This could work really well if it replaced GPT-X-judged performance ranking with human-in-the-loop ranking of prompts, but that’s not as exciting, I guess.
GPT-Prompt-Engineer
141–150 of 166 posts
Re: GPT-Prompt-Engineer
#142This tool doesn’t benchmark based on how a model actually responds to the generated prompts. Instead, it trusts GPT4 to rank prompts simply in terms of how well it imagines they will perform head-to-head. Thus, there’s no way to tell if the chosen ‘best prompt’ actually is the best, because there’s no ground truth against actual responses. Why is this so popular, then (more popular than promptfoo, which I think is a…
Because this is the upside if not the peak of the hype bubble. All you have to do is use GPT for a task, and nobody cares whether it actually works, you still get whatever VC funding is left after the interest rate hikes.
The market correction is gonna be gnarly.
Re: GPT-Prompt-Engineer
#143This could work really well if it replaced GPT-X-judged performance ranking with human-in-the-loop ranking of prompts, but that’s not as exciting, I guess.
I think the human in the loop evaluation of prompts is a red herring since the LLMs were trained with HFRL, which optimized them to be convincing. Humans aren't reliable for evaluating these things because it was literally trained to make us think it's doing well.
Re: GPT-Prompt-Engineer
#144Earlier quoted context omitted.
You’re missing the point here. It’s not even getting the LLM’s opinion on evaluating the responses to the prompts (which itself is fraught for some tasks, and benchmarks are known to be limited —even OpenAI admits this, it’s why they made evals). It’s one level abstracted from that. It’s evaluating what the LLM thinks of how well the prompt will do, in purely hypothetical terms. That’s hogwash —different LLMs perform…
> Here is what HN was talking about, nearly three months ago -the exact same type of ‘auto-prompt-gen’ tool. I was reminded of the same thing. What a lot of it boils down to is that LLMs have no innate ability to self-reflect . They can pretend to do it, but no more effectively than an untrained human would.
> They can pretend to do it, but no more effectively than an untrained human would.
Which is exactly as much as Generative AI should be trusted.
Re: GPT-Prompt-Engineer
#145This tool doesn’t benchmark based on how a model actually responds to the generated prompts. Instead, it trusts GPT4 to rank prompts simply in terms of how well it imagines they will perform head-to-head. Thus, there’s no way to tell if the chosen ‘best prompt’ actually is the best, because there’s no ground truth against actual responses. Why is this so popular, then (more popular than promptfoo, which I think is a…
There is a paper on arxiv saying that GPT4 correlation with human evaluators on a variety of tasks with strongly positive. I am also uncomfortable with it, but using GPT4 as a grader is not as bad as you think.
If we could use GPT-4 to grade prompts, we wouldn't need to talking about grading prompts to use for GPT-4, since this solution requires that the problem doesn't exist. The question then becomes, how do you grade the prompt grading, objectively? At the bottom, there has to be a ground truth.
You can't use the thing you're testing to evaluate its own performance. This applies to rulers, speedometers, and AI. It's the difference between a "subjective" and "objective" metrics. If you want an objective metric, you need to have it based on something external, based on reality, objective. Otherwise, you have metrics and ideas that have to held themselves up.
Source: My day job is test and measurement. These concepts go back centuries. You never trust your measurement system, you verify it against a standard.
Re: GPT-Prompt-Engineer
#146Earlier quoted context omitted.
You're forgetting what you may consider to be factual, self-evident and a priori is your opinion. You may be under the impression that annual U.S. deaths from medical errors being in the hundreds of thousands miscommunicates but that is truly your opinion. You are merely jumping to conclusions at places another person might not. And going on to rely on the LLM to validate your perspective is a lossy process. It may n…
This is an excellent example. The post you replied to was saying that the deaths were caused by miscommunication, but you interpreted it to mean that stating the number of such deaths is somehow a miscommunication itself!
Doesn't help that I had just recently woken up but yes, most definitely.
Re: GPT-Prompt-Engineer
#147Doug Adams would've had so much fun these days.
Imagine what Russell or Wittgenstein would think of all this.
Re: GPT-Prompt-Engineer
#148This tool doesn’t benchmark based on how a model actually responds to the generated prompts. Instead, it trusts GPT4 to rank prompts simply in terms of how well it imagines they will perform head-to-head. Thus, there’s no way to tell if the chosen ‘best prompt’ actually is the best, because there’s no ground truth against actual responses. Why is this so popular, then (more popular than promptfoo, which I think is a…
Because this is the upside if not the peak of the hype bubble. All you have to do is use GPT for a task, and nobody cares whether it actually works, you still get whatever VC funding is left after the interest rate hikes.
This seems to be some guy who made a MIT licensed project, so I'm not sure there's any reason to complain about VCs.
Re: GPT-Prompt-Engineer
#149Earlier quoted context omitted.
I don't get where this sentiment comes from. I build software specifically on the concept of predictable results from llm's being composable. Sure, the results are not deterministic in that 100% of the time the exact prompt returns the exact same result, but you can tune your prompts so that 100% of the time they give you a valid result in the result category you were seeking, and with a specific probability distribu…
but you can tune your prompts so that 100% of the time they give you a valid result in the result You can't though, that's the issue. Illustrative here are tokens like "SolidGoldMagikarp", but this does happen to "normal" sequences of tokens as well. There is no filter you can build to keep out such mistakes, any set of otherwise normal tokens could trigger the model to produce wrong output. Because of how large thes…
you could do a sort of validation of output by prompting the llm repeatedly with the same prompt and then compare the responses to eliminate outliers. I do feel like this stuff is magic though, just wanted to provide a counterpoint.
Re: GPT-Prompt-Engineer
#150This tool doesn’t benchmark based on how a model actually responds to the generated prompts. Instead, it trusts GPT4 to rank prompts simply in terms of how well it imagines they will perform head-to-head. Thus, there’s no way to tell if the chosen ‘best prompt’ actually is the best, because there’s no ground truth against actual responses. Why is this so popular, then (more popular than promptfoo, which I think is a…
Agreed. This is a pretty terrible idea.