4 ms·

Speaking as someone not really involved in LLMs: i get the strong impression that prompt engineering is a scam (in the classical sense: taking advantage of an i

Speaking as someone not really involved in LLMs: i get the strong impression that prompt engineering is a scam (in the classical sense: taking advantage of an ill-understood new field to milk clients for cash). It’s difficult to believe that in a system that’s (a) non-deterministic, and (b) is prone to sometimes serious hallucinations, prompt engineering would be generally effective. Are there any good counter-arguments to this?

WhatsName 3y ago
It's as much a scam as SEO used to be, or still is. Yes there are some techniques, that help you get better results, like providing examples and being precise. Especially local LLMs tend to have their own litte quirks, which you'll have to work around. Same with SEO, yes, putting all you content in an image is bad, but as a rule of thumb ranking = relevancy * content * honest backlinks * google secret sauce. If you are able to avoid the obvoius footguns, there is no need to ever hire a Prompt-Engineer /SEO-"expert"
ickelbawd 3y ago
While they’re non-deterministic at a token level, they’re also still pattern following systems. Prompt Engineering is about (a) getting more deterministic outputs from the model, (b) using the models’ own bias to get more useful output, (c) refining the prompt to shorten it and reduce costs at scale, and (d) using them as building blocks for higher-level functionality. As a disciple (if it can even be called that) it may very well disappear as the tech progresses. Edit to also say: you are correct that there are a lot of people selling secret sauce or magic techniques and those are pretty scammy.
slowmovintarget 3y ago
First, you're right in that articles like this are spam for click-throughs. Writing prompts is like writing good search queries, and the techniques developed today will fall away tomorrow as the models and generation parameter settings get better. This has happened rapidly with diffusers, and will likely happen with language models. The system isn't non-deterministic, although there are a fair number of parameters used to try to make it seem like it (temperature in particular). But given the same prompt, zero temperature, and the same seed, you'd get the exact same output each time. The injected pseudo-randomness is deliberate. Having said all that, there are a handfuls of key phrases or instructions that do allow you to get more than you would from simple conversation English. For example, asking a model to use Tree of Thoughts and Chain of Thoughts in calculating its answers does have an effect. The same with asking "are you sure" for some models can yield corrections and refinements.
sottol 3y ago
My novice understanding of prompt-engineering is that it "guides" the LLM to certain parts of the output space. It has all this information related to a topic from a variety of sources, by creating some context (i.e. prompting) one can increase the likelihood of the LLM outputting text that's more related to that part of the training data that you'd want. For example, let's say the LLM learnt about color theory from: 1. academic articles, 2. artistic discussion forums, 3. reddit ( :) ). If you wanted to steer the output from a certain "niche" you could create context in the prompt that makes the LLM more likely to create output (ie a continuation of the prompt, sort of). Tricks I've seen is stuff like "You're a super-smart, best in your field XYZ, tell me something about...". But I bet you could similar good output by using lingo and vocab specific to either niche to "prime" the LLM. Take with lots of salt-grains, just my pet-theory.

WhatsName 3y ago

It's as much a scam as SEO used to be, or still is. Yes there are some techniques, that help you get better results, like providing examples and being precise. Especially local LLMs tend to have their own litte quirks, which you'll have to work around. Same with SEO, yes, putting all you content in an image is bad, but as a rule of thumb ranking = relevancy * content * honest backlinks * google secret sauce. If you are able to avoid the obvoius footguns, there is no need to ever hire a Prompt-Engineer /SEO-"expert"

ickelbawd 3y ago

While they’re non-deterministic at a token level, they’re also still pattern following systems. Prompt Engineering is about (a) getting more deterministic outputs from the model, (b) using the models’ own bias to get more useful output, (c) refining the prompt to shorten it and reduce costs at scale, and (d) using them as building blocks for higher-level functionality. As a disciple (if it can even be called that) it may very well disappear as the tech progresses. Edit to also say: you are correct that there are a lot of people selling secret sauce or magic techniques and those are pretty scammy.

slowmovintarget 3y ago

First, you're right in that articles like this are spam for click-throughs. Writing prompts is like writing good search queries, and the techniques developed today will fall away tomorrow as the models and generation parameter settings get better. This has happened rapidly with diffusers, and will likely happen with language models. The system isn't non-deterministic, although there are a fair number of parameters used to try to make it seem like it (temperature in particular). But given the same prompt, zero temperature, and the same seed, you'd get the exact same output each time. The injected pseudo-randomness is deliberate. Having said all that, there are a handfuls of key phrases or instructions that do allow you to get more than you would from simple conversation English. For example, asking a model to use Tree of Thoughts and Chain of Thoughts in calculating its answers does have an effect. The same with asking "are you sure" for some models can yield corrections and refinements.

sottol 3y ago

My novice understanding of prompt-engineering is that it "guides" the LLM to certain parts of the output space. It has all this information related to a topic from a variety of sources, by creating some context (i.e. prompting) one can increase the likelihood of the LLM outputting text that's more related to that part of the training data that you'd want. For example, let's say the LLM learnt about color theory from: 1. academic articles, 2. artistic discussion forums, 3. reddit ( :) ). If you wanted to steer the output from a certain "niche" you could create context in the prompt that makes the LLM more likely to create output (ie a continuation of the prompt, sort of). Tricks I've seen is stuff like "You're a super-smart, best in your field XYZ, tell me something about...". But I bet you could similar good output by using lingo and vocab specific to either niche to "prime" the LLM. Take with lots of salt-grains, just my pet-theory.