Comment by lkjdsklf
2 days ago
This is the only approach that seems even remotely reasonable
“Prompt engineering” just seems dumb as hell. It’s literally just an imprecise nondeterministic programming language.
Before a couple years so, we all would have said that was a bad language and moved on.
It's technically deterministic, but it feels nondeterministic in chatbots since tokens are randomly sampled (temp > 0) and input is varied. Using the right prompt makes the model perform better on average, so it's not completely dumb.
I like task vectors and soft prompts because I think they show how prompt engineering is cool and useful.
https://arxiv.org/pdf/2310.15916
https://huggingface.co/docs/peft/conceptual_guides/prompting
> It's technically deterministic, but it feels nondeterministic in chatbots since tokens are randomly sampled
Are you not aware the random sampling makes something non-deterministic?
I'm saying LLMs are deterministic and because of that, prompt engineering can be effective. You knew what I was trying to say, but chose to ignore it.
You should follow the HN Guidelines. I'm trying to have a discussion, not a snarkfest.
> Be kind. Don't be snarky. Converse curiously; don't cross-examine. Edit out swipes.
> Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith.
I don’t believe that when he wrote that, he was using his own intelligence
2 replies →
A few years ago we didn't have an imprecise nondeterministic programming language that would allow your mom to achieve SOTA results on a wide range of NLP tasks by asking nicely, or I'm sure people would have taken it.
I think a lot of prompt engineering is voodoo, but it's not all baseless: a more formal way to look at it is aligning your task with the pre-training and post-training of the model.
The whole "it's a bad language" refrain feels half-baked when most of us use relatively high level languages on non-realtime OSes that obfuscate so much that they might as well be well worded prompts compared to how deterministic the underlying primitives they were built on are... at least until you zoom in too far.
I don't buy your past paragraph at all I am afraid. Coding langues, even high level ones, are built upon foundations of determism and they are concise and precise. A short way to describe very precisely, a bunch of rules and state.
Prompting is none of those things. It is a ball of math we can throw words into, and it approximates meaning and returns an output with randomness built in. That is incredible, truly, but it is not a programming language.
Eh, how modern technology works is not really the part I'm selling: that's just how it works.
Coding languages haven't been describing even a fraction of the rules and state they encapsulate since what? Punch cards?
It wasn't long until we started to rely on exponential number of layered abstractions to do anything useful with computers, and very quickly we traded precision and determinism for benefits like being concise and easier to reason about.
-
But also, the context here was someone calling prompting a "imprecise nondeterministic programming language": obviously their bone is the "imprecise nondeterministic" part, not distilling what defines a programming language.
I get it doesn't feel warm and fuzzy to the average engineer, but realistically we were hand engineering solutions with "precise deterministic programming languages", they were similarly probabilistic, and they performed worse.
4 replies →
Do you mean, like, scripting languages? Are the underlying primitives C and machine language? "Might as well be well worded prompts" is the overstatement of the century; any given scripting language is far closer to those underlying layers than it is to using natural language with LLMs.
Sure doesn't seem like it. https://x.com/jarredsumner/status/1999317065237512224
And forget scripting languages, take a C program that writes a string to disk and reads it back.
How many times longer does it get the moment we have to ensure the string was actually committed to non-volatile NAND and actually read back? 5x? 10x?
Is it even doable if we have to support arbitrary consumer hardware?
5 replies →
> nondeterministic programming language that would allow your mom to achieve SOTA results
I actually think it's great for giving non-programmers the ability to program to solve basic problems. That's really cool and it's pretty darn good at it.
I would refute that you get SOTA results.
That has never been my personal experience. Given that we don't see a large increase in innovative companies spinning up now that this technology is a few years old, I doubt it's the experience of most users.
> The whole "it's a bad language" refrain feels half-baked when most of us use relatively high level languages on non-realtime OSes that obfuscate so much that they might as well be well worded prompts compared to how deterministic the underlying primitives they were built on are... at least until you zoom in too far.
Obfuscation and abstraction are not the same thing. The other core difference is the precision and the determinism both of which are lacking with LLMs.