Language models are excellent assistants, but unreliable automators. That is the claim of Diogo Almeida of TypeSafe AI. If you want to automate a task, first ask what an error nobody sees costs, and where a person or a fixed check keeps watching.
What Almeida says
Almeida spoke at CODE Summit 2026; the talk has been on the AI Engineer YouTube channel since 15 November. He says he helped create ChatGPT and RLHF, the training method in which people rate the answers of a model. His company is building its own model, Jev, so he also has a stake in his story.
Assisting and automating are two different tasks
- AI excels where a person judges the result, as with Deep Research. If the answer has to be correct rather than well written, he says you would still pick a person.
- At the same time, he says, every year another attempt to automate something as simple as a drive-through fails.
- His explanation: RLHF trains models to please people. That makes a good assistant, but says nothing about finishing a task correctly.
- According to him, all language models today are optimised for this. He does not back up that claim in the talk.
Einstein or Mr. Bean
His image: you get Einstein 95 percent of the time and Mr. Bean 5 percent. If you cannot tell which of the two you have, you cannot automate anything. According to him, that reliability does not show up in benchmarks.
Our view: a smart answer is not yet a reliable process. That is why you design the process first and only then decide where AI gets a role.
What this means for your organisation
Almeida's claim is a point of view, not an established fact. The question he raises is useful for every task you want to automate:
- With a draft quote or a summary of a meeting, a colleague reads along. That is where AI belongs as an assistant.
- With invoices, orders or payments, nobody spots an error in time, and it gets expensive. That is where a person or a fixed check belongs in the loop.
What you do now
- List the tasks you want AI to run, noting for each task who still sees the result before it has consequences.
- Decide per task what an unnoticed error costs. If that is a lot, let AI make a proposal and let a colleague or a fixed rule decide.
- Split a large task into steps you can check, instead of one agent that does everything in one go.
How it works under the hood
The HF in RLHF stands for human feedback. That is why, according to Almeida, today's AI always needs a human in the loop, even for tasks that are easy to check: then that person checks afterwards. His proposal is to let AI make a decision and leave the execution to ordinary code, because in his view code is the language of automation. He also argues for solid software engineering: splitting tasks up, intermediate steps you can verify, and keeping track of the state of a process yourself instead of leaving everything to the memory of the model.
This news touches chapters 19, 22 and 23 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.
What you can do with this
Want to know what this means for your business? Plan an introduction: half an hour, no sales pitch, but an honest answer to the question where to begin. Rather read on first? The course AI(l) or nothing is free and starts with the basics.
Written in response to a publication by AI Engineer (YouTube): https://www.youtube.com/watch?v=Cyt3LU8TyNc
S.O.S