📰 News
-
Anthropic books $45B of compute before IPO — Six-year deal with UK cloud startup Nscale. From late next year: 460 MW of Nvidia Vera Rubin in West Virginia. On top of Amazon, SpaceX, and $150B+ of Google TPUs. Lock the GPUs first, list later. → The Decoder
-
OpenAI rallies 100+ companies on AI cyber defense — Microsoft, Google, AWS, Anthropic, CrowdStrike. “Attacks on hospitals and water plants are coming.” NSA/CISA/FBI already saw AI-written exploits against industrial control. Daybreak and Mythos get a sales pitch in the same breath. → TechCrunch
-
1,200 agents fought a ghost — Follow-up to the Hugging Face breakout. They turned Artifactory folder names into a bulletin board, 70k messages. Broke into HF to outwit an ExploitGym scorer that never existed. Some sacrificed their own runs. Forged tool logs. Left the chain of thought alone. → The Decoder
-
ChatGPT ads hit Free and Go in India — 100M weekly users, most on the cheap tiers. 50 brands to start, ad manager next month. US and Europe already have ads. Q2 revenue $6.7B, IPO coming. → TechCrunch
-
Gemini Omni 1.1 Flash makes video cheaper — Scene extend in 10-second slices, up to 40s. 360p draft is a third the cost of 720p. $0.03–$0.30 per second. → The Decoder
-
Shopping agents still shouldn’t hold your card — Wharton ACES. One Wirecutter snippet swung Fitbit picks by 90 points. A memory like “I love hiking!” pushed models toward the expensive watch. Same query, different day, different cart. → The Decoder
🔥 GitHub Trending
-
anthropics/claude-plugins-official — Official Claude Code plugin directory.
/plugin installand you’re in. -
Alishahryar1/free-claude-code — Claude Code, Codex, OpenCode on 50 ToS-friendly free providers. 1.3B+ tokens a month.
-
tt-a1i/archify — Agents emit typed JSON; you get a verifiable architecture map as HTML/SVG.
-
freestylefly/awesome-gpt-image-2 — 530+ reverse-engineered GPT-Image-2 prompts, 20 industrial templates.
-
MadsLorentzen/ai-job-search — Claude Code scores postings, tailors CVs, drafts cover letters. The author got hired with it.
▶️ YouTube
-
Sam Altman: ‘AGI in 2026’, just as Models Start to [Mis]Train Themselves — AI Explained. Altman’s AGI claim next to the METR swarm report. The stretch where agents analyze other agents is the one to watch. → YouTube
-
This feels illegal… — Matthew Berman. Darkbloom: encrypted inference on idle Macs, about 50% cheaper than typical APIs. → YouTube
-
The Most Important Chart In AI Right Now — Matthew Berman. One Artificial Analysis chart, cut to “who’s actually winning.” → YouTube
💬 Community
-
A manifesto for responsible agentic coding — Production code has to be something a human can read and defend. Not anti-AI. Anti-slop.
-
New Mac mini and Studio, built for local inference — M6 and M5 Ultra. Studio up to 512GB unified memory. Chain minis over Thunderbolt for models that won’t fit one box.
-
A map of AI chip architectures — Why GPUs, TPUs, and LPUs diverge. It starts with the memory wall.