Matthew Berman breaks down a Vercel chart showing open-weights models flipping past closed models in token volume. DeepSeek eclipses Anthropic in token share (25.2% vs 24.5%) but Anthropic still captures 64.6% of spend (23x DeepSeek) — open weights win volume, the frontier wins value. Plus platform risk, the Harvey/Thomson Reuters adoption story, and the China chip dependency risk.
OpenAI cut GPT-5.6 Luna's price by 80% — and the gains came from pointing GPT-5.6 Soul at its own infrastructure. Matthew Berman on the 20% serving-cost drop the model found for itself, the Karpathy auto-research parallel, why the flagship conspicuously didn't get cheaper, and what recursive self-improvement means for everyone trying to catch up.
LocalLayer breaks down Claude Opus 5's launch — same $5/$25 pricing as Opus 4.8 but within 0.5% of Fable 5 on Cursor Bench, with adaptive thinking, migration gotchas, and why all benchmarks are self-reported. The real story: a cost-per-task race at the mid-tier.
Matthew Berman analyzes Moonshot AI's Kimi K3 (2.8T params, biggest open-source model ever) which tops Arena AI's front-end benchmark at 76% vs Fable 5's 63%. Half the price but twice the tokens. Covers industry reactions, Anthropic's distillation accusation, why US labs are still 8-10 months ahead, and why Chinese open source benefits the entire AI stack.
Moonshot releases Kimi K3: 2.8 trillion parameter open-weight model rivaling Fable 5 and GPT-5.6. #1 Frontend Code Arena, #1 writing benchmarks, top 3 overall. $3/$15 per million tokens (half Sol, third of Fable). MoE with 896 experts (16 active), 2.5x scaling efficiency. Native multimodal: motion graphics, video editing, scientific research (2hrs vs 2 weeks). First open-weight frontier model raises export control questions.
A deep dive into GLM 5.2 by Z.ai — a 744B MoE model (MIT licensed, ~40B active per token) with 1M context via sparse attention and Index Share (2.9× fewer ops), multi-token prediction (~20% faster inference), and 74.4% on Frontier SWE — all at roughly 10× cheaper than Claude Max.
While the AI industry doubles down on transformers and autoregression, DeepMind is quietly investing in entirely different paths toward AGI — from Griffin and Titans to diffusion LLMs, world models, and Gemini Omni. Pourya Kordi maps the full strategy and the Hassabis vs LeCun debate at its center.
After the US government banned Fable 5 overnight, Greg Isenberg makes the case for local AI — walking through runtimes (Ollama, LM Studio), model-to-hardware matching, quantization, the agent loop with Hermes, and five startup ideas that only exist because intelligence now sits on your desk.
Pope Leo XIV released a 40,000-word encyclical entirely dedicated to artificial intelligence. Matthew Berman walks through the key arguments — from consciousness and human dignity to the concentration of technological power in private hands.
Nate B Jones argues that Opus 4.7 and OpenAI 5.5 represent a qualitative shift that makes traditional prompting obsolete. He introduces the AI Question Method — a framework for complex agentic knowledge work where users don't know what to ask.
