lebek

The harness is just another codebase for the model to write and optimize. The value is still very much in the model.

Not a bad analogy because the bigger your backpack the slower you walk. With models a big context and tool set degrades performance. So you want the smallest harness/backpack that can do the job.

Codex and Claude historically had more bloat in their system prompt and tools. Pi is minimal by design so more adaptable. But to be fair Claude Code is moving in the Pi direction with a small system prompt.

The point is that it's general. It can do this task and many other tasks and it doesn't need custom development like OpenCV does. Of course if you only want to count pills and you want it to be cheap/fast you're still better off using OpenCV.

I'm building one! Features:

- /command to get current sentiment

- /command to submit feedback

- automatically tell you current sentiment when you start a session

> Then human work is changed to figuring out new things and the AI solves all old things, that seems much more fun than most white collar work today.

But it's not fun to be figuring out new things all the time. Some amount of routine work is necessary to 1) exercise mastery (feels good), and 2) recover energy. This is why a lot of people find agentic coding exhausting and less fun, you're basically always having to be creative (what's the next feature?) or solve the hardest 5% of issues the LLM can't handle.

I understand that, but it seems like even the MVP "shitty" flagellum would require many mutations that individually have no benefit. But I suppose with enough generations/parallelism you get enough stacking of useless mutations to reach the useful ones.

I think he's saying, random mutation wouldn't produce all required components at once. One mutation gives you a bit of a flagella, another gives you bit of a nose, but how does the flagella mutation survive to coexist with the nose mutation that makes it useful.

I suspect the answer is that having flagella without a nose is still better than having no flagella. If so it suggests evolution isn't good at accessing groups of mutations that aren't individually beneficial.

Link to said research?

> nobody at this point expects a 13B parameter model to succeed with the same accuracy at the broad range of tasks supported by what may be a 1T parameter model

I think a lot of people believe exactly that. To take one example from the "We Have No Moat" essay:

"It doesn’t take long before the cumulative effect of all of these fine-tunings overcomes starting off at a size disadvantage. Indeed, in terms of engineer-hours, the pace of improvement from these models vastly outstrips what we can do with our largest variants, and the best are already largely indistinguishable from ChatGPT." - https://www.semianalysis.com/p/google-we-have-no-moat-and-ne...