Ex-OpenAI researcher cites RLHF limits in AI automation
InstructGPT co-author Diogo Almeida warns that RLHF models suit human assistance, not unsupervised enterprise automation. Here is why current tools fail.
InstructGPT co-author Diogo Almeida argued at Modal’s Runtime conference that foundation models were designed for conversational help, not unsupervised workflows.
Silicon Valley routinely promises autonomous agents that will take over full corporate roles. Yet one of the researchers who helped build today’s flagship models says the underlying math does not support that pitch. Diogo Almeida, a former OpenAI researcher and co-author of the seminal InstructGPT paper, brought that skepticism to Modal’s Runtime conference in San Francisco, pointing directly to RLHF limits. In remarks covered by The Information, Almeida cut through industry hype, asking why genuine automation remains scarce across corporate deployments. His answer: commercial models were trained to satisfy human raters in dialogue, not to run multi-step tasks without adult supervision.
RLHF limits in enterprise workflows
Almeida spent more than four years at OpenAI working on model alignment and reinforcement learning from human feedback (RLHF). That technique gave ChatGPT its characteristic polish, training systems to favor responses that human raters find helpful and articulate during an interactive chat. It works remarkably well for answering queries, drafting memos, and troubleshooting ideas.
The breakdown happens when enterprise teams plug those same conversational engines into backend processes that demand strict logical boundaries. Because RLHF rewards plausible, reassuring language, the model lacks hard operational constraints. In complex multi-step pipelines operating without human review, small inaccuracies compound until the entire workflow breaks down. For Almeida, the tech industry mistook conversational fluency for deterministic execution.
The onstage clash: assistance versus autonomy
Almeida’s argument sparked debate among the roughly 800 developers gathered at The Midway. Other speakers at the event pushed a very different timeline. Scott Wu, CEO of Cognition, the company behind the Devin software development assistant, maintained that recent advances in reasoning models are rapidly bringing full autonomy within reach for programming tasks.
That friction mirrors what engineering teams see inside companies. Tools that sit alongside developers and suggest code have enjoyed massive adoption. Meanwhile, ambitious projects meant to replace customer support or tier-one operations teams frequently stall under heavy debugging costs, uncertain handoffs, and the stubborn necessity of keeping humans around to catch hallucinated outputs.
Structured outputs and the commercial test
After leaving OpenAI, Almeida founded TypeSafe AI, which raised a $40 million seed round led by venture firm DCVC in September 2026. The startup is building Jev, a system trained with reinforcement learning from compiler feedback (RLCD). Instead of simulating a digital employee, Jev targets typed, deterministic outputs designed to integrate directly into existing enterprise software rules.
That alternative strategy must now prove itself against the massive developer ecosystems of OpenAI, Anthropic, and Google. Industry teams are also debating whether inference-time reasoning tricks can patch these reliability gaps without abandoning familiar chat architectures altogether. OpenAI has not issued a formal response to Almeida’s remarks.