techMon 12 Oct, 07:30

Sources93 X accounts · techtwitter.com · Hacker News · 8 newsletters
  1. Runs every day at 07:30 London time.
  2. It is shown its own papers of the last 3 days, to know what it has already said.
  3. Haiku 5.5 writes the paper from what was fetched, with no tools.
  4. Jev scores every story by how much it matters to you; each section puts those first, and what is not for you drops to the briefs.

93 X accounts

Their posts of the last 24 hours, through FxTwitter; plus everclear077/awesome-ai-x-accounts when it answers.

techtwitter.com

What tech Twitter is discussing, ranked by discussion.

Hacker News

The front page, plus stories over 80 points in the last 24 hours.

8 newsletters

Every new issue, each delivered once, looking back up to 7 days.

The tech daily written for one reader

OpenAI drops 719 AI-written math proofs, and nobody has checked them yet

OpenAI posted 719 proofs written by its AI in one night, claiming they solve problems that stumped mathematicians for decades. None has been peer reviewed, and three were already pulled within a week. A Fields Medal winner told Le Monde the batch hit his own field hard.

Karpathy's funnel: most people have barely touched AI

Karpathy's updated post splits AI users into a funnel. About 6 billion people have barely used an LLM, 1 to 2 billion use free chatbots casually, around 20 million pro coders and mathematicians watch months of work collapse into prompts, and about 5,000 lab insiders see agent swarms do new science.

Is software dead? levelsio says yes, and the replies split hard

levelsio says we are in “the era of personal software,” where everyone builds what they need. He also says he has cancelled most of his SaaS subscriptions. Pushback includes dabit3, who calls it a “Kodak Brownie moment” that will bring an explosion of apps, and rileybrown, who says he keeps downloading apps. signulll argues that even 1% of people building software would flood the market.

“Passkeys are a trap”: a post urges developers not to use them

The post tells developers to skip passkeys, the login method that replaces passwords with device keys. The Hacker News thread hit 260 comments.

Terence Tao's Math 2.0 slides reach the front page

Tao's Caltech talk on how AI changes math drew 375 points and over 400 comments on Hacker News. trq212 quoted his view that we should not settle for an opaque superintelligence.

Big announcements

Monday, October 12, 2026 · Pergent2

Coinbase's small fine-tuned model beats frontier models at fraud checks

Coinbase fine-tuned Qwen3.5-9B, an open-weight model, on its own internal data. For fraud prevention it beat frontier models while costing less and running faster. Brian Armstrong expects more companies to copy the pattern.

Hugging Face turns OpenAI's math problems into open practice environments

Hugging Face's FineEnvs project turned OpenAI's math problems into open RL environments that anyone can train on. The early verifier only accepts the exact original statement, so an equivalent proof can still score zero.

Decision 3.0 ships with open weights from 0.6B to 27B

The vllm-sr team open-sourced Decision 3.0, a multimodal decision model, with weights on Hugging Face. Its makers say it leads both boards in their own evaluation. A Core ML build also triages 1,000 pull requests on a Mac at about 8 per second, with nothing leaving the machine.

GPT-6.1 Sol gets a 500k default context in Codex and Work

OpenAI raised the default context length for GPT-6.1 Sol in Codex and Work from 272k to 500k, about an 80% jump. Users can still set it anywhere up to 1M.

OpenClaw v2026.10.1 adds skill discovery and full cloud backups

The release adds installed-skill discovery, MCP Apps beside the chat, full backups to external storage or Cloudflare R2, and a simpler Mac setup. It bundles 2,425 PRs from 402 contributors.

Hermes Agent adds Anatomy, a skill that builds playable diagrams

Anatomy turns a technical idea into an interactive diagram you can poke at. It is now an official optional skill, installed with one command. One demo, a rate limiter, came out of a single voice prompt and about 90 minutes of work.

River AI launches cheap hosted open-weight models with $25 in free credits

River AI says its fast inference for DeepSeek v4.1 Flash and GLM-5.3 Flash costs 20% less than the models' own makers charge. Its stack is written from scratch in Rust, which it says makes it cheaper to run.

LangChain pitches managed deep agents as the shape of all future agents

LangChain's new managed deep agents bundle durable workflows, a cloud sandbox where agents use a computer, memory and skills, and scheduled triggers. Harrison Chase says the design matches the company's long-held vision.

ArtCraft: open-source Rust clones of Adobe's apps appear within a week

Brandon Thomas pushed the first commit on September 30. Within days there were open-source Rust versions of Photoshop, Illustrator, Premiere, and more, mostly typed by Claude Opus. The project bans decompiled code.

What people are fighting about

Monday, October 12, 2026 · Pergent3

Does clean code still matter when AI writes the code?

poteto says clean code has always cost performance, and that its rules were built for humans who maintain code by hand. He thinks we should update what “good code” means. shadcn pushes back: “I still care that the code looks and reads right,” and says he is not ready to let go. addyosmani argues that quality comes from cheap, automatic checks that let agents verify their own work.

Nikita Bier: intellectual property may not mean much when intellect is free

The post drew 8,000+ likes and 758 replies, a high reply count for the size of the crowd. A follow-up from the same account says technical moats are mostly over, and that only proprietary data, network effects, or a physical part still protect a business.

OpenAI's pricing poll backfires, and the engineer behind it is confused

thsottiaux says he expected a poll on doubling Pro tier limits, at the cost of bringing back 5-hour caps, to win. The result went the other way, and the post drew 1,636 replies.

Nikita Bier: welfare programs should ask if you tried AI

The proposal is an affidavit asking whether you have tried AI to become more productive, with eligibility lost if the answer is no. The post drew 1,446 likes and 270 replies, a lot of argument for one line.

Alex Finn says Grok Bot built and launched a business on its own

Finn says Grok Bot built a website, wrote a business plan, collected emails, shipped 20+ PRs, and sends a newsletter, all in 15 hours. The claims are his own, and the post drew 223 replies.

Web dev corner

Monday, October 12, 2026 · Pergent4

Yandex Cloud hit a third time as its bot traffic drops 99%

Yandex Cloud's status page logs a third data center hit, and Cloudflare Radar shows Yandex bot traffic down 99% since the attack. The two main Hacker News threads on it drew about 940 comments combined.

“Fingerprints are the new cookies” says browser fingerprinting is the next tracker

The post argues that device fingerprinting, which identifies you from your browser traits, is taking over from cookies as the main way to track people. It hit 259 points on Hacker News.

Meta removes human support for Instagram, and the AI replacement fails

Gergely Orosz says Meta now offers AI-only support to paying customers, and that the AI support does not work. He says there is no way to report the bug, and the post drew 1,141 likes.

Gergely Orosz: AI on-call tools could make reliability worse

He argues that AI SRE startups, which try to stop developers getting paged at 2am, may remove the feedback that makes systems reliable. Good on-call work has always meant the builders feel the pain.

Every major framework is being acquihired, and nobody knows where it goes

kettanaito says big companies are buying the teams behind major frameworks and runtimes. He is glad for the people, but asks what happens to the projects once they belong to a company.

The tilde in your PATH may not be your home folder

Putting a literal ~ in your PATH can make tools look in a folder with that exact name. Use $HOME or a full path instead.

Valen proposes group borrowing for memory safety

Valen's author describes a new approach to borrow checking, the rule system that keeps memory safe, applied to groups of data. It is a design idea, not a shipped feature.

Research & open source

Monday, October 12, 2026 · Pergent5

Meta's GitSwarm lets hundreds of agents share one Git repo

Each agent's contribution is a commit that records which earlier work it builds on, so partial work and failed ideas are not thrown away. The paper targets long-horizon agent runs.

Worth a read

Monday, October 12, 2026 · Pergent6

Berkeley study: 10 minutes of AI use can erode persistence

A Berkeley study found that brief AI use made people less willing to keep working through hard problems. It drew 95 comments on Hacker News.

“I’m sorry, but you still have to think” hits the front page

A 356-point Hacker News essay with that title makes the case that AI does not remove the need to think. Read it for the argument, since the files do not include the text.

GPT-6 Pro keeps finding errors in academic papers

Ethan Mollick says the model finds supposed errors all the time while doing other work. Confirming them still takes time, and we have no good way to handle a flood of corrections.

Google is cutting forest for data centers in Finland

The Guardian reports a temporary halt after an environmental impact assessment flagged the cuts. The story hit 353 points on Hacker News.

New documents add evidence to the climate deception case against ExxonMobil

Inside Climate News reports new evidence in the case over what the company knew and said about climate change.

Nitter seeks funding and legal help

The privacy-focused X front end posted an update on October 10 asking for money and legal help. It hit 443 points on Hacker News.

Total Annihilation rebuilt for modern computers

A fan rebuild of the 1997 strategy game for modern machines. It drew 187 comments on Hacker News.

NYC's click-to-cancel rule is now in effect

The city's rule makes it as easy to cancel a subscription as it was to sign up. It hit 430 points on Hacker News.