Unlearn to Learn
Matthias Lübken•ArticleIf you have listened to one of my talks, you have probably heard me say this before: I believe that in AI software engineering, there are more new things to learn than things to settle on.
Last week's AI Engineer World's Fair in San Francisco was a prime example of this. There are two talks in particular I'd like to recommend in this context:
Geoffrey Litt: Understanding Is the New Bottleneck
The first talk is by Geoffrey Litt. He has written many great articles, so I was eager to see his talk. The recording is not up yet, but he posted a write-up here: Understanding Is the New Bottleneck.
As Geoffrey does, he connects his thoughts to core principles. In this talk, he reasons about how we can better review software in the age of coding agents, where we get slammed with code. He proposes different techniques for better understanding. He has even published a skill /explain-diff in which he formalizese his review process.
It made me realize that I've thought about reviews the wrong way. Instead of reviewing the code, or the specs and prompts (which still might be useful in certain cases) why don't we use an agent to help us understand the important parts?
The fun thing is that this might differ from case to case and reviewer to reviewer. If you are familiar with the code, you might ask different questions than if you are new to it. And if you want to make sure you learned it, you might want to take quizzes, as Geoffrey suggests. Fun stuff. So much to explore, especially on the collaborative part. I can think of many very catered use cases.
Thariq Shihipar: Being Unreasonable
The second talk I highly recommend is Thariq's talk, titled Field Guide to Fable. Despite its title, I think this applies to all frontier models. Maybe Fable just stretches it even further.
And he posted a write-up of "Finding Your Unknowns" here: A Field Guide to Claude Fable: Finding Your Unknowns.
It essentially guides you on how to use the model to think outside the box. Like Geoffrey, Thariq is using interactive artifacts, like HTML, to help him reason and learn about things. Including quizzes.
What stuck with me even more was the idea of "being unreasonable". Everyone building with coding agents has experienced this before. We are getting things done that would have been unreasonable to even think about before.
I think this is a good guiding principle as we try to unlearn things. Question old beliefs and start with a fresh mindset. That doesn't mean everything will be magically done. Quite the contrary. The bar shifts, e.g. toward better learning. But I'll try to be more ambitious with my new work, and learn on the way.
This is also part of the thinking behind TAVON.ai. We help companies turn agentic workflows from demos into production systems. If you have a workflow that should be an agent, or an agent stuck between prototype and real use, please reach out. Happy to chat.