Social Media Briefing — July 3, 2026

Chollet predicts AI will converge toward intuition-guided symbolic world modeling, i.e. deep-learning-guided program synthesis, for compact and generalizable models.

John Carmack clarifies that GPU utilization measures non-idle time and drops only when the GPU waits on the CPU, not from thermal or power throttling.

Mollick argues continual learning is the biggest barrier to explosive AI adoption and has implications for recursive self-improvement, since amnesiac models keep humans in the learning loop.

Mollick says claims about Anthropic Mythos and Fable models handling autonomous cybersecurity work were not hype, based on hands-on use.

Continuing yesterday's Social point on building your own benchmarks, Mollick argues you need your own benchmarks because different tasks favor different models, and warns against swapping models purely to cut cost without testing.

Continuing yesterday's Social thread on Bridgewater's Tinker fine-tuning, now with the joint results published, The Rundown summarizes joint Thinking Machines Lab and Bridgewater results showing that fine-tuning an open-weight model on expert judgment via the Tinker API reached 84.7 percent accuracy on news-filtering, beating frontier models with fewer errors and far lower cost.

The vLLM project announces native support for DeepSeek's DSpark semi-autoregressive speculative decoder, detailing implementation reuse of SparseMLA backends, CUDA graph capture, FP8 KV cache, and benchmarks of ~250 tokens/s at batch size 1 with 12-42% higher acceptance than MTP on 8x B300 GPUs.