talk feedback #4
Description
I just watched your talk on youtube and had a couple thoughts.
this likely fits better in discussions but I don't' see that enabled.
Regarding markdown diff.
- I was poking around the Zep paper on bitemporal memory systems and it seems like you could use that to get the current state and state history for code and markdown entities. The fact state history is essentially the decision history. you could link in the jsonl lines from the agent, the code, the spec in the graph. Claude was able to oneshot entity extraction from the Zep paper (https://arxiv.org/abs/2501.13956) and I think you could easily adapt it to experiment with.
- I wonder about doctests in general for your system. Have you tested model accuracy when the code, tests, and spec are spread out vs consolidated?
- Why share your code? I think this is a very interesting question. With software being much more disposable (due to the lowering of the cost ) I also feel like open source changes in a big way 3-5 years out. There will be a huge struggle with disposable software that no one is quite ready for we are already seeing the impacts from in the number of rejected PRs from AI generated code. Rejecting the PRs make sense but it will only push that code out into isolated islands and while a lot of trash gets filtered out some real gems will be lost.