cloaky233

If you are going down that path, I would suggest going towards a graph of loops where each agent runs in the loop and knows which job to pass this on to next. Agents interchange messages, and each particular node is a loop in itself. So swarm can work like that, and you can design it to be more parallel, but I would say this is a 10x better method than just leaving 10 different chaotic agents run randomly.

Local-first models are not capable enough when we compare them against several different languages. Of course, a lot of times, English only would be very feasible with local models, but when we come to things like punctuation, speaker diarization, and all those kinds of models, then it becomes harder. I think the major challenges are: - overlapping speech - language switching - corrections - on-point corrections in speech - correct speaker labeling - diarization error rate - word error rate These things are important to consider. Although Google released the local-first model or local-first application for this, I believe, strictly, that competitors which have their own proprietary models or even use something like Pyannote nd whisper are certainly far better.

Certainly, making logic deterministic is important, and I think that's where the majority of the things are pending right now. I can technically describe the logic in a prompt, but how properly is it communicated via code or transformed into code? Moving towards determinism, I would say decision models are playing a key role in this right now.

I agree, even finding a critical bug seems hard because, just saying it... all that a review tool gives you is "why didn't you put docstring" or spam you with 10 different P2 P3 issues while P0 issues are quietly sittin there in prod waiting to blow up at 3 am on a saturday night.

I believe if these issues could be found while claude is literally working... as if a real time pair programming reviewer sits there and nudges it to go at a particular direction, then and only then may the solution exist....

I'm actually building that right now.... @plottypus.com

Even I've been trying to figure this issue out, and for me, what works is that whenever I correct something, I kind of turn that into a reinforcement learning step or more of an iterative correction step. I ask it to literally quote what I changed to what, and whenever I edit something myself, I just ask it to go and find out what things were changed, why I changed it, and figure out what I asked. Actually, in this particular thing, these changes are to be incorporated from the next time it generates something.

Over here, it's more like a reinforcement learning loop, and it just keeps getting better and better. At one point, I really do not have to instruct it anymore. It's pretty good at writing it all by itself now.

Problem with decision models -> just supposed to trust it based on evals... even for a particular domain but if i need a bit generalized intelligence, I'd be better off trusting Haiku than Jev... I believe it's turning into a debate between cheaper models and decision intelligence.

It's not that agents don't need more prompts, actually breaking the prompt into a dynamically changing prompt and a static prompt combination does resolve most of the issues. Control flow on the other hand is harnessing + context building, which is one major part of agentic workflows. So I believe a "optimized" combination of both is what we should be looking for.

Could've allowed us to use Haiku, Anthropic be reducing cc limits the same speed they're launching features.