8 ms·

If this stuff is so powerful, or going to be so powerful, is positioning like this wise? It kinda feels like it is aiming squarely at its own obsolescence. That

If this stuff is so powerful, or going to be so powerful, is positioning like this wise? It kinda feels like it is aiming squarely at its own obsolescence. That is, if LLMs turn out to indeed be "the way forward," than surely the esoteric/specialized science of prompt engineering will fall away precisely in step with how much better next-gen LLMs are, right? Or to put another way, what is the ultimate value of LLMs in general if they are not aiming to make the very idea of "prompt engineering" obsolete?

jasongill 3y ago
I understand what you are saying. When OpenAI's API was still new, there was a LOT of "prompt engineering" required - a lot of proper formatting and tricks to get it to produce the required output, proper ways of formatting prompts with "few-shot" examples, etc. Packages like LangChain sprang up because trying to chain together multiple prompts with different inputs and outputs was becoming a bear. But then, within a few months, the LLM's got better and better to the point where the average ChatGPT user doesn't need to be a "prompt engineer" to get what they want. Sure, there are tricks with adding specific phrases to a prompt to get it to do a better job, but what we now consider "prompt engineering" and what was "prompt engineering" pre-ChatGPT are very different. LangChain, for example, has reacted by pivoting away from just prompt chaining and templating and into an ecosystem for agents and automated AI "bots". I think that may be overkill for many use cases, and there are good alternatives (Rivet is now my go-to) if you still want to build an LLM-powered app that just requires a specific logical path and specific inputs and outputs, but honestly if I was working at an "AI startup" I would probably be trying to focus my time on getting away from "prompts" and more just to "doing a good job of what you are told". I do think that "prompts" (especially chat-style conversations) lack a lot of the nuance that you need for a production-grade app especially when it comes to specific input and output formats, etc. Function calling helps a lot but I don't see why xAI is focusing on an IDE for writing a great prompt instead of building an LLM that can make good, structured output with just an "ok" prompt. The next year will be interesting - it hasn't even been a full year since ChatGPT was released, and look how far the industry has come since the closed-beta release of GPT-3.
- heisgone 3y ago
  For me, I would just like a simple way (no technical knowledge) to invoke LLM sub-steps. Exemple "Find the 4 articles of laws most relevant to this case, then anaylse each of those articles separatly". So, a simple non-technical way of doing auto-gpt style substeps. This would greatly increate ChatGPT precision.
jejeyyy77 3y ago
what?
- beepbooptheory 3y ago
  Others have commented to maybe help with my bad writing, but: I am basically trying to say that the promise of LLM models are in the long term at odds with the idea of "prompt engineering" even being a thing in, like, two years. Every improvement to existing models can only really make such a thing more and more obsolete. Otherwise, what are we even doing here? Expensive regex?
ianbicking 3y ago
If you want standard responses then presumably LLMs will start to converge on those responses. But if you want a non-standard response then you have to ask for it, and I think that will continue to be a kind of "prompt engineering". For instance I'm working in math problem generation right now and you could imagine saying "make the problem harder" and you'll get back... something. But what does "harder" really mean? It's very underspecified. If you work in making educational material you probably have some much clearer criteria, and if you are building a product you may have a clear point of view about what you want in the next problem in a sequence... and "harder" is just punting on all these details. The tricks may stop being as necessary, but the deciding what you want and how to explain what you want seems essential no matter how smart the LLM is.
- darkwater 3y ago
  If you were to talk to a human that has enough context in the problem space, telling them "make it harder" should achieve something that makes sense... easily, no? I think that at some point a generically trained LLM will be able to do just the same. Prompt engineering will probably be a niche for security (both blue and red teams).
numpad0 3y ago
I suppose it's always a correct decision to vertically integrate and invest heavily in tooling, from a pure monopolist standpoint. Doing so reduces stakeholders and improve control, which should grant more power to management than constantly having to succumb to engineering, market or baseline reality. But e.g. very few laptop manufacturers are even capable of designing and fabricating own display panels(not sure Apple does?), let alone ones that provide any advantages, so there's that.

jasongill 3y ago

I understand what you are saying. When OpenAI's API was still new, there was a LOT of "prompt engineering" required - a lot of proper formatting and tricks to get it to produce the required output, proper ways of formatting prompts with "few-shot" examples, etc. Packages like LangChain sprang up because trying to chain together multiple prompts with different inputs and outputs was becoming a bear. But then, within a few months, the LLM's got better and better to the point where the average ChatGPT user doesn't need to be a "prompt engineer" to get what they want. Sure, there are tricks with adding specific phrases to a prompt to get it to do a better job, but what we now consider "prompt engineering" and what was "prompt engineering" pre-ChatGPT are very different. LangChain, for example, has reacted by pivoting away from just prompt chaining and templating and into an ecosystem for agents and automated AI "bots". I think that may be overkill for many use cases, and there are good alternatives (Rivet is now my go-to) if you still want to build an LLM-powered app that just requires a specific logical path and specific inputs and outputs, but honestly if I was working at an "AI startup" I would probably be trying to focus my time on getting away from "prompts" and more just to "doing a good job of what you are told". I do think that "prompts" (especially chat-style conversations) lack a lot of the nuance that you need for a production-grade app especially when it comes to specific input and output formats, etc. Function calling helps a lot but I don't see why xAI is focusing on an IDE for writing a great prompt instead of building an LLM that can make good, structured output with just an "ok" prompt. The next year will be interesting - it hasn't even been a full year since ChatGPT was released, and look how far the industry has come since the closed-beta release of GPT-3.

heisgone 3y ago
For me, I would just like a simple way (no technical knowledge) to invoke LLM sub-steps. Exemple "Find the 4 articles of laws most relevant to this case, then anaylse each of those articles separatly". So, a simple non-technical way of doing auto-gpt style substeps. This would greatly increate ChatGPT precision.

jejeyyy77 3y ago

what?

beepbooptheory 3y ago
Others have commented to maybe help with my bad writing, but: I am basically trying to say that the promise of LLM models are in the long term at odds with the idea of "prompt engineering" even being a thing in, like, two years. Every improvement to existing models can only really make such a thing more and more obsolete. Otherwise, what are we even doing here? Expensive regex?

ianbicking 3y ago

If you want standard responses then presumably LLMs will start to converge on those responses. But if you want a non-standard response then you have to ask for it, and I think that will continue to be a kind of "prompt engineering". For instance I'm working in math problem generation right now and you could imagine saying "make the problem harder" and you'll get back... something. But what does "harder" really mean? It's very underspecified. If you work in making educational material you probably have some much clearer criteria, and if you are building a product you may have a clear point of view about what you want in the next problem in a sequence... and "harder" is just punting on all these details. The tricks may stop being as necessary, but the deciding what you want and how to explain what you want seems essential no matter how smart the LLM is.

darkwater 3y ago
If you were to talk to a human that has enough context in the problem space, telling them "make it harder" should achieve something that makes sense... easily, no? I think that at some point a generically trained LLM will be able to do just the same. Prompt engineering will probably be a niche for security (both blue and red teams).

numpad0 3y ago

I suppose it's always a correct decision to vertically integrate and invest heavily in tooling, from a pure monopolist standpoint. Doing so reduces stakeholders and improve control, which should grant more power to management than constantly having to succumb to engineering, market or baseline reality. But e.g. very few laptop manufacturers are even capable of designing and fabricating own display panels(not sure Apple does?), let alone ones that provide any advantages, so there's that.