Social Media Briefing — June 29, 2026

The vLLM project announces support for Baidu's Unlimited-OCR using Reference Sliding Window Attention to keep KV cache constant, enabling 40+ page one-shot parsing and claiming 35 percent faster throughput than DeepSeek-OCR.

Delangue lays out a detailed case for regulating frontier API models for government transparency while leaving open-source AI unregulated, arguing closed black-box APIs are the real risk.

Mollick assesses that GLM-5.2 trails GPT-5.5/Opus 4.8 and is far from Mythos, but notes open weights have reached GPT-5.2-level capability, which is considerable.

Geoffrey Hinton recommends an excellent lecture by Adam Brown on AI's future impact on physics.

Mollick argues model routers underweight non-math/coding tasks, yet innovation, marketing, and qualitative work benefit most from smarter models.

Ethan Mollick argues AI model routers underestimate non-math/coding task difficulty, depriving non-verifiable tasks of smarter models.

Shumer argues that if frontier models like Fable and 5.6 are held back, open source will not save US users because the government is unlikely to allow downloading powerful Chinese weights.

Mollick asks whether Gemini 3.5 Pro is being export controlled, implying significance if it is not.

Gary Marcus argues the AI industry resembles airlines with thin margins and huge costs, suggesting the trillions invested may not pay off.

Nathan Lambert expresses optimism about the diversity of companies building open models, noting much untapped value lies under the shadow of frontier models.

Mollick observes that GLM sits right on the capability curve alongside Qwen, Kimi, and MiniMax, implying Mythos-class open models may arrive within 6-12 months if released.