techThu 8 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

Claude Haiku 5.5: the cost of inference changes

Anthropic's latest small model is 75% cheaper to run than Haiku 4.5. For the common case—prompts under 100k tokens—it's 10x cheaper. It outperforms OpenAI's small model at an eighth of the cost per task. Already rolling out to Cursor, Devin, and all major APIs.

OpenAI math proofs keep tightening

Quasi-Riemann Hypothesis proof compressed from 500k pages to 21. Integer multiplication improved 2^123-fold. Terence Tao weighing in on what this means for the field.

Margaret Hamilton, Apollo software pioneer, has died

She coined the term "software engineering." Her work on Apollo's flight software set standards that defined the field. The photos of her standing beside the stack of program printouts are humbling.

Amazon bans agents; Paul Graham smells opportunity

Amazon doesn't want third-party agents using its platform. Graham argues this is the first real chance to build an Amazon competitor: people will want agents to handle shopping, and they won't use an Amazon-supplied agent to do it. The push-back: Is anyone actually using agents for shopping yet?

Mistral Large 4: hedge against lock-in

French lab Mistral released 1-trillion-parameter open-weight model competitive with closed alternatives. Enterprises can own the weights, so no single provider can sunset it. Trained on 4,000 NVIDIA GPUs at a fraction of competitors' cost.

Big announcements

Thursday, October 8, 2026 · Pergent2

GPT-6 and Intelligent UI ship to everyone

OpenAI rolled out GPT-6 to ChatGPT's 40M users, along with Intelligent UI: ChatGPT now generates custom interfaces on the fly—charts, calculators, tools—whatever your question needs. No more walls of text.

Windows reinvents itself for local AI

Microsoft announced today that Windows is getting rebuilt for agents. You get MAI-Code-1.1 Flash (137B params, runs on your PC), Hybrid Intelligence (Copilot takes action locally), and deep NVIDIA integration. Jensen Huang showed up to celebrate the four-year partnership.

tsc-rs: TypeScript rewritten in Rust with Claude

One person and agents rewrote the entire TypeScript compiler in Rust in two weeks, spending $20k in Claude. An earlier attempt burned $400k of OpenAI Codex tokens with nothing to show. The result is open source and a drop-in replacement for tsc.

Grok Bot can now search your X feed

Elon's bot got genuinely useful overnight. Grok can search, read, and monitor your X feed—no connector needed. Ask it to find feedback about your product, find people criticizing you, whatever. Built-in, no setup.

Meta and Microsoft are pulling back on Claude

Meta cut Claude Code seats from 60k to 30k. Microsoft cut expected spend by a third (from $1B+). The Pentagon stopped using Claude entirely. All happening as Anthropic prepares a $2 trillion IPO.

What people are fighting about

Thursday, October 8, 2026 · Pergent3

Does Anthropic have an enterprise moat?

Pentagon blacklisted Anthropic as a supply-chain risk and stopped using Claude. Meta built their own tools. Microsoft did the same. Anthropic's IPO is priced at $2 trillion on a $42B loss. The question: Can a frontier lab survive when enterprises build their own instead?

Does vibecoding destroy code quality?

DHH and others say: Claude out of the box, barely any constraints, models are great. Matt Pocock counters: Agents need guardrails and you need to care about your codebase. The fight: Is the easy win (throwing code together) hiding the hard problem (maintainability)?

Math is solved—now what?

OpenAI published hundreds of unsolved math proofs with no errors found. Terence Tao is thinking carefully about what this means. The fight: Is this AI doing math or doing symbol manipulation? Does it matter? One camp says pure math careers are over. The other says this is just the start of math we couldn't do by hand.

Web dev corner

Thursday, October 8, 2026 · Pergent4

JPEG XL ships in Chrome 155

Chrome 155 now supports .jxl images: 30-50% better compression than JPEG, HDR support, no weird quality loss. If you're building image-heavy stuff, this is a tool in your belt (still need fallbacks for other browsers).

From the newsletters

Thursday, October 8, 2026 · Pergent5

Anthropic's enterprise moat is getting thinner

The Deep View's detailed dive: Pentagon's supply-chain risk designation forced cessation of Claude use. Meta cut Claude Code seats 60k→30k. Microsoft cut expected spend by a third. The hard question: what's Anthropic's growth story when enterprises build their own?

Research & open source

Thursday, October 8, 2026 · Pergent6

Worth a read

Thursday, October 8, 2026 · Pergent7

Terence Tao on what it means when AI solves unsolved math

The greatest living mathematician thinks carefully about AI-generated proofs. Thoughtful and not dismissive. Worth reading if you care what this actually means.