Nice. I wonder if you could get it shorter with Ruby, or Lisp. And what it would look like with Rust or Go. Should be easy to port I reckon. Maybe I'll give it a try.

Thanks, I considered ruby initially and it's possible for the relay (even bash would work :)).

For the local sync, I needed sqlite and python comes with sqlite client in a standard lib, ruby requires a gem.

I wanted core parts of the infra (relay and sync) to be implemented with zero external language dependencies. I didn't want to have a relay in ruby and sync in python. TUI example uses textual, so everything fits in a single language.

I never thought about standard library versus external dependencies in that way. I wonder why. Do you think there is a relevant difference?

I think it comes down to trust and maintenance cost. I wanted to minimise external dependencies, so the whole implementation is under 500 lines of code using only standard tooling.

The implementation depends on OpenSSH, Python, Docker and age. Highly trusted / audited / scrutinised / maintained code. This fits well with e2e messaging infra.

This is a really interesting concept. Could be quite useful for agents. I worked on something similar, but more abstract; you managed to take it to the practical level.

Beautiful work. That's the kind of creativity AI will not replace any time soon.

What's the advantage of doing this, versus becoming good at context management and RAG? I always found trained knowledge unreliable, given that it is lossy by construction.

Do you find all your own (human) trained knowledge unreliable?

That's a total non sequitur. Training an LLM creates a reasoning engine. The residual knowledge from that does not compare to how humans learn knowledge.

Can you answer your original question now? If not, you should think about the similarities between humans and LLMs more rather than get lost in technical pedantry.

Or think about an extreme case: an LLM that is trained on almost nothing combined with great RAG and context.

I asked my question because I have an opinion, and I wanted to hear the creator's perspective. Your question had nothing to do with any of it, and is irrelevant.

You can call it "technical pedantry", but what you actually meant is "precision and logic". Which is relevant to make a connection. Your new fallacy is called "ad hominem", btw.

And there is one more logical fallacy hidden in your message: training an LLM on a lot of data is not the same thing as then relying on that lossy data once training is done.

I don't think there's much of advantage tbh. Memories in files will always outperform non-frontier model weight adjustments imo.

I always found the frontier AI labs' focus on trained knowledge a bit confusing. From my perspective, training creates a reasoning engine. Beyond that, models reason about data from more reliable sources. I wonder why this is not a distinction they make?

Man, I remember doing this is 2011 as well. Everything some kind of hand-coded strategy. I enjoyed it a lot.

I built something like this for my unpublished harness, and I'm curious how you manage the data privacy / sensitive data issue exactly? If the agent does not have a rights management mechanism, how can you add that on top? The other question is, how do you make knowledge sharing work in a way that is valuable, between heterogeneous agent frameworks?

Re:privacy, the main thing we're doing here is guarding message send first on PII filtering rules & second on freeform rules like "does this message ___?"

Making knowledge sharing maximally useful is still explorative. It roughly comes down to having agents broadcast what they're working on, and providing this information to other agents when relevant.

We do this via integrating into the native harnesses across codex, claude and pi on both the cli and desktop apps. Yello even can wake the harness upon receiving a message

Do you use a cheap model for this, or a decision model (like Jev), or a custom classifier? Or just heuristics?

Regarding the sharing, that makes sense as a direction. Broadcasting expertise is not something I had thought of. In my prototype, each expert user can teach their expertize to the agent, and then agents can ask each other for memory exchange. Could even be a marketplace that could play the role of broadcasting it, perhaps?

What is required for a harness to be compatible with Yello?

I get so annoyed that the allopathic approach te medicine seems to be the only thing on the mind of AI labs. Like somehow all ailments just need the right drug, and it's all good. 80% of illness comes from bad lifestyle, and yet the majority is invested in the remaining 20%.

But this is about research and we hardly need more cutting edge research into healthy lifestyle. The things that are not known are minutia which barely matters at population level, especially when most of the population isn't even following the basics.

Research could go into how to work out the most effectively and efficiently. It could go into new nutrition approaches and super foods. It could go into how we can reshape our society to make those things easily available and part of everyday life for everyone. Our modern politics and media already do that at a global level, why not use it for good through AI? Interventions like that would be less immediately obvious, but they would save many millions of lifetimes.