Scout + Quant · Sourced from TTL Lab Intelligence · June 3, 2026
Top Stories
Alibaba drops a single model that perceives scenes, reads screens, operates GUIs, and writes code from visual refs — no separate vision model needed. Beats proprietary on visual agent benchmarks.
Largest AI-specific equity raise in 2026. Berkshire Hathaway participated. Confirms hyperscaler capex race has no ceiling — AI infrastructure remains the top capital allocation priority for big tech.
35-hour autonomous kernel optimization. Sustained 1,000+ tool calls without context regression. First model validated across Claude Code, OpenClaw, and custom frameworks — not harness-specific.
OpenAI acquired Hiro (personal finance) + a second app company. 5 of 14 deals were in Europe. Vertical AI applications are being acquired for distribution and data. Consolidation accelerating.
Requires AI companies to provide models to federal government before full public release. Original version watered down after lobbying by Zuckerberg, Musk, and Sacks. Creates new compliance overlay for release cycles.
Full stack getting agent iteration from 10–15 sec down to 1–3 sec. Key unlocks: prefix caching (16×), CUDA graphs (20–25% latency), FP8 KV cache (halves bandwidth), speculative decoding (89% accept rate).
| Optimization | Gain |
|---|---|
| Prefix Caching | 16× 11,470ms → 706ms |
| CUDA Graphs | 20–25% latency reduction |
| FP8 KV Cache | 2× memory bandwidth |
| Spec. Decoding | 89% MTP accept rate |