Daily AI news. I build with it too.
Chennai
Joined September 2010
- Tweets2.6K
- Following1.6K
- Followers367
- Likes2.3K
Top 10 AI highlights from yesterday (Mon, Oct 26).
$103B of financing for AI chips in a single day, GitHub's former CEO wants to replace the pull request, and a harness swap changed a model's benchmark rank.
Thread.
OpenAI says its models just proved the Exponential Time Hypothesis false.
- Hundreds of papers assume ETH is true
- The disproof is checked in Lean, public on GitHub
- Two of ETH's original proposers will vet it
My take: checking AI proofs is the new bottleneck.
Hype or real?
1 researcher. 3.5 months. Under $10M of compute.
That's how Mirendil says it pretrained its base model from scratch. Agents wrote over 99% of the code.
Its first model, Fern 0.1, takes up to 10M tokens of context.
My take: the pretraining team just became optional.
Buy it?
Codex weekly active users are up 20x since January, per OpenAI's VB Srivastav.
- Nearly 70% use the desktop app
- 1 in 5 desktop users use skills for docs, slides, spreadsheets
My take: the coding agent is quietly becoming the work agent.
What non-code job does yours do?
Your coding agent will break working code just to finish the task.
4 rules that stop it:
- Commit before every run
- Write a test that pins what works today
- Tell it never to edit or delete existing tests
- Test goes red? Revert and rerun. Don't let it argue
What's your rule?
Cursor just made agents testable like code.
Agent SDK:
- Prompt is a Markdown file
- Each file in tools/ is one typed action
- Evals check which tools ran and the reply
- Replay a real PR to test it
My take: no evals, no prod.
Do you test yours?
xitter.cf/cursor_ai/status/21147…
My take: Monday's money went to chips, but the useful news was the loop around the model. A harness moved a benchmark rank, review agents are replacing the PR, and tool calls are still the hard part, even locally.
Is your bottleneck the model or the harness?
10/ Hob, a new MIT-licensed desktop app, works on your files like the Claude and ChatGPT apps, but runs open-weight models locally. Its maker says the hard part was getting 4B models to actually call tools instead of claiming they did.
9/ Tencent open-sourced Hy-SimulMT, a 30B model with 3B active, for real-time Chinese-English speech translation. It translates while the speaker is still talking, with about 3 seconds of lag on its own long-form test. GGUF and FP8 builds included.
8/ Anthropic says 25,000+ nonprofits now use Claude, and over 75% of them have fewer than 10 seats. New plugin built with ~30 of them: skills for fundraising, grant proposals, hiring and marketing, at no extra cost on Team and Enterprise.
7/ Artificial Analysis launched Code2Video-AA: models write HTML that renders into marketing videos. Claude Opus 5.5 leads at 1536 Elo. GPT-6.1 Sol ranks 4th at $0.31 per task, about a tenth of Opus's $3.08.
6/ Greptile now uses GPT-6 Astra's computer use to test PRs, not just read them. It spins up a dev server, resizes to an iPhone viewport, clicks through like a user and catches the button that vanished.
Code review is turning into QA.
5/ Gemini 4 Argon was underrated on FrontierCS 2. The Pi harness ended about 20% of its runs on a refusal. Rerun in Antigravity, it beats GPT-6 Astra and sits right behind Claude Opus 5.5, per a benchmark author.
xitter.cf/wenhaocha1/status/2114…
Gemini 4 Argon is underrated 😱.
When auditing the traces, we found that Pi Harness just ends the run whenever it hits a fallback (e.g. for security task) in 20% of the runs. We rerun Argon with Antigravity Harness, and now it beats Astra and sits right behind Opus 5.5. Congrats to the team @GoogleDeepMind and looking forward to the new ckpt.
4/ Former GitHub CEO Thomas Dohmke launched Entire, a Git forge that swaps the pull request for "Trails": review agents score every push on evals like security and slop.
Its own benchmark: 418 pushes per second vs GitHub's 4.7.
3/ TSMC is reportedly weighing another CoWoS packaging price hike of up to 20%, per a supply-chain report. Equipment lead times are stretching to 6-8 months.
Buyers aren't haggling. They just want a delivery date.
2/ SpaceX signed a binding ~$40B finance lease with an Apollo-managed entity for Nvidia Vera Rubin GPUs. Leases start in the first half of 2027.
Its CFO claims payback periods under one year on its compute investments.
1/ Banks and Blackstone lined up $63B to buy Broadcom AI chips and networking gear and lease them to Anthropic through a new SPV. Broadcom backstops the first-lien part. Deliveries start in 2027.
My read: the chip vendor is now also the credit guarantor.
Anthropic's S-1: over 90% of contributions to its codebase in September were written by Claude. Daily code per engineer is over 45x the 2024 average.
What share of the code you shipped this month did an agent write?