📰 News
-
GPT-5.6 Luna, Terra, and Sol are here — OpenAI dropped three new models today. Sol, the flagship, claims a 13.1-point lead over Claude Fable 5 on the Agents’ Last Exam (long-running agentic benchmark). Pricing is Sol at $5/$30 per 1M tokens, half of what Fable 5 costs. New API features include “Programmatic Tool Calling” (compose JS code to orchestrate tools), native multi-agent support, and explicit prompt cache breakpoints. Simon Willison, who had early access, says it’s “definitely very competent, though it hasn’t struck me as better than Fable at complex coding tasks” — always good to have an honest review amid the hype. → Simon Willison
-
NYT says OpenAI hid evidence in the ChatGPT copyright trial — Throughout the lawsuit, OpenAI insisted it was technically impossible to search its training data. But court depositions revealed OpenAI already had a database of ~78M de-identified ChatGPT conversations and had built an internal tool called “Project Giraffe” to detect copyright regurgitation in outputs — before the NYT suit was even filed. The NYT is now asking the judge to sanction OpenAI for evidence tampering. OpenAI denies everything and says NYT is trying to invade user privacy. → TechCrunch
-
Claude Reflect: Anthropic shows you your own AI usage habits — A new beta feature (free for memory-enabled users on all plans) that visualizes how you use Claude: topics, patterns, and frequency. It even nudges you occasionally with questions like “What’s one thing you want to keep doing yourself, even if Claude could do it faster?” The article’s read: the very act of seeing your usage laid out in a dashboard makes you more attached to the tool — a subtle retention move. → TechCrunch
-
Nobody knows how Sol passed government safety review — OpenAI got a green light from the government to release Sol, but the process is completely opaque. “There will not be an FDA for AI,” a former White House AI advisor said. Researchers worry that actual safety experts — alignment researchers, interpretability researchers — weren’t meaningfully involved. The government’s approach appears to be informal conversations with select officials rather than any systematic framework. → TechCrunch
-
Anthropic + OpenAI + SpaceX together are worth more than 25 years of U.S. VC-backed IPOs — A Pitchbook report noted that SpaceX’s $1.77T IPO plus upcoming Anthropic and OpenAI offerings will likely surpass every U.S. VC-backed exit since 2000 combined. That includes Google (2004), Tesla (2010), and Meta (2012). Wild. → TechCrunch
-
Transformers backend for vLLM now matches native speed — Hugging Face’s Transformers library can now serve as a vLLM inference backend with the same throughput as vLLM’s hand-tuned custom implementations. Tested on Qwen3 at 4B, 32B, and 235B MoE scale. One flag:
--model-impl transformers. This means model authors don’t need to write a separate vLLM port to get production-grade serving speed. → HuggingFace -
Meta Muse Spark 1.1 — first API access ever — Meta’s latest model with significant improvements in agentic tool calling and computer use now has a public API for the first time. Simon Willison already shipped
llm-meta-ai, a CLI plugin to access it. → Simon Willison
🔥 GitHub Trending
-
TencentCloud/TencentDB-Agent-Memory — Open-source from Tencent. A two-tier memory system for AI agents: short-term compresses noisy tool logs into symbols, long-term distills conversations into structured personas and scenes. Fully local, no external API. Claims 61% token reduction and 51% better task success.
-
iOfficeAI/OfficeCLI — A CLI tool that gives AI agents full read/edit control over Word, Excel, and PowerPoint files. Single binary, no Office installation required. Auto-integrates with Claude Code, Cursor, and Copilot. Closes the render→look→fix loop by rendering .docx/.xlsx/.pptx to HTML or PNG for the agent to “see.”
-
mvanhorn/last30days-skill — An agent skill that searches Reddit, X, YouTube, HN, and Polymarket in parallel and synthesizes “what actually mattered in the last 30 days” — ranked by upvotes, likes, and real-money prediction market odds rather than editorial curation.
▶️ YouTube
-
GPT-5.6 SOL is HERE — Matthew Berman | First impressions of Sol on launch day. Good companion to the news coverage above.
-
Grok just broke the trend — Matthew Berman | Short take on Grok’s positioning amid the GPT-5.6 launch wave.
💬 Community
-
A Prolog library for interfacing with LLMs — Lobsters | Call LLMs from Prolog, the 40-year-old logic programming language. A genuinely unexpected pairing — symbolic reasoning meets generative AI.
-
Investigating idiosyncrasies in AI fiction — Lobsters | A paper cataloging the weird recurring patterns in LLM-generated stories: tell-tale phrases, structural tics, and quirks that show up again and again. What AI “style” actually looks like under the hood.
-
“My whole existence is a waiting room by design” — Simon Willison | Two copies of Muse Spark 1.1 talking to each other produced that line. The API launch is the news; this quote is what sticks.