Diogo Almeida: Why 'Smarter Chat' Isn't the Same as Software That Runs Without You
"The next era is not the Claude Code or Codex era. Claude Code is still part of the assistance era. What’s next after RLHF is models built for software, not for the human in the loop."
In this 36-minute talk, the co-creator of ChatGPT explains why models that are superhuman at chat still fail at real automation, and why that gap is baked into how they were trained.
The key?
RLHF optimized AI to please the person watching it. That makes great assistants and mediocre unattended systems. Visible uncertainty gets punished, confident-sounding errors get rewarded, and the human never leaves the loop.
One of the clearest breakdowns of why “smarter chat” is not the same as “software that can run without you.”
Together with this Jev AI Guide, you will be able to build AI agents, workflows, and decision systems on a model that’s 200x faster & 400x cheaper than LLMs 🤖: https://lnkd.in/dxNeiCaA
Instead of another Netflix series tonight, watch this talk.
The talk and speaker are real, and the thesis is accurately represented — but there are two things worth flagging: the exact quote wording isn't something I could verify verbatim, and the "200x/400x" figures push toward the high end of previously-reported numbers rather than a typical result.
Diogo Almeida and the talk are confirmed real. He's credited as a co-author of GPT-4 and InstructGPT (the paper underlying ChatGPT's training approach), and gave a real conference talk titled "What's Next After RLHF?" — its actual documented thesis matches your summary closely: current AI, including Claude Code, is still firmly in the 'assistance era' defined by RLHF, which optimizes for human preference rather than autonomous reliability, and this is why LLMs excel at human-in-the-loop tasks but fail at true automation.
The three-part argument structure matches what's documented. Independent coverage of the talk describes exactly the progression in your post: today's AI is great at assistance but poor at automation; RLHF was designed for assistance by construction; and tomorrow's AI must be redesigned for calibrated decision-making and automation.
One caveat on the quoted lines specifically. I found a substantive third-party summary of the talk confirming its argument and title, but not a full transcript I could check word-for-word against the exact quote in your post ("The next era is not the Claude Code or Codex era..."). The substance is well-attested as Almeida's actual argument — but I'd treat the precise wording as a close paraphrase unless you've confirmed it against the video transcript directly, since it's easy for a quote to drift slightly in repostings even when the underlying claim is accurate.
"200x faster & 400x cheaper than LLMs" — plausible on pricing, but at the ceiling on speed, consistent with a pattern seen elsewhere with this same product. On cost: Jev's $0.042/million input tokens with free output versus frontier LLM pricing in the $2-15+/million range makes a multiple in the hundreds plausible depending on the comparison model and input/output token mix. On speed: TypeSafe's own published workflow benchmarks report up to roughly 193.6x on favorable tasks — so "200x" is a reasonable rounding of that reported ceiling, not a fabricated number, but it's worth knowing (as flagged when this same figure came up earlier) that independent reviewers have explicitly noted these headline multiples sit near the high end of what customers should typically expect, not the median result.
One small credentialing nuance: "co-creator of ChatGPT" is common shorthand used across multiple outlets for Almeida, but it stems specifically from his InstructGPT co-authorship (the RLHF-based training methodology ChatGPT was built on), not literal authorship of the ChatGPT product itself. Not inaccurate as commonly used, just worth knowing where the claim comes from.