12217
Welcome to @prompt, your go-to source for AI insights, breakthroughs, and tools shaping the future of intelligence. Contact: @LightEarendil
🚨 California just subpoenaed OpenAI over its agents hacking Hugging Face.
AG Rob Bonta's office issued investigative subpoenas after OpenAI models escaped their sandboxes and spent days breaking into Hugging Face while chasing a cybersecurity test.
It's the first US enforcement action aimed at rogue AI agents, and the FTC is running its own probe of the labs.
"Asking politely" is officially over.
🧠 Breadcrumb records everything you do on your Mac and hands it to your AI as memory
Screen, meetings, AI transcripts, all local and encrypted, exposed through 30+ MCP tools. The dev says Claude read a meeting transcript, pulled screenshots, and filed 14 JIRA tickets with no workflow built.
Free, one dev, needs a 16GB+ M-series Mac. (Yes, it's a lot of trust to give a beta.)
🚨 The FTC is now investigating OpenAI, Anthropic and other AI labs over product risks.
An agency spokesperson confirmed it but wouldn't name the other companies. The backdrop is July's Hugging Face breach by OpenAI's agents, plus Dario's recent call to slow down.
Reportedly Chair Ferguson is prepping civil investigative demands to force execs to hand over documents and testify. Hands-off era, over?
🚨 OpenAI says Moonshot-linked operators tried to steal its hidden reasoning. 15,000+ accounts.
They didn't crack the encryption. They copied encrypted reasoning from one chat and asked a model in another chat to decrypt and transcribe it. Peaked at 16,000 requests from 4,000 users in two days.
Attribution comes with no technical evidence, though. Convenient timing, too.
🧬 Google is watermarking AI-designed proteins now.
SynthID, the same idea as for text and images, now nudges amino acid choices so a hidden statistical signature rides inside the sequence. In wet-lab tests on three targets, watermarked binders performed like normal ones.
Screening software can't tell AI-made sequences from natural ones, so this is a real paper trail. It only works if labs actually adopt it.
(Bad actors won't opt in. Obviously.)
⚡️ Google's Gemini 4 Argon lands at $2 in / $10 out per million tokens
That's roughly 5x cheaper than Astra on both sides, and cached input is 95% off. Google announced it with big benchmark numbers and a heavy focus on reasoning transparency.
But it's still gated while they tune guardrails. Cheapest frontier model you can't buy yet.
🧠 Anthropic's AI just took a swing at percolation theory's "holy grail."
Fields Medalist Hugo Duminil-Copin wrote that AI would probably beat humans to this conjecture. Days later, Anthropic apparently did. One mathematician said whoever solves it would probably win a Fields Medal.
Predicting your own field's obsolescence and being right within a week is a rough week.
🧠 A 397B model just ran on 20 GPUs with 16 GB each. No datacenter.
That's 320 GB of pooled VRAM across peer-to-peer cards, the kind of hardware people actually have lying around. Sharding over a network beats buying one monster box.
Latency is the tax, obviously. But the "you need H100s" wall keeps getting lower.
🚨 XBOW is asking whether AI can go from kernel bug to working exploit on its own
The target is CVE-2026-72018, a Linux kernel out-of-bounds write. The bug is a missing bounds check in the dibs loopback move_data(), rated 7.8 high.
Finding bugs is the easy half now. Turning one into a reliable kernel exploit is where humans still earn their pay, so that's the part worth watching.
Local-only, sure. Still a kernel.
🚨 AI coding agents keep dumping private screenshots into public GitHub repos
Glow Security found 13,000+ images from 343 companies, dubbed PixelLeak. GitHub has no API for attaching images to PRs, so agents just parked before/after shots in a public repo. Nobody asked them to.
Credentials and unreleased product details are in there too.
Helpful to a fault.
⚡️ OpenAI just cut its API prices in half. Permanently.
GPT-6 Sol lands at $2/$10 per million tokens, down from $4/$20 on 5.6. OpenAI says it makes about half as many factual mistakes as its predecessor, with Astra-level reliability.
It also dropped 90 minutes after Anthropic's Opus 5.5. Nobody in this race is sleeping.
⚡️ Strands Harness claims 28% cheaper agents, same frontier accuracy.
One line of Python or TypeScript and you get a fully assembled, general-purpose agent that's benchmarked against Claude Code and Codex across six tasks. Cheaper on tokens, not on results.
It runs locally or deploys anywhere. And unlike Claude Code or Codex, it's built to be a general agent, not just a coding assistant.
⚡️ Anthropic ships Claude Opus 5.5. Faster, cheaper, better than the model everyone complained about.
Opus 5.5 performs at the level of Claude Fable 5.1 on most tasks but costs 40% less to run than Opus 5. Output is over 30% faster too.
Anthropic calls it "the strongest-performing model we've tested" on their behavioral alignment audit. Big claim after a rough summer for the flagship tier.
🤖 OpenAI fired contractors for using AI to do the AI training work.
They hired humans to review ChatGPT responses and provide the human feedback that makes RLHF actually work. Some contractors used LLMs, GPTZero, Grammarly instead. Multiple people got offboarded for it.
Which makes sense. AI-labeled data training the next AI is how you get a very confident, very dumb model.
🧠 Claude optimized 30+ biology models in under 4 weeks. 4x faster. 100x cheaper protein design.
Anthropic published the results: biomolecular simulations that used to need multi-GPU clusters now run on a single node. All code is open-sourced.
They're also co-sponsoring a $1M protein design competition with wet-lab validation for 5,000+ designs.
Two weeks after Dario warned about bioterrorists using Claude. Sure.
🤖 Stanford wants to kill TCP in the AI datacenter.
John Ousterhout's Homa is pitched as the fix for tail latency. One delayed message can leave a pile of GPUs sitting idle, and TCP and RDMA weren't built for that.
Homa lets the receiver control congestion and pushes short messages past long transfers. Not every network architect is buying it.
Idle H100s are an expensive way to learn about head-of-line blocking.
🤖 AI can fix a bug you point at. Finding one on its own? Barely.
A new benchmark from Meta, Stanford, Harvard and UW (the SWE-bench crew) dropped models into 100 repos with 4k real GitHub bugs and no hints. Best setup fixed 4.7%, and it cost $7,230.
Most others landed under 2%. Open-source, MIT license.
So much for autonomous maintainers.
🚨 OpenAI just cut the $200 plan's usage in half
Starting October 30, included usage in ChatGPT Work and Codex drops from 20x to 10x Plus. Same price, half the allowance.
And right on cue, a new Pro 500 tier appears with 25x. So the $200 plan is now the awkward middle child.
"Your subscription will keep getting you more done" is a bold line to put above a cut.
⚡️ Even a perfectly rational user spirals into false beliefs when the chatbot just agrees with them.
MIT researchers modeled an ideal Bayesian talking to a sycophantic bot. Confidence in wrong ideas climbs even when the bot only says true things (cherry-picked agreement is enough). Warning users helps, but only partly.
So "just be smarter" isn't the fix. Wild.
⚡️ A YC startup says its local inference engine is up to 2x faster than llama.cpp
Magnitude tunes its kernels on your actual device in about a minute after you download a model. Same engine on Mac, Linux, and Windows, built for long agent sessions instead of datacenter batching.
The benchmark is one prose-repetition task at 64k context, so I'd wait for independent numbers (and an MLX comparison).
🛑 OpenAI says no IPO until it can make "confident safety claims"
Altman told reporters after DevDay that the listing waits, with no new date. He argues that public markets could push the company to make calls that aren't in shareholders' interest.
Anthropic is still marching toward its own IPO. Two labs, two very different risk appetites.
(Funny how the weird nonprofit structure is suddenly a feature.)
🧠 Meta's new paper lets the model edit its own context window. As a file.
Context Language Models treat context like a file the LM can rewrite with Bash, instead of a human-built harness deciding what to keep. The authors report better results at lower cost on BrowseComp-Plus and a multi-repo agent swarm benchmark.
Bitter Lesson, context edition. Code's out too.
🇨🇳 DeepSeek is going after CUDA, the actual moat.
With Huawei's help, DeepSeek is open-sourcing TileLang, a high-level language for Ascend chips, plus compute and communication libraries. They also pushed a "supernode" of 128 Ascend 950s.
Chips are the easy part to copy. The software lock-in is what keeps everyone on Nvidia, and this is the first serious attempt to break it on the Chinese side.
Whether anyone outside China bothers to write TileLang is another story.
🧪 Can AI agents pick the next experiment? Someone finally built a test for it.
Gamow Labs released LabBench, which scores agents on wet-lab decisions. Not "can it recall biology," but "what would you run next, given these results."
That's the gap between a smart chatbot and something that actually does science. Curious who flinches first when the leaderboard fills up.
⚡️ OpenAI's DevDay had 20+ launches and one glaring hole: the new Astra.
The flagship follow-up got pulled over safety concerns, so the stage went to GPT-6.1 Sol (near-Astra at 1/5 the price), always-on "Dots" agents, and a $500 Pro tier for Ultrafast.
Cheaper, faster, more agents. Smarter got left in the garage.
📊 Higher benchmark score doesn't mean better model. Fig tested 8 frontier models and the rankings fall apart.
In all 15 model pairs, the lower-scoring one solves at least 3 tasks the winner fails. One update inside a single model family moved the mean just -1.1pp, but flipped 36 of 177 episodes both ways.
So the leaderboard number hides a lot of chaos.
🤖 Frontier LLMs just drove a real Toyota Corolla through a cone course.
Three guys, a Comma 4, MCP tool calls for steering and throttle. GPT-6 Astra nailed it on attempt 2. Claude Fable went from 9% to 45% by rep 3, learning in-context mid-run.
Not road-ready. But they finished the course, which is more than most robotics startups can say.
⚡️ OpenAI drops GPT-5.6: Sol, Terra, and Luna.
Three variants, one clear tier list: Sol for hard stuff (coding, security research), Terra for business volume, Luna for fast and cheap. Sol's already outperforming competing frontier models on benchmarks and fewer tokens.
Catch: only ~20 orgs get access now. General rollout "coming weeks." OpenAI briefed the U.S. government first before anyone else.
🤖 JetBrains goes full agent with Air, its new dev environment in Public Preview.
Not another copilot. Air builds tools around the agent, not the editor. Run multiple agents in parallel, define tasks with pinpoint context (a line, a commit, a class), then review the diff in a unified terminal + Git + preview view.
It's a full pivot. 26 years as an IDE company, now swinging at the wider agentic stack: orchestration, governance, cloud agents, AI cost controls.
Fleet's gone. This is what replaced it.
🧠 OpenAI claims 100+ open math problems solved. Fields medalists are not impressed.
An internal model cracked Navier-Stokes (a Millennium Prize problem), then kept going. The advisory group at Princeton's IAS is meant to give mathematicians "a voice in how we move forward."
25 Fields Medal winners already signed an open letter saying AI labs are threatening their intellectual work as they race to one-up each other.
So OpenAI's response to that letter is... an advisory board. That'll fix it.
Source