> Have benchmarks for every single task and sub-task that will happen everywhere a LLM is involved, so you can measure improvements
exactly. are there standard approaches or tools specifically for instrumenting LLM utilization, for applications rather than model development?
i use claude code a bit. i relentlessly push it to keep its work ephemera in source controlled and discoverable locations, following a lot of ozbrain's motivation. i also have claude generate task logs for all spawns. the theory with the task logs is to try and build some standard measures so i can see whether a thing i do (writing skills, experimenting with third-party MCP tooling, etc) actually helps in a way that is more robust than gut. i have yet to turn that effort into something useful. i'm just starting to get enough data to try to do some diving.
the problem is that this feels like the classic problem of "how do we measure productivity"? i was thinking very basic efficiency like "tokens in per tokens out" (maybe with a domain-specific value weighting function on the output tokens, e.g. code tokens > memory tokens > chat tokens) could be an possibility.
the point being, seems like you'd want to become pretty sophisticated about measurement to sell a product like ozbrain, as a basis for continuous improvement across major versions. or convincing people to use it, quantitatively rather than qualitatively.
curious if you're using something mostly-out-of-the-box to layer on visualizations for your dashboards?
relatively new to duckdb, love it so far, looking at alternatives for downstream visualization. so far just exporting datasets and piping into python scripts.
For a schema-first (vs. code first) approach (which I think would be a sweet spot for agent driven dashboarding), I'd suggest looking at https://vega.github.io/vega-lite/ or https://vega.github.io/vega/. A little higher level than full D3 but gives you a little higher level approach.
that could easily be trojan-horsed with links to malware if you are viewing it in a poorly secured setting (like public wifi), because you can't verify the origin. so the best we can say about the author is that we are getting inconsistent signals on how seriously they understand and implement security concerns. so better review that code carefully before use, rather than assuming their expertise from release notes.
> we need a way to preserve judgment that used to be developed through the struggle.
it's true. once you've gone "deep" for a few years in at least one technical domain, that depth transfers pretty well to the next big new thing you didn't know you'd have to learn when you started. i think the fear about the new regime is that people will be denied the opportunity to obtain depth in anything. like we'll encounter the human equivalent of domestication syndrome.
i remember when certain loud individuals believed that {managed memory | IDE auto-complete | statistical db optimizers | programming languages higher than assembly level | ...} were going to make everyone stupid. but the higher-order systems have continued to present rich problems to engage the mind and spark creativity. this era feels different though, the worry more pressing.
exactly. are there standard approaches or tools specifically for instrumenting LLM utilization, for applications rather than model development?
i use claude code a bit. i relentlessly push it to keep its work ephemera in source controlled and discoverable locations, following a lot of ozbrain's motivation. i also have claude generate task logs for all spawns. the theory with the task logs is to try and build some standard measures so i can see whether a thing i do (writing skills, experimenting with third-party MCP tooling, etc) actually helps in a way that is more robust than gut. i have yet to turn that effort into something useful. i'm just starting to get enough data to try to do some diving.
the problem is that this feels like the classic problem of "how do we measure productivity"? i was thinking very basic efficiency like "tokens in per tokens out" (maybe with a domain-specific value weighting function on the output tokens, e.g. code tokens > memory tokens > chat tokens) could be an possibility.
the point being, seems like you'd want to become pretty sophisticated about measurement to sell a product like ozbrain, as a basis for continuous improvement across major versions. or convincing people to use it, quantitatively rather than qualitatively.