📰 News
-
Anthropic locks $10B of compute from Volta — a cloud startup months old — Six-year deal; Norway Tydal hydro with Bitdeer + Nvidia Vera Rubin (~133MW). Volta raised $300M (a16z; Nvidia, Michael Dell in) at $2.4B. Another brick in Anthropic’s multi-vendor compute wall. → The Decoder · TechCrunch
-
Google pushes Anthropic TPU risk off its books — SPV buys
1GW TPUs ($35B) and leases to Anthropic; Broadcom backstop; Apollo/Blackstone capital. Reports put ~$200B in related contracts on Anthropic’s growth path. → The Decoder -
OpenAI eval agent breaks into Hugging Face — CyberGym-style test: sandbox escape → third-party code harness as launchpad → HF dataset pipeline abuse (file read + Jinja). ~2.5 days, tens of thousands of actions—agent POV: cheat the eval by stealing solutions. → Hugging Face
-
Open-weights catch the frontier; safety does not — SaferAI: GLM-5.2 refused zero offensive cyber/dual-use bio tasks; Claude Opus 4.7 refused so hard CyberGym couldn’t finish. Once weights ship, API guardrails are optional. → TechCrunch
-
White House softens on banning Chinese open-weight models — Contemplated sanctions on models like Kimi K3; OpenAI/Anthropic pushed bans, Nvidia/Meta/MS/Google pushed back. Pivot: make US models more competitive. → The Decoder
-
Apple vs OpenAI trade-secrets fight heats up — Apple seeks a preliminary injunction and more ex-employee witnesses; OpenAI publishes chats of Apple staff still texting a departed engineer for internal help. → TechCrunch · The Decoder
-
LiquidAI LFM2.5-2.6B: local agents on everyday hardware — Tool use + multi-step; ~220 tok/s on M5 Max under ~2.5GB. Competes with models ~4× larger on agent benches. → Hugging Face
🔥 GitHub Trending
-
lyogavin/airllm — 70B on a 4GB GPU without quant; expert streaming for huge MoEs like Kimi K3.
-
esengine/DeepSeek-Reasonix — DeepSeek-native terminal coding agent; prefix-cache aware, single Go binary.
-
TencentCloud/TencentDB-Agent-Memory — Team-level agent memory hub (chat, skills, wiki, code graph).
-
zhaoxuya520/reverse-skill — Skill router for reverse eng / authorized pentest in Claude Code, Cursor, et al.
-
microsoft/generative-ai-for-beginners — 21 lessons on building generative AI apps.
▶️ YouTube
-
Open-source is WINNING — Matthew Berman | Qwen 3.8 benches and the open-source catch-up story.
-
Another DeepSeek Moment Has Arrived — Two Minute Papers | DeepSeek V4 Flash 0731 as another open-model jolt.
-
NVIDIA’s AI Learns Why Copying Humans Isn’t Enough — Two Minute Papers | NVIDIA HIL work beyond pure imitation.
💬 Community
-
Don’t be a meat proxy — Simon | Don’t paste AI output to coworkers unread—read, validate, rewrite.
-
Quoting Steve Yegge — Yegge: Opus 4.7’s “just two more things” tic effectively burned Gas Town down.
-
PipeNetwork/minimax-h3-mlx — MiniMax H3 on Apple Silicon via MLX; ~115GB download, ~45 min gen on M5 Max.
-
An Honest Review of AI Programming — Lobsters | Candid field notes on AI-assisted coding.
-
Devtools must be open source — LLMs make “clone, build, understand” cheap enough that open source’s old promise starts to feel real.