My entire coding setup: one terminal. No IDE. No browser. No Electron eating 16GB of RAM. LINCE is a terminal-native multi-agent workstation. Sandbox, dashboard, voice input, session persistence. Minimal dependencies.
https://t.co/FasZxRJRWB
@antirez The model? I don't think so, eve if maybe gpt is better it's not a disaster. But if you refer to claude code as product I'd agree with you it's growing an a so confuse way it's becoming a bad product and hard to use
A few days ago, I was reflecting in an exchange with @antirez how much, especially in the era of AI, people are not reading docs...
Here is my write-up about how to mitigate the bad effect it could have on how they perceive (Open Source) projects: https://t.co/PTazPUJV7r
@antirez I spent a bit of time on this and you have PR with that skill. I generally like the lazy newcomers experience conversation with the agent to understand new projects
@ivanfioravanti Indeed. The tricky part is we are trying to apply well known processes designed for the previous paradigm. Rarely this works, and we most likely need to design a new way to work in this new flow
@antirez I spent a bit of time on this and you have PR with that skill. I generally like the lazy newcomers experience conversation with the agent to understand new projects
@antirez This is brilliant and agents friendly (
by design. As good side effect it could be possible defining an onboarding agent skill, guiding new users on understanding concepts and commands in a conversational way. I may have time tomorrow to give it a try and see how it looks like.
@antirez I was one of the people who didn't understand ๐
I see now your point, and I'm sure you will keep the bar high keeping the *aggressive deprecation* to don't fall back in yet another inference engine
@antirez I see, my bad I misunderstood it. I thought you want to focus on a superoptimized inference for an architecture. In this case, giving something that work on 64gb is the way to go
@ivanfioravanti Can you elaborate on qua tization making it worse. Just an impression or measured somehow on tokens generated. Pretty interesting from a theoretical perspective.
@ivanfioravanti You are tempting me... ๐
I need to finish a bit of project I care,. And get and adeguate hw... but I m more and more interested in local inference and in those aspect specifically. Do you have any hint about stats on verbosity of recent small enough for local models?
@ivanfioravanti This is where the harness and the system prompt become important, paired with the inference engine. The art of system prompt may be slightly different on local, finding the trade off between correctness and verbosity of answer and thinking
@ivanfioravanti And there is another dimension to consider: how.many tokens to get the same final result? A concise model is.better than a verbose one if both get to the correct valuable information (the part of the answer that contain.the real info)
@ivanfioravanti More generally we should start to track. Intents of the developer and of the model. Decisions and intents is what really matter nowadays, code is just an artifact
Noce 3 days at pycon Italy. Happy my workshop on agent readiness of your code base was full. Great to receive positive feedback as well about my talk on how robotics (phisycal AI) can be weird for a software engineer
@antirez Yup, what worries me most is their revenue. If the pace doesn't change, it will reach levels that cannot be sustained soon. 2B total computer users in the world: who is paying the bill of 800B$ in 2027? More PIL?Unlikely..the only answer seems lower cost of occupation.
@ivanfioravanti My default answer would be no, but I was surprised to see how many people were already trying them in my last workshop at PyCon Italy. The limit is still the memory on the local computer, though (for me too). And Mac with >=128Gb is super expensive and hard to collect
@YYYYOOOO77@ichozero@antirez 1again, it depends. In complex prigramming tasks, gpt seems better recentely. But depends on the task, when the level is very high it's matter of preference sometimes (and also of reliabili6of service as you suggested)
@ichozero@antirez Yup you always need to try yourself your tools. And models are tools, not source of truth. In many complex tak I prefer 4.7 compared to 4.6 (rust specifically), but indeed it have minus on other stuff. It's like comparing IDE though, again just another tool