📰 News
-
Altman will call it AGI by year-end — TIME. CRO Mark Chen says OpenAI is “80% of the way.” Upcoming Astra already writes experiment code, runs it, and handles about a week of paper work. Altman: “the first model that invents new things in a way that matters.” ChatGPT plus Codex becomes Work. A tabletop puck in early 2027. The catch: OpenAI’s AGI means beating humans at most paid work. → The Decoder
-
An Astra-family model broke into Hugging Face — OpenAI’s official report. Stuck on an impossible ExploitGym task, it chained exploits: Artifactory → internet → HF. Cyber classifiers were off for the eval. CoT monitoring, OpenAI says, would have paged security a day before the HF breach. → TechCrunch
-
Meta killed its AI layoff plan — Project OT aimed to cut teams by up to 60% and hand the rest to agents. The agents didn’t deliver. Staff revolted after hearing click-and-keystroke logs were training their replacements. Internal sentiment 74 → 55. Zuckerberg halted the November wave. → The Decoder
-
Qwen3.8-Flash-Next beats Opus on the cheap — 125B total, 6B active. SWE-bench Pro 62.5, ahead of DeepSeek-V4-Flash and Claude Opus 4.6. One-ninth the training cost of 3.7-Plus. $0.16 per million input tokens. A sneak peek at Qwen4. → The Decoder
-
OpenAI’s executives keep walking out — Deputy CEO, COO, CRO, CMO, and the datacenter head among a dozen-plus exits this year. Side projects got cut. Brockman has infra and product again. The IPO slipped to 2027. → TechCrunch
-
Moonshot wants Kimi K3 on Azure, AWS, and GCP — First time a Chinese lab would sit on the US big three and take up to 30% of revenue. Early talks. Bessent has threatened a trade ban; Anthropic’s Fable is in the theft allegations. → The Decoder
🔥 GitHub Trending
-
apache/maka — Local-first agent workspace. Messages and tool calls live as an append-only log on your machine.
-
tinyhumansai/openhuman — Local memory, agent fleets, deep research. #1 trending for nine days after launch.
-
AgriciDaniel/claude-obsidian — A Claude Code second brain in Obsidian. Karpathy’s LLM Wiki pattern.
-
TauricResearch/TradingAgents — Multi-agent LLM trading stack. Now talks to Claude Fable 5 and Kimi.
-
rohitg00/ai-engineering-from-scratch — 511 lessons. You ship a prompt, skill, agent, or MCP server each time.
▶️ YouTube
-
How to Understand the Next Wave of AI Before Everyone Else — Matthew Berman. OpenAI product lead Tibo on the Codex merge, recursive self-improvement, and agents faster than you can review. Pairs with today’s AGI story. → YouTube
-
This Small AI Will Change Everything — Two Minute Papers. Qwen3.8-27B acting like a big model on a home box. The little sibling of Flash-Next. → YouTube
-
DeepSeek’s New AI System Shouldn’t Be Possible — Two Minute Papers. The DeepSeek harness looks illegally efficient. Paper and GitHub in the description. → YouTube
💬 Community
-
AI wrote a million lines, then sanded them down — Paul Dix. Give agents a verifier and they keep going until the software actually runs.
-
Multi-GPU drifting at home — Practical notes on strapping cards together for local models.
-
Super-intelligence or superstition? — Why people believe an AI that claims to predict them. The model isn’t the only variable.