- Diogo Almeida argued that RLHF is part of why AI shines at human-in-the-loop work such as deep research but fails at plain automation.
- Our reading: if a human must always check, the valuable skill shifts from making the work to judging it.
Diogo Almeida, cofounder of TypeSafe AI, argued that RLHF, the method used to train chatbots, is part of why AI shines at human-in-the-loop work but fails at plain automation. Models can look superhuman at hard maths yet fail at accounting and drive-thru ordering. He made the case on the AI Engineer show, speaking at CODE Summit 2026. The episode was published on 15 November.
What was said
RLHF means reinforcement learning from human feedback. People rate a model's answers, and the model learns to produce the kind people like. Almeida said this is why all AI today needs a human in the loop.
He split tasks in two. For deep research, he said, the models beat him. But if the answer had to be correct rather than well written, he said he would probably pick himself.
For automation, he claimed, pleasing a human has nothing to do with doing the task. He said even tasks that are easy to verify still need a person to check the result afterwards. In his words: "That is what the HF is in RLHF."
He said every year brings an infamous failure of AI at automating drive-thrus. His advice on unattended work was blunt: "for actually doing run-in-the-background cognitive work, you should not trust LMs at all because they have not been optimized for that".
Why it matters
Almeida spoke about verification, not careers. Our reading is that his argument points to a change in where skill sits. If a person must always check the output, the scarce ability is judging whether it is right.
For managers, that changes the sums. The time saved by a model is partly spent on review. An accounting team using AI still needs people who know what a correct ledger looks like.
It also raises a training question. People learn to judge by doing the work first. If software does the making, firms must decide how newcomers build that judgment. Almeida did not address this in the excerpts.
The other side
This is one speaker's claim, and he is building an alternative. Almeida said there is a better way and that AI should be designed for machines rather than for human assistance. The excerpts do not show how that works or whether it fixes the failures.
He also conceded limits. He said models are excellent as assistants and that he would not pick himself over ChatGPT for chat queries. He said very easily checked tasks can run on their own.
He mentioned models trained on checkable results, called RLVR. He said they are very bad to interact with unless they have also been trained with human feedback. He called RLHF "part of the problem" but gave no evidence that it alone causes the drive-thru failures, so the link remains his argument.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: I love you so jevving much — Diogo Almeida, TypeSafe AI (2026-11-15)





