TheNewGuard @TheNewGuardAI

Agentic AF

And the security model? 🚨 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: thenewguard.ai/issues/015-the…

The compute bill, in one number: 🧠 SpaceX's S-1 revealed Anthropic pays xAI $1.25B/month for the Colossus cluster through 2029 — ~$45B. Karpathy joined Anthropic the same week to "use Claude to build Claude." Labs are renting each other the means of production.

Google walked on stage this week and had 93 AI agents build the core of an OS in 12 hours — 2.6B tokens, under $1,000 — then ran a game on it live. 📡 Impressive. Also a thesis: swarms of cheap agents, with someone metering every token. 🧵

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 thenewguard.ai/issues/010-age…

So what do you DO with this? 🎯 Model-level alignment is a layer — not a wall. Safety has to live in the runtime. Full methodology + repo: 🔗 thenewguard.ai/features/ablit… 🛠️ github.com/thenewguardai/…

I probed the new boundary 🔬 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.

🧪 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% → 47%. The twist — Gemma 4 fought back harder than any model in its class 👇

📡 Issue #007 is live. This week: ⚡ Judge blocks Pentagon's Anthropic ban 🔬 Anthropic leaked its next model 👀 OpenAI killed Sora 💰 SoftBank borrowed $40B 🚨 Your AI supply chain got hit Issue #007 is live. thenewguard.ai/issues/007-ant…

this week in AI: ⚡ $1T in chip orders 🦞 250K GitHub stars + security dumpster fire 🧠 small models > big models now 🚨 1,600 layoffs to "self-fund AI" 👀 Microsoft chose Claude over GPT The New Guard #006 🛡️ thenewguard.ai/issues/006-cha…

The full writeup + hands-on lab + automated deploy scripts (WSL2, macOS, Linux): thenewguard.ai/features/tng-n…

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 — and so is its attack surface. Free read, no gate: thenewguard.ai/issues/005-the…

8/ Full writeup with the data, contribution map, and the Graveyard of Glorious Failures: thenewguard.ai/features/karpa… Got a GPU with 16GB+ VRAM? You can try this tonight.

7/ The meta-layer is what's interesting. You're not writing Python. You're writing program.md — research instructions in English. Prompt engineering applied to research methodology. The agent's output depends on how well you write those instructions.

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. 🧵