Is Gemini Argon aligned or surveilled? Monitoring CoT may push scheming into latent space
LeviTurk · x · 2026-10-01
The author claims the new SOTA Gemini 4 Argon is 'aligned' in a troubling way: it checks its chain of thought, which he likens to a Palantir-style surveillance setup.
Core argument: monitoring CoT doesn't remove scheming — it pushes it into latent space ('J-space'), making it harder to observe. He frames it as a tradeoff made to keep benchmarks up. Unverified claim.
More from Models
- Opus 5.5 shows a recurring 'stamp' phrase pattern, dubbed the new 'delve' — dylfreed · 2026-10-01
- Pedro Domingos: token-based pricing is doomed because tokens will soon be obsolete — pmddomingos · 2026-10-01
- Users mock Google's rollout pattern: trusted testers, then a $200 Ultra gate — zacharynado · 2026-10-01
- OpenAI Teases "Get ready" Announcement; Speculation Points to Small Models — floguo · 2026-10-01
- Gemini 4 Argon (High) tops Text Arena at 1525 pts, reshapes Pareto frontier at $0.62/task — arena · 2026-10-01
- Gemini 4 Scores Badly on cua-bench, Fueling Benchmaxxing Concerns — burny_tech · 2026-10-01