Co-founder & CTO @ nu:legal. Building AI that does real legal work, with lawyers in the loop. Legal tech since 2017. Co-founder simplesurance (acq. Allianz).
@davis7 We switched to strict TDD and humans only review contracts. For me how a model talks matters less than if the tests pass. But for sure last thing I want is to go back to Claudish :)
@mattshumer_ Just gave another try to "6.1 Sol xhigh" in 2 different tasks. Not very complex refactoring tasks on testing setup, just measure and decide... Absolute garbage results... OpenAI really has to do something about coding performance.
@GergelyOrosz End of summer we rewrote our platform and completely over engineered it on purpose. It took about a month to decide and document what we wanted, and we shipped it in 3 weeks. The features you would normally have at 2nd or 3rd year of your company, we have it at our first months.
@kunchenguid We already figured this out 2 months ago and instructed our agents not to do so. They are only allowed to write unit tests in case of complex internal logic that can't be easily tested by e2e user journeys.
@biz84 The longer and harder working part is the same for us, but this is a common complaint I heard multiple time from outside of software as well. On the contrary we communicate person to person a lot, I mean as in hours a day a lot.
@TripInChina The amount of workers we see in this photo is probably higher than the total number of construction workers currently working on Germany's autobahns.
@GergelyOrosz We switched to strict TDD, and humans only review contracts. We run only 1 round of AI review on the implementation. 2 months in, working really well.
Codex & GTP5.6 Sol combo is like therapy when it comes to communication with the AI model. Anthropic should learn a lot from OpenAI about it. Talking with Claude is an anger management session.