Honestly think one shots, the blue moon when they work, are usually the best output because it’s following one coherent set of goals and assumptions. Every other case is the model fighting contrasting decisions
For really small projects, I’ll often delete and restart
Everything else, which is most projects, have to fight against existing bias and wrangle models
@mstockton Honestly great of older too, extends the capabilities of a lot of the older generation who want to keep their mind active. Fun to see what my mom is even doing with it!
@krzyzanowskim@caguilar51 Binary search also requires the collection to be sorted, which obviously may or may not have been the case, but would be curious to see its approach
@themoneygps@DanielLockyer The crazy thing is more reasoning on the same model can be worse on the same task, so it’s very fractured as to what is best at any given times. Kind of the dark ages that we have to choose at all though
@elonmusk This is genuinely cool, curious on the cooling solutions that end up getting developed for this kind of stuff, pushing the laws of physics to the edge
I think gtp Astra is genuinely the first model where I would be okay just with unlimited use of it if models stop advancing
Far from perfect, but is good enough to start executing on some of my more challenging ideas
@burkov I often ask it to query the session history for what it does, it’s crazy this is still an issue though, but seems alignment has moved away from just hallucinating
Sad to see codex down on a Friday, it’s funny how things always happen on a Friday
It’s got to be a crazy infra challenge keeping all this up, would love to see into how they handle it
@RyanCarniato It’s crazy, but I’ve almost always fallen back to choosing rust recently, I think it’s typing just works well in the ai era, and used to only favor typescript for FaaS support and sharing code across the stack, but ai reduces that cost