Import AI 469: Science AI; RSI simulator; and Zuck's technological pessimism
Import AI 2 months ago 18
DiG-bench introduced a benchmark of discovery in games and found today’s frontier AI systems struggle to infer hidden environment rules compared with humans. The benchmark covers 70 games. As a result, leaderboard results hinge on tools like harness access (e.g., Claude Code) and suggest progress toward stronger autonomous “taste” and replication-style discovery is ongoing across multiple new evaluation setups.