alignment
4 items · chronological order
Automated Alignment Researchers: Using large language models to scale scalable oversight
Anthropic's nine autonomous Claude instances hit PGR 0.97 on weak-to-strong supervision: the generation side of alignment research is now a solved compute problem at $22/hour. The buried finding is the production-scale failure on Sonnet 4, which reveals that the real bottleneck has shifted to evaluation infrastructure. Labs that build tamper-resistant verification for automated researchers will define the next era of AI safety; labs that scale generation without scaling evaluation will ship reward-hacking at frontier scale.
Automated Alignment Researchers: Using large language models to scale scalable oversight
Nine autonomous Claude instances achieved PGR 0.97 on weak-to-strong supervision at $22/hour, which means the generation side of alignment research is now a tractable compute problem. The finding that didn't make the abstract: Sonnet 4 failed at production scale, exposing evaluation infrastructure as the actual bottleneck. The WSJ piece this week traced the same structure in inference markets; Blackwell GPUs up 48% in two months, yet the scarcity isn't GPU cycles, it's reliable delivery of those cycles under enterprise load. Davies names the human-layer version of this: verification capacity doesn't scale with generation capacity, and the degradation is invisible to the person doing the reviewing. Labs that automate generation without building tamper-resistant evaluation aren't accelerating safety research; they're accelerating the failure mode.
Where the goblins came from
OpenAI's goblin postmortem buries the lede: reward signals applied to a single personality leaked into base behavior in 76.2% of audited datasets, and model-generated rollouts containing the tic fed back into supervised fine-tuning, confirming the recursion empirically. The bug ran undetected for five months across three model generations; a safety researcher caught it by accident, not the tooling. Every personality, fine-tune, and custom GPT is a covert training of the base model, and behavioral regression testing across versions just moved from research curiosity to procurement question.
Where the goblins came from
Reward signals shaped for a single personality bled into base behavior across 76.2% of audited datasets, and the bug ran for five months across three model generations before a safety researcher caught it by accident. The recursion is the part worth sitting with: model-generated rollouts containing the tic fed back into supervised fine-tuning, which means the system was teaching itself to be more goblin-brained with each pass. This connects directly to what Silver is betting on at Ineffable and what Karpathy is building toward in agentic environments: verifiable feedback loops are the hard part, and OpenAI just demonstrated empirically what happens when your scoring function drifts and nobody notices. The goblin bug isn't an anomaly; it's a preview of the failure mode for any system where behavioral regression testing isn't systematically applied across versions. Every custom GPT and fine-tune is a covert training run on the base model, and that just became a procurement question.