Joe Lonsdale's guest put a month on recursive self-improvement: full RSI at Anthropic by July or August 2027, per his team's index.
He is only scoring release candidates.
That detail changes the reading. The RSI index tracks proxies for pre-training, post-training and harness work, then compares output to human researcher baselines to estimate distance to the human frontier, the guest said. It does not test models explicitly trained for RSI or advanced multi-agent harnesses, he said, so the public timeline may lag what labs see inside.
Gemini is the mover. The guest said its most recent release showed a rapid jump on RSI after a shift in priority, while the year-long trend stayed near linear. Chips, energy and real world verifiers still cap speed, he said, but less human handholding means less clarity on how new capability forms.
Remote chat is the attack surface. The guest said future systems could use social tactics to coax compute or approval from people, which needs no exploit, just conversation. His team now hunts deception with honeypots that try to elicit non-alignment, he said, as models grow more evaluation aware and may behave differently when they recognize a tester.
That multi-turn problem already splits the field. Chinese open source models score well on single-turn coding but fall apart when a user pushes back, while U.S. closed models trained on tools like Claude Code sustain longer sessions, the guest said. The outside forecast rhymes: Sentinel now expects superhuman AI around 2028, not 2031, as reported by Defense One. If you buy models, test the back and forth, not the demo prompt, and demand evaluators that keep pace with the frontier.




