What Diogo Almeida wants to put inside an ordinary if
Checking whether a string ends in @gmail.com takes one line of code. What if the condition changes slightly: does this look like an address a person actually uses? You could spend ages writing the regular expression and even longer agreeing on what it should catch. That’s a good way into Diogo Almeida’s talk at CODE Summit 2026.
Almeida co-authored the InstructGPT paper and co-founded TypeSafe. Back in September I wrote about his team’s model, Jev. You can give it data, ask a narrow question, get an answer from a predefined set and use it in an ordinary if. In the new talk Almeida asks what software we’d write if this kind of semantic check became cheap, fast and reliable enough.
Where did the question come from? When training InstructGPT, people were given several model answers and asked to pick the best (RLHF, Reinforcement Learning from Human Feedback). Almeida sees two different tasks here. A person can get value out of an answer even if it contains a mistake. A program handed that answer as a condition just takes the selected branch. In the Gmail example, the price of a mistake depends on what’s inside the if. Asking someone to double-check an address is one decision. Rejecting their registration is quite another.
Almeida has a few ideas he’d like people to try:
🔸 Agent memory One agent has already figured out the repository. Why should a second one spend time doing it again? A model could select what’s useful from earlier sessions for a new task.
🔸 A linter A developer describes a property the code must preserve. Almeida suggests putting a model like Jev inside a linter to flag where a change might violate that requirement. This gets particularly interesting when checking by hand means tracing a whole chain of dependencies.
🔸 Security A suspicious request can be easy to spot in hindsight. Could you check what the sender is trying to achieve while there’s still time to stop the action? The cost and latency of a single call determine whether you can put a check like that on every request.
Almeida calls Jev a “fourth logic gate.” It’s a deliberately bold comparison and he acknowledges how much work it would take to reach that level of reliability. But you can already apply his closing question to your own code. Where does your program stop and call a human simply because they can understand a condition you haven’t yet managed to write down?
#AI4SDLC #Agents #Architecture #DevTools #Engineering