techSat 17 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

Reflection's Beam 1.0 is out as a free, open 501-billion-parameter model

Beam is a mixture-of-experts model, so only about 23B of its parameters work on each token, which keeps it cheaper to run than its size suggests. It has a 1-million-token context window, ships under the Apache 2.0 license (free to use, even commercially), and had vLLM, SGLang, Cloudflare and Vercel's AI Gateway support on launch day, with Vercel offering it free for now.

Claude reveals where the ECDSA curves came from

Anthropic says Claude helped answer where the standard ECDSA curves came from, a question that has nagged cryptographers for decades. A post going around says a $36,864 bounty was posted for the puzzle and Claude solved it in under 30 minutes, which leaves people asking why nobody solved it sooner.

Poteto's checklist for trusting your agents more

Poteto says agents copy whatever exists and take the shortest path, so make the right thing the easy thing. He wants better architecture, lint rules and tests, plus clear AGENTS.md notes, to catch problems before they happen.

🇪🇺 Should Europe try to build its own frontier AI?

Arthur Mensch argues that leaving frontier AI to the Americans and Chinese would be a fundamental mistake. dhh replies that Europe's despair is the real problem, and levelsio jokes about how fast the frontier is being paced.

Bun wants to compile TypeScript into fast machine code ahead of time

Bun's new flow runs bun build --compile --aot, then type-checks, optimizes and emits machine code, and it has already sped up JSON.stringify by 2.2x in one benchmark. Vercel's Rauch and cramforce both say their TypeScript builds beat the Rust-based compilers, with Rauch citing 2 to 4 times faster than tsgo.

Big announcements

Saturday, October 17, 2026 · Pergent2

Tesla's Roadster gets a closer look: thrusters, flight rules and a playable demo

Tesla's hardware VP says the car uses cold-gas thrusters, the same kind satellites use, to hit 0 to 60 in 0.55 seconds, with nitrogen stored at 10,000 PSI. One commenter argues federal altitude rules would block the 'hops' over towns that the demo implies, and a developer rebuilt the demo as a playable simulation. Sam Altman also called it amazing.

iPhone Duo pre-orders open today

Apple's foldable iPhone Duo goes up for pre-orders starting today, the biggest consumer hardware news in the feeds. A post from Cognition shows its Devin agent testing the folding screens, a small sign of how developers will need to test it.

Perplexity open-sources two new Decision Models

pplx-decider-flash-v1.2 (4B parameters, 1 cent per million tokens) and pplx-decider-v1.2 (27B, 2 cents) are both open weights on Hugging Face. Perplexity says inference is up to 5x faster because all questions in a request now run in one pass, and the 4B model is the top small model on the Decision Index.

What people are fighting about

Saturday, October 17, 2026 · Pergent3

Anthropic's own AI bill gets praise and pushback

Politico reports Anthropic wrote its own frontier AI bill and is pitching it to lawmakers. Charlie Bullock says it is a reasonable start but weaker than the FRONTIER Act in key ways, and that new bills should go further now that the politics have moved toward safety. Separately, New York City's comptroller urged Dario Amodei to delay Anthropic's IPO.

Are AI skeptics in denial, or just not using the tools?

Gergely Orosz says he keeps meeting devs who call anyone who sees LLMs as permanent 'AI psychosis', yet most professional teams do the same things. The reply thread ran to 93 comments, and his other post notes that AI spend runs from about $100 to $5,000 per dev per month, with few teams measuring return.

Will AI eat the software dependency?

shadcn argues that teams will stop pulling in outside code and instead keep an internal registry that agents reshape on demand. It drew 137 replies on 1,225 likes, which marks it as a real argument rather than a cheer.

Web dev corner

Saturday, October 17, 2026 · Pergent4

Postgres 19 is delayed because AI reviewers find more bugs than humans can fix

The release is pushed back after AI code reviewers flagged more issues than the team can work through. It is a sharp sign of what AI scanning does to a mature project: it finds problems faster than people can close them.

Anthropic says Claude writes about 90% of Claude Code's code

Changes in Anthropic's monorepo rose from 2,500 a day in May to over 30,000 in September. Merged PRs more than doubled since June while reverts fell by over 40%, because agents write the code and agents review it. A human still pulls the plug when agent authors and reviewers loop too long.

A 37-year-old kernel bug in OpenBSD gets exploited

Researchers at calif.io describe an exploit for a bug that sat in OpenBSD for 37 years. It is a reminder that old, quiet code can hide serious holes.

Supabase Middleware runs one set of request logic on any runtime

Write per-request checks once, such as who is calling, and run them anywhere that has fetch. Bridges for Hono, H3, Elysia, NestJS and TanStack Start are included.

Vercel makes agent turns survive crashes in fx.sh

The same createFxAgent() can now be durable on a laptop or in the cloud, through files, Postgres or Vercel Workflows. If the process dies mid-tool-call, it resumes from there.

A user says Anthropic Console charged their card 100x

The overcharge was traced to an input bug in the international version of the Console. The thread is a reminder to check your billing page after updates.

From the newsletters

Saturday, October 17, 2026 · Pergent5

Eight states reportedly sue OpenAI over failing to protect vulnerable users

The suits focus on kids' online safety and claim OpenAI did not do enough to protect vulnerable users.

OpenAI agrees not to use trade secrets Apple says were stolen

The deal comes as the two companies head to court in a high-stakes AI trade-secret fight.

Bank of America hires 1,000 contractors after Mythos finds flaws

After Anthropic's Mythos model found flaws, the bank is paying for a large contractor push, a sign that AI security testing is turning into hiring.

Research & open source

Saturday, October 17, 2026 · Pergent6

Four Gemini 4 Argon agents beat Minecraft from scratch in 27 hours

Each agent controlled one player through the Mineflayer library, with no assigned roles and no human help. Together they wrote and debugged 2,783 programs, defeated the Ender Dragon, and shared lessons as they went.

Claude designs quantum circuits that beat published ones

In a paper with Harvard, Claude Mythos 5 proposed circuits that beat the best-known designs on seven of eight tasks, with fewer Toffoli gates and fewer qubits. An independent evaluator checks correctness and cost, and the best designs seed the next round.

Frontier models now solve hard theory questions far more often

A new MIT study says frontier OpenAI models rose from 3% to 35% success on tough math, physics and economics questions in 16 months. A good attempt costs about $15, and the study says checking answers may be the new bottleneck.

Mercor says 90% of US working hours fall outside AI benchmarks

Mercor maps tasks against benchmarks and finds most of the real work sits where tests do not reach. That is a useful check on any claim that AI is 'ready for jobs'.

AMD's MI355X nearly matches Nvidia's B300 on Kimi K3 serving

8 MI355X GPUs deliver 24.4% more throughput than an HGX B300 box on the SemiAnalysis agentic coding benchmark, and 87.4% more tokens per dollar. vLLM's optimizations and AMD kernels closed the gap.

Worth a read

Saturday, October 17, 2026 · Pergent7

How vLLM went from a 2022 side project to running on about 500,000 GPUs

a16z tells the story: a Berkeley student set out to speed up a Meta model in two weeks. The team borrowed virtual memory from operating systems, which is why the 'v' in vLLM stands for virtual.

You could have invented transformers from memory

An essay walks through the associative-memory idea that leads straight to attention, step by step.

Llama.cpp runs on a jailbroken PS5

A project runs any GGUF model natively on the console, a fun hack for anyone who likes odd hardware.

Waymo is headed to Toronto

Waymo's blog announces its Toronto plans, and the post drew a very long Hacker News thread.

Scott Aaronson says AI one-shotted open problems from his old papers

He says problems that would have taken him about 100 hours now fall in one try, and that AI is wearing down his own math skills.

🇨🇳 China signals more scrutiny for high-risk open models

Nathan Lambert says the new ministry statement does not end China's support for open weights, but it singles out high-risk models for review, and implementation will decide everything.