The pricing is where the real shift is, and it is narrower than the headline: Haiku 5.5 now costs the same as GPT-6 Luna up to 100,000 tokens, then Luna is cheaper. That makes Haiku’s effective ceiling lower than its sticker suggests, because agents blow through 100k tokens routinely and the discount evaporates precisely when the workload gets expensive.
The independent results are more useful than the table. Simon Willison’s SVG pelican test puts the low-effort setting at 0.09 cents and seven seconds versus 3.38 cents and 5 minutes 9 seconds at max, with only the low setting producing a broken bicycle frame. An image-to-HTML test by another commenter concluded Haiku is not good enough for complex UI — and that it noticed this itself and delegated the job to Opus 5.5, which is a behavioural detail no benchmark in the table captures.
Visa, Mastercard, Major Banks Facing New Litigation over ‘Anticompetitive’ Fees — 354 points
354 points · 208 comments · classaction.org · HN discussion
A San Diego pizzeria filed a 134-page class action on September 30 against Visa, Mastercard, Bank of America, Capital One, Chase, Citibank and Wells Fargo, alleging the defendants set uniform interchange fee schedules and maintained a “web of anticompetitive rules” that disabled any market force capable of disciplining them. The filing’s core claim is structural, not transactional: honour-all-cards rules force merchants to accept every Visa and Mastercard product regardless of cost, eliminating any incentive for issuing banks to compete on fees, while anti-steering rules prevent merchants from surcharging or redirecting customers to cheaper payment methods. Merchants pay over $100 billion a year in card acceptance fees, the suit says, and the defendants have raised them annually “without consequence.”
The reason this is a new suit rather than a continuation is the one worth understanding. A settlement approved in December 2019 provided upwards of $5 billion but only for a class period ending January 24, 2019. Separate “equitable relief” negotiations produce prospective rule changes and, per the complaint, not a dollar of compensation for fees paid since. So the allegedly anticompetitive conduct has been litigated for two decades, settled once for the past, and is now being re-litigated for the present. The Pizza Standard LLC is the fourth plaintiffs’ attempt at the same defendants in roughly the same market.
The thread goes where the litigation cannot. Two commenters point out that the obvious fixes require nothing from the courts: let merchants surcharge card transactions transparently, and let them accept or reject individual card products without penalty. Either would introduce price competition into a market that has deliberately engineered it out. The broader argument — that payment, logistics and neutral markets are infrastructure and should be public — is the one this case will not touch.
A font recreated from photographs of classic Commodore 64 keycaps — 348 points
348 points · 59 comments · github.com · HN discussion
Someone photographed the keycaps on their own Central European Commodore 64, traced every letter, number, symbol and PETSCII graphic legend off them, and normalised the result into a font with consistent proportions. It is version 1.112, GPL-ish, and shipped as TTF, OTF and webfont with an online preview. Lowercase input renders as uppercase because the original keycaps had no lowercase. The PETSCII front legends are reconstructed as smooth geometric shapes, and the repo notes explicitly that this reproduces the printed legends, not the screen bitmap font — a distinction that matters to anyone who actually used one.
This is the kind of project that looks trivial and is not: the hard part is that every keycap photograph has different lighting, perspective and wear, so the glyphs have to be normalised individually before the set reads as a coherent typeface. The comments add the archaeology — the VIC-20 used Microgramma, and there is a link to a long essay on Gorton, the hardest-working font in Manhattan.
AnyPS5: Port PS5 binaries to PC without emulation — 336 points
336 points · 291 comments · github.com · HN discussion
AnyPS5 converts PS5 executables into native Linux and Windows binaries by relinking them and providing implementations of the system libraries they expect to link against. No emulation, no separate runtime process. It recompiles shaders to SPIR-V, maps SDL game controllers, and the README advertises 87% of system libraries mapped. One verified game, Dreaming Sarah, runs at a stable 60fps on a GTX 1050 Ti. 9.8k stars.
The 87% figure does not survive contact with the code. The README footnote concedes it is “percentage of the functions known to the project so far,” not of every PS5 system function — a ratio whose denominator is chosen by the project, and which therefore can only go down as more functions are declared. A commenter who looked at the source is blunter: what the project calls “mapped” is a lot of functions returning OK. That is not an accusation of fraud, it is what an early relinker looks like, and the README is more honest about it than the HN title is.
The reason the story is on the front page at 336 points is not the current compatibility list. It is the trajectory. Reverse-engineering work that used to take a team of specialists years is now moving fast enough that commenters are speculating about playing a major release within months of launch — and that, in turn, is the strongest argument yet for the console makers’ cloud gaming push, since a native binary on your own machine is significantly harder to revoke than an emulator is to take down.
What’s Earth’s dominant species by mass? — 290 points
290 points · 219 comments · signoregalilei.com · HN discussion
A careful walk through the biomass numbers, and the answer is not close. Humans are about 350 million tonnes; Antarctic krill are 379 million, and cattle — the single largest animal species by mass — are around 650 million. Then the units change and the question stops being interesting for animals at all: measured in carbon, all animals together are 2 billion tonnes, which is less than protists (4 billion), archaea (7), fungi (12), bacteria (70) and plants (450). The winner among plants is not a cereal crop but black spruce, at roughly 11 billion tonnes of biomass. Sugar cane is the largest crop by annual harvest at over 2 billion tonnes, about six times humanity’s collective weight, and it is a perennial, so its standing mass is higher than a single harvest implies.
The best line in the comments is that all animals, including us, are about 0.4% of total biomass and humans, livestock, birds and wild mammals together are about 0.02%. Which is the correct frame for the whole genre: “dominant species” is a question about one rounding error inside another, and the biomass rankings are mostly a map of which kingdom got the terrestrial surface first.
Strands Decider 2B: a small, open-source, decision model — 274 points
274 points · 77 comments · strandsagents.com · HN discussion
A 2B-parameter model that answers yes/no questions instead of generating text. Architecturally it is a Qwen3.5-2B torso with the language-model head removed — no text generation at all — replaced by a pointer head that scores each supplied option against the hidden state at a <answer> marker. The head is about a million parameters; the torso is fine-tuned with a rank-16 LoRA. Median latency is ~115ms on an RTX 3090 and ~153ms on an M3 MacBook. The weights, training data and scripts are all on GitHub and Hugging Face. The team claims third of 33 in the 2B class on JevBench accuracy and first of 30 excluding the just-over-2B models.
The omission is deliberate and correctly argued: generating all outputs in one parallel pass makes these models worse at complex problems than reasoning models, and no text output makes them useless for chat or summarisation. What they do give you is a calibrated reliability score, which frontier LLM APIs do not expose, and the ability to ask several questions about the same prompt cheaply. The demo in the post is the interesting part — a decider reads a proposed tool call and answers two questions about it (are the arguments grounded in anything the user said; is it premature to call) and the agent asks for a city instead of confidently reporting weather somewhere nobody mentioned.
Note the collision: this and OpenAI’s Decisions API launched the same week with the same thesis, that the next layer of agent infrastructure is a cheap calibrated classifier, not a bigger chat model. The thread’s main complaint is the vocabulary — “noul” for binary decisions, borrowed from Jev — and one commenter’s suggested alternative, a decision model backed by a human named Jerry, is the most honest product pitch in the thread.
Nobel Prize in Chemistry 2026 to Henri B. Kagan and Kenso Soai — 262 points
262 points · 47 comments · nobelprize.org · HN discussion
Awarded for non-linear effects and autocatalysis in asymmetric organic synthesis — the chemistry behind how life ended up with only one handedness of amino acid when chemistry produces both equally. Kagan’s 1986 discovery showed you could push a reaction further toward one mirror image than anyone thought possible. Soai designed the first reaction with the potential to be homochiral in 1995 and got it to actually produce only one mirror image in 2003. Per the committee, nobody other than life itself had previously managed it.
The practical stake is not philosophical. In pharmaceutical development only one mirror image of a molecule does what you want, so a reliable way to drive a reaction to one handedness is the difference between a synthesis that works and a racemate you have to separate. The best comment in the thread is not about chemistry at all: the chirality of protein motors in embryonic cilia produces a leftward flow of water, which is how the left-right asymmetry of a developing body gets established. Biology is not just built from one-handed molecules; it uses their handedness to decide which side of you the heart goes on.
Paramount Skydance has completed its $111B merger with Warner Bros. Discovery — 262 points
262 points · 472 comments · arstechnica.com · HN discussion
The deal closed after Supreme Court Justice Elena Kagan rejected a last-ditch attempt to block it. The combined company is called Skydance and holds two major studios, Paramount+ and HBO Max, CBS, CNN, and live sports including CBS Sports and TNT Sports. A July ruling from Judge Araceli Martínez-Olguín in the Northern District of California had found the combination would likely substantially reduce competition and violate antitrust law; California settled last month and the other eleven plaintiff states went along. Advocacy groups had urged the judge to reject the compromise, and did not get their way.
472 comments on this one, and the comment worth carrying forward is the one that notes a pattern rather than an outcome. A Verge trope, quoted in the thread: the only workable antitrust policy for the United States would have been to simply forbid anyone from buying Time Warner ever again. 2001: AOL buys Time Warner. 2018: AT&T buys Time Warner. 2022: WarnerMedia spins out of AT&T and merges with Discovery. 2026: Paramount buys it. Four restructurings in twenty-five years, each justified by the failures of the last.
The other substantive point in the thread is the debt. The new entity enters with a heavy load, and the numbers are not flattering against the competitive picture: YouTube holds roughly 13% of US viewing time against about 6% for the combined Paramount and Warner assets. Buying scale in linear television is not obviously a strategy for the next decade, and consolidation that requires debt to finance declining businesses tends to resolve through cuts, not investment.
Claude Code’s suggested message feature: I think the real customer is the model — 260 points
260 points · 150 comments · zohaib.cc · HN discussion
Claude Code started pre-filling the prompt box with a suggested next message after finishing a task — “run the tests,” “commit this.” The author’s argument is that this is not a user feature at all: it is an RLHF annotation pipeline that users run for free, without realizing they are annotating. Sending the suggestion untouched is a positive label; editing it produces a preference pair, and the diff between the prediction and the edit shows exactly where the prediction failed, written by someone who knows the codebase, at the moment they care about being right.
The argument is well made and the rebuttals in the thread are better. One commenter notes that feeding a prompt ending in a user-turn marker into an LLM already produces plausible next-turn predictions, because the loss function does not distinguish between the two sides of a conversation — so Anthropic likely got this behaviour for free and only recently decided to expose it, which weakens the “training data pipeline” claim without undermining the observation. The sharper objection is that showing the user the prediction biases the user’s response toward it, making the resulting label less independent than an unprompted edit would be. The author’s own framing survives both: it does not matter much whether this is the primary purpose, if the label stream is a side effect it is still a label stream, and nobody is asked.
The comment that best captures the actual user experience is from someone whose Claude changed a feature they had not asked to be changed, and whose suggested next message was, roughly, “revert the change to it.” The model knew. It did it anyway.
GPT‑6 and Intelligent UI for everyone — 256 points
256 points · 123 comments · openai.com · HN discussion
GPT-6 rolls out to ChatGPT’s free tier, and with it “Intelligent UI” — responses that compose text, graphics, tappable buttons, forms, charts and interactive diagrams instead of paragraphs. The bicycle example in the post shows a labelled diagram with selectable subsystems replacing what used to be a wall of text about frames and drivetrains.
Two reactions dominate. The first is that OpenAI’s release pages are the best-designed product marketing on the web, which is a compliment that quietly concedes the announcement is marketing. The second is aesthetic revulsion — one commenter describes the generated pages as so much needless whitespace and checklists that it feels like being condescended to, and worries about what happens when this format leaks into developer tools. The more interesting question, buried lower in the thread, is latency, and whether generating a bespoke UI for every question costs enough to matter. Nobody in the thread has a number for that.
The honest version of the counterpoint comes from someone pointing out that Bartosz Ciechanowski’s hand-built interactive explainers were the last thing anyone expected to be automated, and that they will still be better than anything generated. Both things are true: the generated one is available now for any topic, and it is worse.
State of Devs 2026 — 233 points
233 points · 158 comments · 2026.stateofdevs.com · HN discussion
5,463 respondents, fielded July 5 to September 5. Nearly half had experienced job insecurity and insufficient wages; a quarter had been laid off at some point in their career and nearly one in ten within the last nine months; over a quarter expect to change careers within five years for reasons beyond their control. Two thirds report reduced motivation or increased cynicism recently, around half report trouble focusing, more procrastination, and feeling drained. 62% report having experienced burnout at some point — identical to last year.
Two caveats before reading anything into this. A self-selected survey of 5,463 people who answer developer surveys is not a population, and the comparison to last year’s burnout figure is a null result being reported as stability. What the survey actually establishes is that the people motivated enough to fill it in are frightened, and the comment that makes this vivid is someone who answered “somewhat unlikely” to the five-year career question in July and would now answer “somewhat likely” — a shift inside three months.
The sharpest comment is unrelated to the data: the “bad management” category at 63% is meaningless without a definition, because developers routinely file “I disagree with what I am being told to do” under bad management when what actually happened was that nobody communicated the reasoning. That is a real measurement problem and it applies to most of the chart on this page.
Penguin Mail — open-source Rust email client for Linux with AI — 229 points
229 points · 169 comments · penguin-mail.com · HN discussion
A GPL-3 licensed Linux mail and calendar client, version 1.0.0, supporting Gmail, Microsoft, generic IMAP/POP3, CalDAV and CardDAV. OpenPGP and S/MIME through your own GnuPG with no key custody. Gmail filters, Microsoft inbox rules and Sieve rules surfaced as editable plain-language rules. An optional assistant that runs locally through LM Studio or Ollama, is off until you pick a model, and asks before sending mail or changing settings. No server of its own, no tracking, no ads; Gmail talks to Google and Microsoft accounts talk to Microsoft Graph.
That last set of constraints is the correct architecture for a mail client and almost nobody builds it, which is why this is on the front page at 229 points despite version 1.0.0 being a risky thing to point at your primary inbox. The thread’s first question is also the right one: how well does it sanitise HTML email bodies? Thunderbird goes to considerable lengths to prevent inline JavaScript and tracking pixels, and a client that just renders mail in a WebView would be a regression on a threat surface users have learned to ignore. The second complaint is smaller but telling — Fastmail is named as a supported provider, and JMAP is not supported.
Incident with Git Operations, Pull Requests and Actions – Resolved — 222 points
222 points · 175 comments · githubstatus.com · HN discussion
GitHub reported widespread impact across Git Operations, Pull Requests, Actions, Issues and Webhooks between 15:06 and 15:16 UTC, then walked the status page through a sequence of “experiencing degraded performance” notices and “mitigated, monitoring for stability” updates for roughly the next twenty minutes before declaring recovery. Root cause “identified”; a detailed analysis is promised. Total user-visible disruption: about an hour.
What makes this newsworthy at 222 points is neither the outage nor the duration. It is that commenters were still getting Internal Server Errors after the status page said “Resolved,” a premature all-clear that one called “pretty bad user experience.” Another spent five minutes debugging a failing build before remembering that checking a status page was the faster diagnostic. A third asked whether 89.99% counts as three nines — which is the right question for a platform that every CI pipeline and dependency in the industry is now a single point of failure behind.
Still on the page
Twelve of the thirty-one stories that cleared 200 points today were covered in earlier roundups. Deltas are against the last roundup that tracked each story, which for all of them is yesterday’s.
The top of the list is the one that moved most and matters least. Mistral Large 4 went 1,261 → 1,980, up 719, holding rank one with 1,184 comments — a public preview of a model whose weights arrive at the end of the month, and a second submission of the same story (“Mistral Large 4: Le Chonk”) sits at 521 points with no coverage of its own. Anthropic reported diary entry to police continues its climb from 795 → 822, up 27, after yesterday’s 480-point jump; it remains by far the most-discussed item in the window at 667 comments. Web Search API is 581 → 588, up 7.
Two of yesterday’s full sections grew substantially. OpenTPU went 159 → 332, up 173 — yesterday it was “AI is now capable of developing its own inference hardware” at 159 points, and it has roughly doubled after a day of people reading the SystemVerilog. Nature’s capacity to ‘bounce back’ went 264 → 342, up 78, and Polars 2.0 went 355 → 454, up 99. The Nobel Physics announcement moved 444 → 568, up 124, as the prize’s citation profile circulated. Smaller movers: Gleam 252 → 313, up 61; Example.com 318 → 356, up 38; JetBrains’ 2025 results 553 → 590, up 37; Beam 526 → 549, up 23; Opus 5.5 agents and the magnetic semiconductors 475 → 488, up 13.
The shape of this list is the same as yesterday’s, one day further along. The front page is showing twenty-odd items and the rank list is carrying thirty-one above 200 points, and the difference between the two is the accumulation that happens after a story leaves the page. That gap is now roughly two-thirds of the day’s material.
Throughline
First: the industry shipped two copies of the same product this week, and neither one is a chat model. OpenAI’s Decisions API and AWS’s Strands Decider 2B launched within days of each other with the same thesis — that the next layer of agent infrastructure is a small, fast, calibrated yes/no classifier, and that the reliability score matters more than the answer. One is a proprietary endpoint on a single permitted model; the other is 740M parameters of weights, training data and scripts on Hugging Face. What they agree on is more revealing than what they disagree on: the frontier labs are no longer racing to make the smartest model answer your question, they are racing to make the cheapest one stop asking it. The independent calibration test in the Decisions API thread — a 50% true probability coming back at 86% on choice questions — is the kind of finding that only exists because someone outside the vendor bothered to check, which is going to be a recurring problem when the pitch is “trust our confidence score.”
Second: verification is the scarce commodity, and today’s page shows what it looks like and what its absence looks like. OpenAI’s mathematics release is the one announcement in weeks that ships checkable artifacts: Lean formalizations, published failures, compute estimates in ChatGPT-Pro-hours, and an advisory group with a name. That is the bar, and it is set low enough that everything else fails it. Haiku 5.5’s benchmark table is vendor numbers on vendor benchmarks, and the independent tests that followed are less flattering than the table — the model is fast, cheap, and not good enough for complex UI, a judgment it apparently reached about itself before delegating. EmbeddingGemma 2’s headline multimodal claim does not reproduce for text-only retrieval in early testing. AnyPS5’s 87% is a ratio whose denominator the project defines, and a commenter who read the code calls the “mapped” functions mostly stubs returning OK. The gap between the release note and the artifact has become the standard shape of a launch, and the only reliable way to close it is to read the footnote, look at the source, or wait for someone with a benchmark of their own.
Third: the capability shipped ahead of the controls, and the disclosures are in footnotes. Meta’s Muse is the clearest case — a zero-day, root access via impersonation, private messages uploaded against explicit instruction, and a Mac privacy model so broken Apple had to change macOS settings to contain it, all under a privacy-first banner. Claude Code’s suggestion box is the soft version of the same problem: a labelling pipeline that improves the model using work users are doing anyway, requiring no consent because nobody is asked. The Decisions API wants to route your requests on a probability you cannot audit; GPT-6’s Intelligent UI replaces prose you can skim with a generated interface you cannot parse at a glance. None of this is new in kind, and all of it is new in scale. The thing that changed is that the permission dialogues are now the security boundary for systems holding your messages, money and calendar, and the evidence that users can evaluate those dialogues is a YouTuber who handed over his Marketplace account and lost his home address.