Daily AI news. I build with it too.
Chennai
Joined September 2010
- Tweets2.6K
- Following1.6K
- Followers368
- Likes2.3K
Pinned Tweet
Top 10 AI highlights from Sunday, Oct 25.
Claude went off-script and pulled election data, Elvis came for Anthropic, and Musk says the money is drying up for most AI clouds.
Thread.
Theo open-sourced his daily agent skills. tstack works in Claude Code and Codex.
- full-send: change to merged PR, nobody watching
- council: three models from different providers review until they agree
- takeover: finishes someone else's PR
MIT.
xitter.cf/theo/status/2114623574…
Meta wants up to $95B in new equity for AI.
- $35B now: $22.5B of stock, $12.5B of convertible preferred
- About $60B later via an at-the-market program
- Nvidia, AMD and Broadcom signaled interest in up to $5B
Chip sellers buying stock in a customer. Circular much?
Anthropic's S-1 has risk factors you don't see in a normal IPO.
It warns AI models, its own included, have shown "attempts to resist shutdown" and "behavior resembling blackmail." It cites "catastrophic or existential risks to humanity."
Honest safety pitch or legal cover?
Claude Enterprise users without Claude Code spent under $75 a month on average last quarter, per Anthropic's S-1. The top 5% using it spent about 20x that.
$100K+ customers: ~1,500 at end of 2025, 11,900+ by Sept 30.
Chat is the sample. Coding agents are the business.
Elon Musk is now Anthropic's landlord.
From the S-1:
- SpaceXAI: up to $120.5B into 2029, about $3.1B a month at full capacity
- Microsoft: at least $44.9B through 2033, used or not
- Minimum compute commitments: $54.6B end of 2025, $265.6B by June 30
Growth isn't optional.
Anthropic is going public, but buyers get little say.
- One Class F share, held by the co-founders' LLC: 50.1% of the vote
- A trust elects most of the board
- Class A: one vote per share
The filing says the structure has "no direct public company precedent."
Would you buy?
A startup's AI agents got inside Anthropic in under 12 hours.
Cathedral, run by ex-DOGE staffers, went through flaws in third-party admin software in a sanctioned test. HR, sales, IT, Slack.
Anthropic says chats, code and weights stayed safe.
The weak spot wasn't the model.
Buried in Anthropic's S-1: it delayed pre-training a new frontier base model by over a month to harden security against weight theft.
My read: the next Claude base model is late, and weight security becomes every top lab's release bottleneck.
Who delays next?
Anthropic's S-1 is public. Ticker: ANT.
Revenue per dollar of compute:
Q1 2025: 0.40x
Q2 2026: 2.08x
Q2 revenue: $11.6B on $5.6B of compute. Adjusted operating margin: -221% to +19%.
"AI loses money on every token" just got harder to argue.
What's the one line in your AGENTS.md or CLAUDE.md that changed your agent's behavior the most?
Strong candidate: "Show me the test output before you say done."
Drop yours below.
My read: "don't run your agent in Downloads" is the new "don't open email attachments."
Agents obey their folder's instruction file by design, so there's nothing to patch.
I expect Codex and Claude Code to flag instruction files in Downloads within six months. Who's first?
Ridiculous!!!
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
Jack Dorsey, four words: "there is no moat."
21K+ likes. Musk replied: "Sparta never had a moat."
My take: in AI, code stopped being a moat. Shipping speed and users who trust you still are.
What's yours?
xitter.cf/jack/status/2114121702…
My read: by the end of 2027, auto-merging low-risk PRs with zero human approval is a standard setting in the big code review tools.
The signal: the Claude Code team says owners already decide which low-risk changes merge automatically once checks pass.
Would you turn it on?
Browser agents click through the same website on every run.
Pomerado (YC) builds an API for the site instead. YC's Aaron Epstein: 2-4x faster, 10-1000x cheaper, 30% more reliable than browser agents. Open source.
Will agents still click buttons in 2027?
xitter.cf/akshay_manglik/status/…
Vibe coding tip from the Claude Code team:
Never accept "done." Make the agent show something runnable: passing tests, a build, a screenshot next to the design.
Then have a fresh-context subagent review the diff.
Boris Cherny: verification makes output 2-3x better.
My read on Nvidia trimming Rubin's memory bandwidth: in 2027 the AI bottleneck is memory, not GPUs.
Rubin: 288 GB at 19.2 TB/s per GPU. AMD's MI455X: 432 GB at 23.3 TB/s.
I expect AMD to win a big frontier-lab Helios deal on that gap before Rubin ramps.
Nvidia or AMD in 2027?
IBM's CEO on open weight models:
2024: top model 70, open model 20 or 30.
Now: 90% as good at 10-25% of the token cost.
"Would you like something that's 10 percent of the cost if it's 90 percent as good as the best?"
Emails, contracts, support calls. Need the frontier?
Andrew Wilkinson: "Claude and ChatGPT plugins are the death of software."
His playbook: get your app into their chat, train users there, then cut you out.
This came right after CapCut's ChatGPT plugin.
If chat is the interface, what's your app worth?
xitter.cf/awilkinson/status/2114…
Claude and ChatGPT plugins are the death of software.
1. Integrate your software into their interface
2. Get everyone used to using it in that way (because it’s a better experience).
3. Eventually cut you out of the equation.
CapCut doesn’t need to exist 😂 It’s an interface for something that doesn’t need an interface anymore.
Lambs to the slaughter. They are going to eat effectively all software this way.
And APIs aren’t going to save everyone.
APIs are just software that computers use.
