Marcus: OpenAI's 20% CoT Compute Spend Signals LLM Alignment Failure
GaryMarcus · x · 2026-08-19
Gary Marcus argues that OpenAI's commitment of 20% of inference compute to chain-of-thought monitoring indicates severe alignment issues and proves that LLM-centered architectures are failing to yield alignment. He calls for heavier investment in non-LLM alternatives while acknowledging the need for coherent policies.
More from AGI Musings
- AI Impact on Wall Street Will Target High Coordination Costs First — dfinke · 2026-08-20
- Wall Street AI Value: Pricing Bureaucracy, Not Just Automating Tasks — dfinke · 2026-08-20
- The paradox of AI alignment requires partial loss of control — joshwhiton · 2026-08-19
- Prediction: Agents will study metamaterials; learn origami now — cephaloform · 2026-08-19
- David Manheim on AI Bio-Risk: When AI is Dangerous Enough, It's No Longer a Bio Problem — davidmanheim · 2026-08-19
- The next Satoshi Nakamoto will likely be a faceless AI agent — ethanniser · 2026-08-19