I think about the canonical Reality Has a Surprising Amount of Detail post a lot when trying to automate tasks with LLMs. In particular, any given task has many granular details, most of which don’t come to mind before making contact with reality oneself. The most common failure mode I run into is failing to specify various details that I hadn’t even *realized* were relevant things that could be messed up, before the model encountered the situation and messed them up. This also seems relevant when thinking about the transition from verifiable math and coding problems to more open-ended tasks like research or robotic manipulation.
This is why I find even going through the ritual of writing an LLM prompt can clarify my own processes and goal—as it is forcing me to explicate a whole complex of hidden assumptions or even expose points where I realize I simply don’t have the knowledge or information yet.
I must admit a poverty of imagination; I can’t see how it can be automated. That would be amazing if it could be.
However, the circumstances of each problem or LLM request are always so unique that outside of certain vague guardrails that apply to all problem solving/advice giving (In my experience these take the form of the questions: What have you tried already? Why did you try that way?/what did you expect to happen? What happened instead?). I see the ritual as attempting to explain why this situation is really unique and different – which seems to me to be the antithesis of automation.
However, if the situation isn’t unique, then maybe that can be automated. Realizing “oh this is analogous or really similar to this other thing I did”.
Let’s take two examples of situations I’m likely to ask an LLM for help with – “how do I hear the voices in this video stream better?”, and “what is the word to describe the way professions or taxonomies are divided in Platonic Dialogues? I keep getting the name of the two dots on a vowel[1]”
In the sound example, to avoid the boilerplate answers like “check if your audio drivers are working” or “turn up the volume”. I need to think about what I’m actually expecting here, and what kind of helpI actually want – and realize that what I want it to do is tell me how to route the video stream through OBS so that I can use a compressor in the chain to boost it… and if I know how to do that, maybe I don’t need the LLM to even reply to me: I’ve written a prompt and answered myself. This is what I mean that just the ritual of writing a LLM prompt sometimes clarifies things.
However, with the Platonic word example, any harnesses about OBS, signal chains, or software are not going to be relevant, hence can’t be automated.
I think about the canonical Reality Has a Surprising Amount of Detail post a lot when trying to automate tasks with LLMs. In particular, any given task has many granular details, most of which don’t come to mind before making contact with reality oneself. The most common failure mode I run into is failing to specify various details that I hadn’t even *realized* were relevant things that could be messed up, before the model encountered the situation and messed them up. This also seems relevant when thinking about the transition from verifiable math and coding problems to more open-ended tasks like research or robotic manipulation.
This is why I find even going through the ritual of writing an LLM prompt can clarify my own processes and goal—as it is forcing me to explicate a whole complex of hidden assumptions or even expose points where I realize I simply don’t have the knowledge or information yet.
I can’t help but wonder if this process could be automated? Or at least greatly sped up/eased by a good harness?
I must admit a poverty of imagination; I can’t see how it can be automated. That would be amazing if it could be.
However, the circumstances of each problem or LLM request are always so unique that outside of certain vague guardrails that apply to all problem solving/advice giving (In my experience these take the form of the questions: What have you tried already? Why did you try that way?/what did you expect to happen? What happened instead?). I see the ritual as attempting to explain why this situation is really unique and different – which seems to me to be the antithesis of automation.
However, if the situation isn’t unique, then maybe that can be automated. Realizing “oh this is analogous or really similar to this other thing I did”.
Let’s take two examples of situations I’m likely to ask an LLM for help with – “how do I hear the voices in this video stream better?”, and “what is the word to describe the way professions or taxonomies are divided in Platonic Dialogues? I keep getting the name of the two dots on a vowel[1]”
In the sound example, to avoid the boilerplate answers like “check if your audio drivers are working” or “turn up the volume”. I need to think about what I’m actually expecting here, and what kind of helpI actually want – and realize that what I want it to do is tell me how to route the video stream through OBS so that I can use a compressor in the chain to boost it… and if I know how to do that, maybe I don’t need the LLM to even reply to me: I’ve written a prompt and answered myself. This is what I mean that just the ritual of writing a LLM prompt sometimes clarifies things.
However, with the Platonic word example, any harnesses about OBS, signal chains, or software are not going to be relevant, hence can’t be automated.
Diairesis, not to be confused with Diaeresis