📰 News
-
Hugging Face hit by an autonomous AI agent — A malicious dataset abused code-execution paths in the data-processing pipeline, then pivoted for internal data and service credentials. ~17k attacker actions were forensically sorted with HF’s own AI in hours, not days. Public models, Spaces, and the supply chain look clean; partner/customer impact still under review. Agentic attack left the forecast deck and entered the incident log. → The Decoder · HF disclosure
-
Google’s Frozen v2 hard-codes Gemini’s architecture — Not the weights—the blueprint. Goal: ~6–10× more tokens per watt than current TPUs, deploy around 2028, mostly for internal inference margin (not a broad external SKU). Stock ticked ~3% on the leak. → The Decoder · TechCrunch
-
Moonshot freezes new Kimi K3 subs after a 48-hour GPU crush — Existing users stay up; plans split into general vs Code memberships to spread load. Open weights didn’t shrink compute demand—they exploded it. → The Decoder
-
A “slow-motion ban” on Chinese models — Sanctions lists, security warnings, liability pressure on hosts—less a hard ban than durable FUD so regulated buyers walk away. Kimi K3 and White House personnel shifts reportedly reopened the throttle. → The Decoder
-
OpenAI is scared of open-weight economics — Strategist Dean Ball floated regulatory FUD, then walked parts of it back. Cheap near-frontier Chinese models squeeze closed-lab margins; critics say open software still accelerates the field. The live question: how far policy follows the labs. → TechCrunch
-
Nvidia’s moat thins: Microsoft leans AMD, Anthropic tests too — Azure expands on AMD’s Helios (2H 2026). SemiAnalysis-style reporting puts Anthropic near top AMD priority if software quality clears. → The Decoder
-
Neill Blomkamp ships all-AI short Nightborne — 13-minute sci-fi horror on Seedance 2.0; 32 licensed real faces/voices, human concept art. Full feature and Barley Studios next. → The Decoder
-
NVIDIA Cosmos 3 Edge — 4B open world model for edge robots/vision: real-time scene understanding and action generation on Jetson-class hardware. → HF
🔥 GitHub Trending
-
github/copilot-sdk — Drop Copilot’s agent runtime into your app (Python, TS, Go, .NET, Java, Rust)—planning, tools, file edits without rolling your own orchestrator.
-
kvcache-ai/ktransformers — Heterogeneous CPU–GPU inference and SFT; fast Day0 coverage for models like GLM-5.2 and MiniMax-M3.
-
jamiepine/voicebox — Local-first voice studio: clone, TTS, dictation, agent voices. Privacy-first ElevenLabs + Wispr alternative.
-
KnockOutEZ/wigolo — Local web search/fetch/crawl/research over MCP for coding agents—no API keys, no metered cloud bill.
-
tirth8205/code-review-graph — Tree-sitter code graph + MCP so reviewers only load the slices that matter.
▶️ YouTube
- Anthropic messed up — Matthew Berman | Short take on Anthropic’s latest misstep/controversy—snack-size, not a deep dive.
💬 Community
-
Who’s Afraid of Chinese Models? — Ben Thompson: make training-data collection fair use and ban anti-distillation ToS so US open models can compete. Links Qwen 3.8’s open-weight pivot to Xi’s “seize the open-source moment” line.
-
Quoting Sam Altman (2022) — Board email from Musk v. Altman: ship a local GPT-3-class model first so rivals struggle to raise money. Open strategy as deterrence, not charity.
-
Reverse-engineering is cheap now — Coding agents flip the ROI on home-device hacks; rewrite-and-throw-away stops feeling expensive.
-
Verifiable AI inference — Can you prove the model actually produced that answer?
-
Inventing ELIZA — Open-access history of the first chatbot—useful origin story for today’s agent debate.