@JohnnyNel_ good to know. I don't have intuition on performance yet so I'll need to measure the ones I use to get at this. Many of my CRONs and heartbeats are boringly reptitiive.
We'll se after a week or so.
@loktar00 Thanks! I was on the fence but for repetitive task like what I have, this is a good ratio. I’ll drop $20 and give it a go…
Are you developing SW or agentic tasks?
Yes.
It’s not that we’ve discovered some magic bullet, but rather that JAX, or at least the open source version of it, is mostly optimized for small to medium-sized training runs on Google TPUs, whereas we need to massive training runs on Nvidia GPUs.
Pipeline parallelism is essential and crushes fully-sharded data parallelism at scale.
And C will compile to the most efficient binary short of assembly. Maybe we will do a little assembly too.
4 agents all with purpose. Fairly stable with @openclaw
Alfred: Director of Chaos
Sunny: Joy manager
Scout: Finder of cheap goods
Byte: IT plumber
Using XAi Super Grok, Gemini Free tier, openrouter free, and I’ll add DeepSeek V4 paid.
What agents get what models? Thoughts?
4 agents, all on free tiers, a small local model, due to my modest CPU only setup right now.
Becoming clear, local models are super important for tool calling, scraping, summarizing, etc. Large models for simple Internet tasks appears to be a poor architecture choice at today’s prices.
Converging on an approach that will work for me. More to explore here for sure.
An AI trick I've been using is that when I'm debugging, I collect all the commands I've used in a previous troubleshooting effort. I think offer that as a reference file or simply in the context window... makes everything snappier and avoids pooking around... I now call --help to get all the command options and include it too... seems to help
I’m fighting with my AI agent in OpenClaw because Sunny is convinced it doesn’t have access to the Internet while talking to me over the Internet.
Weird times
Ok I’ll be honest, I have a super modest CPU only machine and I’m having fun with @openclaw
Good job @steipete.
I’m at the point where I can point the files from one instance to another to improve it. Wild.
I don’t have much intuition on how these models behave at various parameter scales yet (might not ever). I do think the smaller ones are a bit limited so I’ve been using them as scrapers and summarizers…
It’s also clear that in many cases tools (ex Python scripts) will outperform the models spending tons of tokens…
Seems Wild West IMHO right now for us non ML professionals… not my field but super fun
I wanted to get back for a comment. For a survey course or even to hams it’s a good story of how many pushed to a limit and where it came from. LDPC has a great story too.
No maths needed, sets bounds and how many smart people got there.
Maybe I need to write this story up. :)
OpenClaw working ok with 5.7. I have two agents: Alfred and Sunny.
I still have random delivery errors and chain of thought issues where the agent announces what it’s going to do rather than just doing it.
Can’t seem to get them to stop with MD files and directives. Details. 75% operational at this point.