Christian Stewart retweeted
Richard Feynman lecture: "The difference between Dick Feynman and other people is that he is more logical and has a greater imagination. Most people who are logical are not very imaginative. As the great Russian physicist Frankel said, "knowledge progresses by illogical steps".
LLMs are like Bruce Lee's version of "Be like water" - they will flow to fit any shape or constraints that you put on them. But, uncontrolled water is chaotic and goes everywhere. Linters, style guides, software engineering skills, and your own preferences should shape it!
LLMs have only short term memory - context - and attempts to bolt-on long term memory are hacks, at best. The model needs a portion of the latent space which is persistent and long-term where it can keep the "mental map" of how everything works in a nonverbal way.
Oh My Python: a fork of Oh-My-Pi with the IPython single-tool kernel approach of prime-agent. The omp tool surface is fully remapped to the IPython kernel, for the best of both worlds. Try it out! github.com/paralin/oh-my-pyt…
Prime-agent feels like the way AI Agents are meant to be. After using it for some time, all the other harnesses seem like side-roads, divergences from the main-road of having a single tool given to the agent: Send input into IPython (interactive python) and get the result out.
I'm working on a lot of enhancements to the prime-agent harness to improve token efficiency and I'm sending PRs upstream, with a few big ideas rolling out now to drastically increase efficiency. But until those land, you can use it today at my fork: github.com/paralin/prime-age…
Christian Stewart retweeted
The weak AI code gen we had until late last year was most useful to low-skill programmers -- it was raising the floor. It was essentially useless to high-skill programmers -- you could move faster and ship better code without.
This has been completely flipped: the strong AI code gen we have now is *most* useful to high-skill programmers, while low-skill programmers are vastly underutilizing it or sometimes drowning in it. It went from a crutch to a power tool.
Richard Feynman lecture: "The difference between Dick Feynman and other people is that he is more logical and has a greater imagination. Most people who are logical are not very imaginative. As the great Russian physicist Frankel said, "knowledge progresses by illogical steps".
I switched from Codex to @_can1357 Oh My Pi because reading thought summaries is such an excellent way to debug agent behavior, and it's missing in Codex today. OMP also has a lot of nice features that aren't overwhelming and roboomp is a genius community-collaboration system.
Saying something to someone is a request for them to parse what you said and expend mental energy trying to understand it. That's why it's better to chose the few most important things you want to communicate instead of brain-dumping your entire line of thinking.
AI is not conscious. It's a computer, and it runs code. Code written in English. The more you think in terms of this framework, the better your experience with LLMs will be!
using LOC as a success metric for LLM use is of course absurd, but using LOC in terms of "lines added equal to lines removed" or even better "lines removed," actually is. deleting things or consolidating or making cleaner patterns is a good metric.
chat feels like telling someone else to do work for you. setting goals and generating + reading documentation feels like programming with English.