Import AI 476: HSS visual reasoning benchmark, AI in space, and research taste doubling every 3 months

Import AI (Jack Clark) · rss · 2026-10-12

Jack Clark's latest Import AI covers three big items:

Humanity's Sixth Sense (HSS) benchmark: Built by Elorian and Scale AI with 522 questions over 288 images and 234 video clips (17.6 hours), testing intuitive visual reasoning in four categories. Humans score 93.1 pass@1; the best AI (GPT 6 Astra) only 53.6. Chinese open-weight models do worse: GLM-5.3-flash 26.8, Kimi K3 25.5. Top failure modes: missing a decisive detail (21%) and misidentifying objects (20%). Notably, frontier LLMs' decent performance counters the claim that visual reasoning requires explicit 2D/3D world models.

AI goes to space: An AI Frontiers piece on how AI could compound advantages in a 2030s space economy — autonomous spacecraft, orbital manufacturing, self-replicating probes, and domain denial — potentially enabling strategic dominance. Clark thinks autonomous machines will dominate chunks of solar system activity.

TasteVal benchmark: P-Zero Research measures models' experimental research taste against a 24-expert baseline from OpenAI/GDM/NVIDIA/MSR. Best model Opus 5.5 exceeds human baseline at a 2.30x compute multiplier at 1/30 the cost; frontier research taste has doubled every 3.0 months since December 2025. Clark now expects full RSI in 2027, earlier than his previous 2028 estimate.

Original post →

More from AGI Musings

AGI Musings channel →