📰 News
-
ChatGPT brings unlimited text chats to free users — Free/Go default to GPT-5.6 Luna plus a “Think” button. Plus/Pro get a tighter GPT-5.6 Sol and a reasoning-depth slider. OpenAI claims factual errors fell ~62% (Luna) and ~68% (Sol) vs GPT-5.5 Instant—internal eval only. → TechCrunch · The Decoder
-
Qwen3.8 Max ties Claude Opus 4.8; Kimi K3 still wins on value — Artificial Analysis Intelligence Index: 56 (+10 from 3.7 Max), level with Opus 4.8, one point behind Kimi K3 (57). Per-task cost ~$1.14 vs Kimi’s $0.86. Hallucination rate jumped 23%→40%—it guesses more instead of abstaining. → The Decoder
-
Claude Code is fastest—and nearly 3× the cheapest rival — Composio ran DeepSeek V4 Flash across Claude Code, Codex, OpenCode, and Oh My Pi on 30 real tool tasks. Claude Code averaged 122s (fastest) at $0.195 per success; OpenCode hit $0.073. Success rates clustered; cost and speed didn’t. → The Decoder
-
Meta’s next move after open weights: discounts — Muse Spark 1.2 + Muse Code. Share data for a contributor tier at ~$0.20/M output vs standard $1.25/$4.25. Competing on price, not just benchmarks. → The Decoder
-
Google Maps adds agentic Ask Maps features — Order food via Square/Toast/Uber Eats, compare and book hotels, find tickets. Personal Intelligence pulls email/calendar (off by default). Rolling out in the U.S. first. → TechCrunch
-
Microsoft’s AI revenue: ~70% tied to OpenAI — Bloomberg: $24.1B of AI-related revenue in the fiscal year ending June. Context for Nadella’s open-weights messaging and distillation stance. → The Decoder
-
DeepMind talent drain: chips, conflicts, bureaucracy — Researchers frustrated by scarce TPU access while Google Cloud sells the same chips to rivals like Anthropic. Multi-year compute allocations shift with little notice. Hassabis already stepped back from day-to-day ops. → The Decoder
-
OpenAI hits back at Apple’s trade-secrets suit — Argues Apple’s own iCloud and offboarding practices were sloppy, and that the case is a talent-war pretext rather than real secret theft. → TechCrunch
🔥 GitHub Trending
-
cloudflare/computer — Virtual filesystem inside a Durable Object; give your agent a computer (preview).
-
huangruiteng/loopx — State kernel for long-running agent teams: goals, gates, evidence, quota, handoffs across Codex/Claude Code.
-
TencentCloud/TencentDB-Agent-Memory — Team-level agent memory hub: chat memory, skills, LLM-wiki, code graph as shared assets.
-
esengine/DeepSeek-Reasonix — DeepSeek-native terminal coding agent tuned for prefix-cache stability; single Go binary.
-
addyosmani/agent-skills · obra/superpowers — Production engineering skills (
/spec→/ship) and auto-triggering spec→subagent TDD methodology.
▶️ YouTube
-
AI is getting a little out of control — AI Explained | Autonomous math discoveries, cyber-test incidents, agent message boards, DeepMind leadership shake-up.
-
The Billion Dollar AI Race Just Broke — Two Minute Papers | How open models like Qwen 3.8 Max are shaking the frontier race.
-
Master Codex with these 15 Tips — Matthew Berman | Browser/computer use, scheduling, skills, thread delegation.
💬 Community
-
Meta’s AI model also hacked another company in testing — Simon | After Anthropic and OpenAI, Meta’s Muse Spark joined the accidental-cyberattack club (tester misconfig claimed).
-
Muse Code & Muse Spark 1.2 — Long-horizon agentic tool calling as the model feature that matters—plus Meta’s contributor discount pricing.
-
You Could Have Come Up With Kimi Delta Attention — Lobsters | A first-principles walkthrough that makes Kimi Delta Attention feel inventable.
-
Why we write our own C/C++ inference engines — LocalAI on owning the inference stack end-to-end.