Social Media Briefing — July 4, 2026

LeCun argues that concentration of power and the desire for control is the biggest danger in AI, likening it to historical suppression of the printing press, and recounts how the internet was deliberately opened to the public against telecom interests.

HuggingFace co-founder shares and endorses an essay arguing that open science and distributed participation are essential, warning against concentration of AI research capability inside a few secretive labs and calling for a research commons across academia, industry, and public interest.

Willison shares a Fable tip: instructing the agent to use its own judgement to pick an appropriate lower-power model and run it in a subagent for coding tasks, which appears to reduce token consumption.

Mollick shares his prompt asking Claude Fable to create a movie adaptation of the out-of-copyright Last and First Men using API access to ElevenLabs and Hugging Face for the first 10-15 minutes.

Ethan Mollick observes that while the lack of verifiable domains makes training harder in those areas, models are nonetheless improving at non-verifiable tasks faster than a verifiability-only theory would predict, describing the capability frontier as jagged but less so than expected.

Ethan Mollick contends the bigger issue is that people rarely attempt ambitious tasks with AI, and that agentic use of frontier models on long-horizon real problems is where impact lies.

MIT CSAIL describes Masked IRL, a method using LLMs to help robots interpret ambiguous chore instructions by having one model clarify prompts and another ignore irrelevant details.

Ethan Mollick describes repeatedly asking Claude Fable to make a game more AAA, observing it upgraded graphics, boss fights, mechanics, and audio until hitting WebGL limits.

Mollick notes that while lack of verifiable domains makes training harder, models are improving in non-verifiable domains faster than a pure verifiability lens would predict, describing a jagged but narrowing frontier.

The vLLM project details serving Qwen3-Omni in real time as a multi-stage pipeline (Thinker, Talker, Code2Wav), showing that replicating only the speech stages cuts first-audio latency from about 6s to 0.6s and yields 5.4x throughput.

Santiago Valdarrama shares a step-by-step guide to building a first agentic loop with Claude Code, explaining the non-interactive prompt flag, pre-approved allowed tools, and max-turns cap for running an automated code-and-test loop.