Diogo Almeida, a co-author of RLHF and member of the OpenAI post-training team who co-authored GPT-4, ChatGPT, and InstructGPT, said on the AI Engineer podcast in mid-2026 that "Today's AI was designed for assistance through optimizing for human preference." He argues the distinction between assistance and automation explains contradictions in AI: models approach research-grade mathematical reasoning but cannot handle customer service without humans, and Claude Code drifts from user intent as it becomes more autonomous. Almeida contends these issues trace back to "minor decisions" in post-training optimization algorithms, requiring an entirely new objective rather than a better version of the same one. He maps the industry into two camps: techno-optimists, who cite every NLP benchmark surpassed and autonomous operating time growing exponentially, and skeptics, who note value creation near zero, everything shipping as a chat app or cloud tool, and circular financing. Both camps cite real evidence, Almeida says. His answer is that successes, like conversational coding and research-grade problem solving, keep humans in the loop evaluating output, while failures are tasks whose goal is to remove the human entirely, running unattended on a server and making defensible decisions.