Social Media Briefing — June 7, 2026

Mollick observes that Gemini Pro models are iterating far slower than Claude or GPT, with the last release in February, creating a widening performance gap that Gemini 3.5 Flash does not close.

jerryjliu0 (LlamaIndex) arguing no frontier lab owns the entire cost/latency/accuracy pareto frontier, driving interest in model routing and cost optimization, relevant to document OCR for AI agents.

Continuing Gary Marcus's commentary from yesterday's Social, Marcus calls a reported government meeting a seismic shift, claiming partial US government ownership of American AI firms would erode global trust akin to distrust of Huawei, benefiting Mistral.

Burkov highlights a Google paper, LEAP, an agentic framework that boosts general LLMs to state-of-the-art formal theorem proving, including autonomously formalizing a subproblem in Knuth's Hamiltonian decomposition.

A counterpoint to Anthropic's claims we covered in Social, natolambert maintaining that despite a recent Anthropic post, serious bottlenecks (organizational, compute, data access) mean AI progress will see linear gains for years, not explosive recursion.

Reports OpenClaw hit 3,000 commits in a day via 60-70 AI agents, with a Chief Architect explaining how the agentic dev factory works and the skill of detecting when agents are bluffing.

Ethan Mollick argues that better AI writing matters because so much software involves text, criticizing repetitive AI cliche phrasing like Claudisms and ChatGPT-isms.

LeCun mocks exponential-pilled enthusiasts for finally realizing real-world processes have irreducible time constants and cannot run faster than real time.

Burkov argues LLMs fundamentally cannot output calibrated numerical confidence estimates matching actual correctness probability, illustrating limits of our understanding of LLMs.