So long “prompt engineering,” we hardly knew ya
1–10 of 76 posts
Re: So long “prompt engineering,” we hardly knew ya
#2I hate approximately everything about this article, but I'm glad that I took a second look through because this is a decent framework for the-thing-which-he-swears-isnt-prompt-engineering:
---
* diagnosis
* decomposition
* reframing
* constraint design.
Diagnosis is discovering the problem that AI can solve. This is the human part of knowing that a problem exists. Learning to ask the right questions, look at the different ways that the problem can be seen.
Decomposition is about splitting the big problems into bite-sized ones. Take the problem apart, examine it, and let AI help you determine your findings since it handles data so well. Instead of tackling the biggest problem, take it apart and work on the smaller parts to achieve small successes.
Reframing is about shifting your perspective and seeking new interpretations. Extrapolating and recombining the parts of the problem in order to identify the meta components. Perhaps a new way of looking at the problem may find a solution hidden in plain sight.
Constraint design is about setting boundaries for the solution. Knowing what to accomplish, and when to know it is done. Setting the length, style, and description of the audience can help AI understand its mission. But we have to know that first in order to instruct.
---
As someone who asks GPT and junior developers for lots of things, there are a lot of similarities. I don't imagine that is going away, at least until we wire LLMs up to a huge amount of rapidly changing, cross-silo context so it could understand "Fix the monitoring that slowed our recognition of yesterday's bug". So being thoughtful isn't going away. The author agrees with that (see above), but doesn't make clear where he draws the boundary between "being thoughful" and "prompt engineering"
Re: So long “prompt engineering,” we hardly knew ya
#3This article’s argument is interesting because it relies on two very questionable principles:
1) That as tools get better, operator specialization is less useful. I understand how this feels like it makes sense, because for for binary tasks like driving it’s true, but for creative tasks, I struggle to think of a single example where this is the case. Which leads to
2) Humans won’t figure out how to use these tools in increasingly complex/weird ways to create increasingly complex/weird outputs.
I think in general, everyone has been so conditioned by the idea of singularity (which to be frank is a completely tangential concept to contemporary LLMs) that they refuse to see these things for what they are: tools built by humans to serve humans when operated by humans.
Add as many layers of “self-prompting” as you want there, but a human still set the original intention and they will be the ones to judge the ultimate outputs.
Re: So long “prompt engineering,” we hardly knew ya
#4i feel like people that get really into prompt engineering
seem to forget that other people are working on building models that are even better. Do they really think that the tricks they developed to make gpt3.5 work well will still be necessary on gpt9 (or whatever new model replaces gpt)?
if you want a long-lasting career in AI you need to work on the actual AI stuff, not just using the AI stuff.
Re: So long “prompt engineering,” we hardly knew ya
#5I hate approximately everything about this article, but I'm glad that I took a second look through because this is a decent framework for the-thing-which-he-swears-isnt-prompt-engineering: --- * diagnosis * decomposition * reframing * constraint design. Diagnosis is discovering the problem that AI can solve. This is the human part of knowing that a problem exists. Learning to ask the right questions, look at the diff…
That's just systems analysis though. To call rediscovering that prompt engineering just because you're dealing with AI rather than programmers siloed away from business processes feels a little silly and just being justified to keep the title when as originally formulated by the people who first started using it, it was about learning the "magic words" for a specific LLM
Re: So long “prompt engineering,” we hardly knew ya
#6i feel like people that get really into prompt engineering seem to forget that other people are working on building models that are even better . Do they really think that the tricks they developed to make gpt3.5 work well will still be necessary on gpt9 (or whatever new model replaces gpt)? if you want a long-lasting career in AI you need to work on the actual AI stuff, not just using the AI stuff.
"if you want a long-lasting career in AI you need to work on the actual AI stuff, not just using the AI stuff."
I think this is more for people, though, that want to maximize the use of AI in their own field, isn't it? The "knowledge worker enhancer"?
I am not sure they really need to work on the actual AI stuff...
Re: So long “prompt engineering,” we hardly knew ya
#7i feel like people that get really into prompt engineering seem to forget that other people are working on building models that are even better . Do they really think that the tricks they developed to make gpt3.5 work well will still be necessary on gpt9 (or whatever new model replaces gpt)? if you want a long-lasting career in AI you need to work on the actual AI stuff, not just using the AI stuff.
I don't agree. There's a huge need for people to figure out how to integrate AI into larger things, and to establish ways we'll work with it - and at every step, including GPT 3.5 prompt engineering. There's a lot you can come up with if you just assume that good generative AI exists, and will probably get better. Different iterations of AI might as well be pluggable modules.
Re: So long “prompt engineering,” we hardly knew ya
#8i feel like people that get really into prompt engineering seem to forget that other people are working on building models that are even better . Do they really think that the tricks they developed to make gpt3.5 work well will still be necessary on gpt9 (or whatever new model replaces gpt)? if you want a long-lasting career in AI you need to work on the actual AI stuff, not just using the AI stuff.
> if you want a long-lasting career in internet, you need to work on the actual internet stuff, not just using the internet
- a C developer that writes modem drivers at a telecom company in the 90s about html/perl developers building first interactive websites
Re: So long “prompt engineering,” we hardly knew ya
#9i feel like people that get really into prompt engineering seem to forget that other people are working on building models that are even better . Do they really think that the tricks they developed to make gpt3.5 work well will still be necessary on gpt9 (or whatever new model replaces gpt)? if you want a long-lasting career in AI you need to work on the actual AI stuff, not just using the AI stuff.
I tend to think that as long as these models remain tools (rather than AI that is self aware, etc.), we will have careers available in both “mechanics and engineers” so to speak. Some will design, others will build, still others will integrate and maintain.
Re: So long “prompt engineering,” we hardly knew ya
#10I think a lot of prompt engineering is magic thinking. It apparently worked once so you keep iterating on it without knowing if it has a positive or negative effect on any continued use. Some folks at work have got really into it and have some elaborate prompts to use before asking code related questions. I cannot see any different in their output vs mine when I just straight ask gpt4 questions.