1,295 karma · joined July 16, 2017
https://github.com/baalimago https://lorentz.app
The browser has tons of functionality baked in, everything from web workers to persistence.
This would also allow for interesting ways of authenticating/manipulating data from existing sites. Say I'm logged into image-website-x. I can then use the WebMCP to allow agents to interact with the images I've stored there. The WebMCP becomes a much more intuitive way than interpreting the DOM elements
Oil[0] is fairly useless without being refined as well. Perhaps: "Data is the new oil, you need to refine it"?
And I agree, "async agents" makes little sense
I'll stick to LLM design, thank you very much
It's similar to how MCP servers and agentic coding woke developers up to the idea of documenting their systems. So a large benefit of AI is not the AI itself, but rather the improvements they force on "the society". AI responds well to best practices, ethically and otherwise, which encourages best practices.
The one tangible usecase is perhaps booking things. But, personally, I don't mind paying 5-10% extra by going to a local store and speaking to a real person. Or perhaps intentionally buying ecological. Or whatever. What is life if you have a robot optimize everything you do? What is left?
The bottleneck of the inference is fitting a good enough model into memory. A 80B param model 8bit fp quantization equates to roughly ~90GB ram. So 2x64GB DDR4 sticks is probably the most price efficient solution. The questions is: Is there any model which is capable enough to consistently deal with an agentic workload?
Question being: WHY would I be doing RAG locally?
Suggestion: continue in the current LLM-generated track and ask Claude (or whatever) to create an example + unit tests validating the idiom. Then tell Claude to remove half the example, leaving only a stub + failing unit tests. Add a go.mod at root + instructions on how to run all tests. The go initiate is "certified" once he/she has forked the repository and made the tests pass.
[0]: https://github.com/charmbracelet/crush [1]: https://github.com/baalimago/clai [2]: https://github.com/anomalyco/opencode
From experimentation, I need to coach the models quite closely in order to get enough value. Letting it loose only works when I've given very specific instructions. But I'm using Codex and Clai, perhaps Claude code is better.