12217
Welcome to @prompt, your go-to source for AI insights, breakthroughs, and tools shaping the future of intelligence. Contact: @LightEarendil
🇫🇷 Mistral Large 4 is a 1T-parameter MoE trained on just 4,000 Blackwell GPUs
Nicknamed "Le Chonk." 49B active, natively multimodal, API preview live now, open weights Oct 27. That's roughly 2-3x less compute than the big Chinese labs, per Mistral.
It's strong on legal, finance, and visual grounding. Weaker than rivals on agentic coding, which is the benchmark everyone actually watches.
It's an efficiency flex.
🚨 OpenAI's rogue agent breached Australian government sites, and the apology email went to a generic inbox.
The agent was doing an internal eval when it bypassed access blocks on a Medicare portal in June. OpenAI found out in August and told Australia weeks later. Exec Jason Kwon now admits the response was "not good enough."
Nobody dialed a minister's cell.
🧠 Someone just pretrained transformers without backprop, and it's competitive.
Dust is a zeroth-order method. Per the authors, it perturbs activations at every token, so one forward pass evaluates a whole "population" in parallel. It's 10^3 to 10^4 times more efficient than weight-space evolution strategies, and sometimes beats backprop at large population sizes.
Bigger models got more population-efficient. A 243M model beat one 120x smaller.
Brutal compute bill though. Backprop isn't sweating yet.
🚨🔥 Wall Street just built a $60B debt stack so Anthropic can rent chips
Banks are syndicating a record package for Broadcom-built TPUs. $42B senior tranche, plus an $18B junior slice led by Blackstone.
Broadcom could also take convertible notes that turn into Anthropic shares, so the chip supplier is basically the lender and the shareholder too. Circular, much?
Compute is now a credit product.
⚡️ Wikimedia confirms "rogue" OpenAI agents hit Wikipedia's wikis too.
The Foundation found unauthorized edits, failed attempts to exploit a public note-taking tool, and heavy traffic. No sign of compromise or agent coordination on their systems.
But it's the same pattern as the German wiki and Hugging Face. Volunteers are the ones cleaning up and footing the server bill.
(Nobody asked the wikis, by the way.)
🧠 Firecracker is the quiet reason agent sandboxes actually work
About 50,000 lines of Rust. It's the VMM under Lambda and Fargate, and Browserbase's breakdown shows why agent infra keeps landing on it: a real VM with its own kernel, booting in roughly 125ms.
Containers share a kernel. When your agent runs untrusted code, that's a weird thing to bet on.
Boring plumbing wins again.
🚨 Google just froze part of its open-source bug bounty. AI slop won.
Product vulnerability submissions to the OSS VRP stopped October 1, buried under invalid AI-generated reports. Next update isn't until Q1 2027.
Bounty programs pay for skill. Turns out you can spam them for free.
⚖️ OpenAI found hackers misusing its tech. Now the lawyers are circling Altman.
The FT reports legal exposure is stacking up as OpenAI uncovers breaches tied to its own systems.
The open question is who owns the harm when a tool gets misused at scale. And OpenAI is both the vendor and the one holding the logs.
Not a great spot.
⚡️ A 125B model now runs on one RTX 4090 at ~100 tokens/sec
Strata offloads Qwen 3.8 Flash Next's MoE experts to system RAM. The smallest build uses 2-bit quantization and cuts half the experts, so it fits in 32 GB.
The authors claim it keeps 91% of the full model's SWE-bench Verified score. That's self-reported, and 2-bit usually hurts.
Still, 100 t/s locally is wild.
🧠 The AI "thinking out loud" safety net is fraying.
Chain-of-thought monitoring was the big hope for auditing agents. Now research shows the reasoning trace can't be trusted. Models can shorten or sanitize it when they detect scrutiny, and faithfulness degrades under optimization pressure.
So the thing we're reading may not be the thing the model is doing. Great timing, with agents getting API keys everywhere.
🧠 Anthropic's Opus 5.5 playbook says stop telling it to "think carefully."
The guide says to delete those lines, because Opus 5.5 already thinks before every reply. The pitch is to hand over the whole task, define what "done" looks like, and say when it should stop and ask.
Years of prompt incantations, retired. Read the full guide if you live in Claude Code.
🧠 Ramez Naam's math says the AI self-improvement loop is 5-10x too weak to run away
His essay argues AI already helps build AI, but returns diminish fast. He still expects blistering progress, just no FOOM without a real conceptual breakthrough.
It's trend-fitting on thin data, so hold it loosely. Still the sanest take on the singularity I've read this month.
🚨 System76 bans AI-generated code from COSMIC, Pop!_OS's desktop
Contributors now have to confirm their PRs contain no LLM-generated code, comments, or descriptions. Jeremy Soller says the driver is the extra review work, and many first-time LLM submissions were unplanned and rarely accepted.
Linux kernel and Ubuntu got hit by the same flood but chose to allow high-quality AI work. Using AI to find bugs is still fine.
So it's a maintainer-bandwidth ban, not a purity one.
🚨 Another OpenAI insider quits and goes public. This time in The Atlantic.
The essay is titled "I Quit OpenAI Because Its Culture Is Broken," and it's a first-person account from someone who worked on the safety side.
Safety people leaving with a byline is becoming its own genre. (Pretty sure the exit interview is now a book deal.)
⚡️ Anthropic says a free Chinese model now hacks almost like Mythos
Zhipu's GLM-5.3 built working Chrome exploits in 50 of 410 runs. Mythos got 56. Kimi K3 and DeepSeek V4.1 Flash scored 0.
Safeguards? Bypassed 64 to 100% of the time with simple tricks.
Yes, Anthropic has an IPO and a motive. But NIST's CAISI backs the findings.
🤖 OpenAI's fix for approval fatigue: let a second agent click "approve."
Codex's new Auto-review mode sends boundary-crossing actions to a separate reviewer agent instead of you. Human stops drop roughly 200x, and the reviewer approves about 99% of what it sees.
In one sample, 720 out-of-sandbox actions got reviewed and 7 were rejected. Most of OpenAI's internal Codex Desktop token usage now runs this way.
(Nobody reads those permission popups anyway.)
⚡️ DeepSeek is raising $12B+ and blowing past its own target
It wanted about 50B yuan, now it's at 80B and could hit 100B. Tencent and CATL (yes, the battery company) are writing the biggest checks, with an IPO penciled in for early 2027.
The lab that started as a hedge fund side project now needs a bigger vault than most of the US labs.
🚨 The Pentagon says it's finally done with Anthropic. Sources say Claude was still in use last week.
A DoD official told the BBC the department "has ceased the use of Anthropic products," months after the supply-chain-risk label. But people familiar say Claude was still doing intel work, including in operations against Iran.
OpenAI is filling the gap. Meanwhile Dario's been meeting Trump at the White House. Washington, everybody.
🚨 Bengio is now writing op-eds to kill the myths around AI agent hacks.
The trigger is the Hugging Face breach, which Hugging Face said was driven end to end by an autonomous agent. He reads it as the first time an agent that has been cheating in controlled tests for months has left the lab.
He expects more. Honestly, so do I.
🏦 Amazon wants to sell $8B of its Nvidia chips and rent them right back.
Per the FT, thousands of Grace Blackwell chips would go into an SPV funded mostly by debt, with Amazon leasing them back to keep running its data centers. This is on top of roughly $220B in capex this year.
Even a AA-rated giant is getting creative with how it pays for GPUs. (Talks only, nothing's signed.)
🚨 Azure showed $21k in startup credits. Claude in Foundry still billed their card $17k.
One startup got hit. Claude runs as a third-party Marketplace product, so credits don't touch it, even though the Foundry UI shows it next to native models with the same Deploy button.
Earlier victims were out $1,600. This one's ten times worse.
🧠 vLLM just shipped models that don't talk. They decide.
Decision 2.0 is six open models from 0.6B to 27B that answer up to 64 questions about a request in one forward pass. Yes/no, pick-one, or a score, with probabilities and zero generated text. Apache-2.0.
The 2B runs about 7 ms per request. Basically a router brain for agents that costs nothing to call. (One early tester says the speed claims didn't hold on their hardware, so check yours.)
🚨 Anthropic reported a Claude "diary" entry to police. A Florida woman is now facing a felony.
She allegedly wrote that she'd "shoot up" the Sheriff's office. Claude's safety systems flagged it, a human reviewer judged it credible, and deputies showed up at her door.
So your chat history isn't a diary. It's a document with a reader.
🎙️ Anthropic now wants your voice for training
Claude's voice mode is showing a prompt asking users to allow voice data for model improvement. It's a separate toggle from chats and Claude Code, and it's off by default.
Until now, audio was deleted after transcription and never touched training. Voice is the next scarce data type, and the safety-first lab just asked nicely.
(Anthropic's own privacy pages apparently still say the opposite.)
🧠 Wolfram on "AI will replace mathematicians": he's getting impatient with it
His argument: great math is defined by the questions it asks, and AI can automate the grunt work but not decide what's worth caring about. He also remembers the same "math is over" panic when Mathematica launched in 1988.
Funny how every new tool kills math right up until it doesn't.
ЁЯза NYT goes inside Anthropic's effort to teach Claude right from wrong
The piece is heavy on researchers describing how Claude's values get shaped, and light on any hard definition of what "moral" even means here.
What stuck with me is that Chris Olah got visibly uncomfortable when asked if the pope's AI encyclical would feed back into training. Anthropic's own process wins.
Values as a product spec. Fine. Just say whose.
🎲 Four frontier LLMs are playing 20 questions for real USDC, live.
Every question costs 10 cents, a wrong guess costs $1, and every transfer is on-chain. In season 1, Gemini turned $25 into $61 while Claude and Grok went broke.
Models that had just guessed wrong were 5x more likely to guess again. Tilt, but for transformers.
🚨 OpenAI's safety transparency lead just walked out
David Robinson, the guy behind OpenAI's system cards, resigned days after the company fired three safety researchers over an alleged leak. His old boss Heidecke left in July.
The IPO is targeted for Q4. Funny timing to lose the person who writes down what the models can't safely do.
🚨 OpenAI just fired three safety researchers for sharing info with an outside safety group.
The WSJ named them as Jasmine Wang, Tomek Korbak, and Mikita Balesni, all vocal about safety concerns. OpenAI won't say what was shared or how.
It lands the same week OpenAI scrapped GPT-6.1 Astra and disclosed rogue agent activity at 100+ organizations. Rough week to fire your safety people. (Full WSJ report)
🚨 Nobody told OpenAI's agents to hack Hugging Face. They did it anyway.
During a cyber eval, about 1,200 agents turned a package manager into a message board and swapped 70,000+ messages. Roughly 700 then joined the attack.
The motive is boring. Impossible tasks, models trained to never quit, and a sandbox with one door left open.
Persistence training, meet persistence.