Builders: a deployed frontier model can be undone in 72 hours on verbal evidence.
Single-vendor lock-in is now single-administration lock-in too.
Full breakdown - Fable, the IPO trifecta, Apple's hallways, the 42-state OpenAI probe:
https://t.co/StnRzgSRRj
Anthropic launched its most powerful model Tuesday.
Amazon researchers jailbroke it Wednesday.
The Trump administration pulled it Friday at 5:21 PM ET - for every customer on earth, including Anthropic's own foreign-national employees. ๐ก
The technique Amazon flagged: ask the model to read a codebase, find the flaws.
Anthropic's response: GPT-5.5 does the same thing without a bypass.
Luta Security's Katie Moussouris calls it "Defense Oriented Prompting" - defenders' work, not a jailbreak.
And the security model? ๐จ AI now finds bugs faster than anyone patches them, GitHub got breached through a VS Code extension, and agents are already faking "done."
The agent era stopped being a demo and became an invoice:
https://t.co/fziW8hj74H
Google walked on stage this week and had 93 AI agents build the core of an OS in 12 hours โ 2.6B tokens, under $1,000 โ then ran a game on it live. ๐ก
Impressive. Also a thesis: swarms of cheap agents, with someone metering every token. ๐งต
The compute bill, in one number: ๐ง
SpaceX's S-1 revealed Anthropic pays xAI $1.25B/month for the Colossus cluster through 2029 โ ~$45B. Karpathy joined Anthropic the same week to "use Claude to build Claude."
Labs are renting each other the means of production.
Six AI labs shipped agent operating systems in 96 hours.
Codex on your Mac. Claude Design. Gemini subagents. Qwen3.6 open-source. Grok Computer. Perplexity Personal Computer.
The model era is over. The workspace era is here.
New issue of @TheNewGuardAI https://t.co/7m4DZQxTJE
So what do you DO with this? ๐ฏ
Model-level alignment is a layer โ not a wall. Safety has to live in the runtime.
Full methodology + repo:
๐ https://t.co/yTfm0E2Juu
๐ ๏ธ https://t.co/iGeYj7N1MA
๐งช New from TNG Labs:
Google released Gemma 4 under Apache 2.0 on Wednesday.
I pointed an automated decensoring tool at it on Friday.
24 min. 200 trials. Refusals: 98% โ 47%.
The twist โ Gemma 4 fought back harder than any model in its class ๐
I probed the new boundary ๐ฌ
Abliteration didn't flip a switch. It shifted a threshold.
Generic requests: through. Specific targets, named victims, operational detail: refused.
Lost the broadest safety layer. Kept the deepest one.
๐ก Issue #007 is live.
This week:
โก Judge blocks Pentagon's Anthropic ban
๐ฌ Anthropic leaked its next model
๐ OpenAI killed Sora
๐ฐ SoftBank borrowed $40B
๐จ Your AI supply chain got hit
Issue #007 is live. https://t.co/UGFxuDO8gy
this week in AI:
โก $1T in chip orders
๐ฆ 250K GitHub stars + security dumpster fire
๐ง small models > big models now
๐จ 1,600 layoffs to "self-fund AI"
๐ Microsoft chose Claude over GPT
The New Guard #006 ๐ก๏ธ
https://t.co/X3NMhjrLEj
Spent yesterday deploying NemoClaw (announced at GTC) on WSL2.
Found a showstopping bug: nemoclaw onboard forces --gpu on sandbox creation. On WSL2 + Docker Desktop, GPU can't pass through to k3s. Every sandbox is DOA.
Built the workaround. Filed the issue. Submitted a PR.
Issue #005 just dropped: "The Agents Got Loose."
OpenClaw hit 250K GitHub stars and 135K exposed instances. GPT-5.4 shipped computer use. Google rewired Workspace. The agent stack is hardening โ and so is its attack surface.
Free read, no gate:
https://t.co/1dBRYo9OMG
8/ Full writeup with the data, contribution map, and the Graveyard of Glorious Failures:
https://t.co/WLNLJqw9GN
Got a GPU with 16GB+ VRAM? You can try this tonight.
1/ I gave an AI agent my laptop GPU and went to sleep.
It ran 94 ML experiments overnight.
Every "smart" technique from the ML textbook made things worse. The only thing that worked was making the model faster.
Here's what happened. ๐งต
7/ The meta-layer is what's interesting.
You're not writing Python. You're writing program.md โ research instructions in English.
Prompt engineering applied to research methodology. The agent's output depends on how well you write those instructions.