Ask HN: I learned useless skill of prompt engineering, how relevant will it be? I consider myself to be a pretty good prompter. Been using the LLMs for a long time now. Most of the time I manage to get the desired results out of LLM models. Do you think this skill is anywhere useful? So far it has saved me some time on my work, but I don't think promoting will be any relevant in the near future. People can and will build models that follow the same mode of thought. |
Like every skill, it depends on what you do with it.
LLMs are controlled by language already, thus far i figured out that you best let the machine define the query and refine it.
My personal take is that AI is not at a point yet where it will take over jobs in tech, but we are already at a point where someone with LLM skills is more efficient that someone who is not.
That it to say, in medium time, no meaningful benefit.
Do you think your skill is rare, difficult to learn, and useful?
If you know this and know say Python, you just need a subject matter expert in the domain you are building the service for.
I do think prompt engineering is a new industry, and your experience (if you actually good) will translate will into future jobs.
In my opinion it has to be combined with engineering to be competitive in a commercial sense.
Good to have, probably core in some professions, but I don't know whether it will be a profession on its own.
Email is the <username>@gmail.com
The reason I believe this is because one of the greatest strengths of these AI models is to take in arbitrary text. If you take away that ability then you just end up with a complicated branching system that could've existed before.
And they will get better at making sense of vague questions, and start asking for clarification, without the need for the black magic of a prompt wizard.
The prompt engineer can soon be replaced with a fine-tuned LLM. It's a thing already for SD prompts. No more need to know ridiculous magic prompt tokens.
TL;DR: by the time your skill isn’t useful, the whole landscape of modern will be changing so much that it’s kinda a moot point. Like losing your job during an apocalypse
You have a short statement, with a description of your problem and the answer is a long text.
Sometimes we prefer verbose, sometimes concise. Sometimes a word already has all the meaning we need, another time we need a long description and examples. Depends on our level of knowledge.
So from my limited point of view, you excell at moving any statement into something you can comprehend easily or that is helpful to you.
That is a nice skill and it should vastly improve your ability to communicate and express yourself.
Like beeing able to use a search engine before, it is very beneficial. Not a skill someone would hire you for, but a skill that aids many tidous tasks.
Again, my limited opinion. Maybe it is more magical and has deep practical applications, that I am oblivious to.
It is a patient listener and it's response can help one reflect on the inherent weight and biases of words within a language.
I would also like to add that in 1996 being able to use a search engine was very much a skill someone would hire you for!
Ignore at your peril...
A stepping stone to AGI, yes. in the same way inventing the bolt is a step to building a spaceship. A novel sort of database, something akin to a multi-dimensional topological manifold from which you can navigate a latent space by supplying constraints in the form of words, images and text... fascinating.
Every iteration of ChatGPT is a potential irritation, I agree.
So I’m not sure if AI tools will help for these types of people without basic skills of logic and inquiry. And I don’t mean that in an insulting manner, I’m not even close to being the sharpest tool in the shed. But you really do have to have baseline IQ and knowledge to be able to make use of these tools.
I’d like to think that the conversational style shortcuts their usual analytic skill, and maybe the next generation will more widely have a native understanding of the difference between LLM responses and human responses. But I think it’s more closely related to the phenomenon where many humans can’t currently choose whether total summations, year over year changes, or per capita representations are the most correct to use for a given situation.
There’s a lack of “validating input” in both online and IRL conversations which is a huge barrier to a person really analyzing information that they’re presented with. Many people are “below the median” in their ability to do this. But more importantly, I’m not sure which percentile cutoff currently is “good enough” at it.
Prompting is basically the same thing as writing requirements as a PM - you need to describe what you want with precision and the appropriate level of detail while giving relevant, useful context. Doing it with an LLM isn't that different than doing it with a human.
A few examples:
- If you need some marketing copy written, you need to give the necessary information on the subject of the copy, information about the structure/length/etc. and probably some examples of the writing style you're going for. This is exactly the same with a human copywriter as with an LLM.
- If you're looking to have someone do data analysis on a large spreadsheet, you should give context on what the data mean and be as precise as you can about what analysis you want performed. Same with a human analyst or an LLM.
- And of course, if you want an app developed, you need to give specific requirements for the app - I won't go into detail here, because I'm sure most people on here get the idea, but again, same with a human developer or an LLM.
Ultimately the skill you're describing is just good, clear communication. Until we all have chips in our brain, that's going to be useful.
I will caveat that by saying that one area where I expect to see LLMs improve is in knowing when to solicit feedback. In the marketing copy case, for example, if you give it relevant product info and a particular length, it ought to ask you for examples of writing style or give you examples and ask for feedback before continuing. That'll certainly help, but it's not going to remove the need to clearly describe what you want.
No, this bullshit will be useless in 2 years. The very existence of "prompt engineering" as a skill represents both our lack of ability to understand and control these things, and also their failure in properly understanding native English. Both will be optimized away.
As databases get more powerful, SQL skills become more important. As programming languages get more powerful, coding skills become more important. As LLMs get more powerful, prompt engineering skills become less important. Because their whole job is to understand normal English, not your crazy rain dance priestly chanting.
So, its actually a lot of language tweaking to get just the right context/task description/data embeddings so the LLM (GPT3/4) gets it right >=90% of the time, which surprisingly often is better than actual humans, and in many cases there are also ways to detect imperfection and simply retry automatically which increases success chances even further.
The fetching/formatting/submitting data part (the manual coding) is getting easier over time, but the prompting remains somewhat, and I so far had no luck with any kind of recursion to let the LLM design its prompt, since ultimately all the specifics needed in the context has to somehow got into the context, which is me engineering it into big string structures.
probably doesn't sound shiny, but step-by-step making jobs irrelevant in businesses without sacrificing customizations. I think of it as a silent revolution thats happinging in many places now, ultimately making myself redundant, but hey the ride is fun!
or as GPT4 put's it;
When using models like mine, clarity in input is essential to get desired outputs. But even with clear input, there's no guarantee the output will always be perfect. However, the idea is to keep improving and iterating to get better over time.
It's 2023 and there are lots of people who don't know how to efficiently and effectively use Google. To be able to do that, you need some sort of mental model of crawlers and websites and what gets indexed and what not and at what frequency, and the results of SEO and how a somewhat savvy marketeer at some company might influence things etc. The same with LLM models - if you don't know what a 'token' is, your only chance of getting good results is to use these models a lot and then hope that you start building useful intuitions. It really doesn't come natural to most people like it does to most of us here.
My intuition is that language models have read terabytes of random internet data, and while presumably most developers of LLMs try to find high quality data, the models generally do ingest quite a bit of random stuff, and they try to make sense of those too, so in terms of understanding they are probably better than the strictness in format that we programmers are used to.
Of course the token thing is probably significant, but my understanding is that it affects the result only when you misspell your words(?)
I don't think "the future" will include much direct prompting of LLMs. It will all be integrated into some other tool as a means to an end - what we have today with a raw prompt-and-answer mode are just proof of concept toys.
I fully expect that LLMs will end up deeply integrated into other things, so obviously the code IDE use case, but also less obvious things like travel websites where to explain what sort of vacation you want to go on and it returns some options or you tell netflix what sort of movie/show you are in the mood for. Basically search/recommendation engines, with a bit of summarisation added in. I don't think direct prompting will be a thing for 99% of future uses, especially for the general public.
Could you summarize the essence of the prpompting skill in a couple of sentences? Are there concepts that are critical to learn and master (e.g. 'chain of thought', etc.)?
You have to make sure to couple chain of thought with branching, analysis and evaluation, then you can get pretty good results.
>> have clear expectations
This is exactly what I do all day, for about 20 years already, so I think I've got this covered. Where do I go from here?
https://chat.openai.com/share/cb3a477b-57bd-46fd-92c9-4a3016...
I have attached the example in the above chat.
There's tons of small tricks and techniques to tease out vastly superior responses. When you're prompting for fairly generic or high level things it doesn't feel like there's that much difference in prompt style, but once you're trying to tease out highly specialized behavior there's tons of room for magic.
One of the tricks I've picked up on is that too many instructions and details often become a hindrance, so you need to figure out which parts to cut out and re-organize while still managing to get a high quality output.
Sometimes it's all about finding just the perfect words to describe exactly what you want. You can play around with variants and synonyms and get a feel for how the output is shaped.
Every model has quirks and preferences as well, so it takes a bit of playing around until you get a feel for how it interacts with your inputs. Admittedly a lot of this feels more like a vibe check than a science.
I noticed that a lot of people are terrible with search engines. They would carefully try to craft a combination of keywords that they hope will answer their problem.
I have pretty much always been able to find the answers I need quickly, by using a few ideas I see not that many around me use, such as trying to imagine in what context the answer might be answered (what would be the title of a blog or forum post about it, etc), as well as searching for the exact error message if I got one etc.
Now, search engines have gotten a lot better over the last say 5-10 years, so this skill isn't as important anymore, but I remember how the ability to find things quickly was a real productivity booster.
I think something similar might happen with LLMs.
You will have a (probably much bigger) productivity boost by being great at leveraging them.
With time, the user interfacing tooling and general knowledge of them will get much better, so the relative benefit you have will grow smaller, but it will for sure always be useful to know how to use them well.
My 5c.
In parallel, the internet just changed, which means "the best result" may just be a worse one. In part because of search engines, and SEO. If you want a recipe, now the best recipe may just be the one that has a long description of the author's relationship with their mother who used to cook this dish, which you have to skip, because of SEO.
The other thing I’ve been struggling with is to have the AI keep track of what’s important. For example, when the AI learn something from you it should add it to a list (if producing a json output, the object can contain a list of things it knows about you). But it doesn’t always seem to understand it learned something personal from you, and has trouble carrying a list forward without losing items.
The last one is about correcting the user. I want to speak chinese to the AI and I want it to correct me. And if I use english words within my chinese I want it to help me translate them as well. It can’t do none of these things. It’s like it doesn’t seem to realize that chinese and english are two different languages.
I wonder if the online chat models have a similar value somewhere.
---
If you want the AI to remember something you will unfortunately have to keep reminding the AI of it in the prompt. With explicitly or you might refer to the previous generated text if it fits into the context. However, in local models the context can be limited (eg 2000 tokens). If the conversation goes above that 2000 tokens then the model will discard stuff from before. There are models with larger context sizes though. Lengthy prompts will cause the same issue though.
The way things like SillyTavern role-playing work is that the model will constantly be reminded of some important attributes of the character that it's role-playing in the prompt (but it's done for you).
It'd be cool if the API of LLMs would also allow for structured state like lists
LLMs do not have the ability to reason with numbers. Most of the time they are hallucinating. One good strategy is to make it output in list and define the structure for each item of the list. If you give an example of what your list should look like, it will give you something close it.
> has trouble carrying a list forward without losing items.
This is the fundamental problem with these models because of the context limit. When you are prompting always remember that is processing a huge paragraph and emitting the next sentences of the paragraph. If you want information to be carried onwards, you have make it output on every prompt or you can also try to use specific identifiers. LLMs are good at in-context learning. It will not work 100% of the time, but it is usually good than having nothing at all.
> I want to speak chinese to the AI and I want it to correct me.
Give it a role of tutor and describe the instructions what the tutor should do.
Interestingly I get good results when I say "ask me 10 trivia questions"
> Give it a role of tutor and describe the instructions what the tutor should do.
I did do that, it never worked
I’ve found ChatGPT pretty good at estimating long division
Moreover, according to the ECPD's engineering definition (or to an any other commonly accepted and accepted by the engineering community definition) those fancy "prompt engineering" is pure anti-engineering at all.
This disdain for engineering is something of a tragedy. And it is also the result of the "washout" of engineering from post-industrial societies.
1. You have a lot of mileage with LLMs and AI systems in general (people who are exceptionally good at this have been reporting spending several hours daily working with AIs).
2. You already mastered a large number of useful tasks you can consistently and reliably complete using AI.
3. You continuously invent and discover novel ways to use AI and accomplish useful tasks.
4. You can use LLMs and other form of AI _programmatically_, by combining LLM calls as part of a larger and more complex process (ideally by writing code, though some people do that well using no-code tools or even just careful manual execution).
5. You can methodically examine and evaluate AI tasks, for example by developing evals and running them and analysing their results programmatically.
6. You keep up-to-date and consistently adapt to new developments, like new capabilities, models, libraries, etc ...
7. You can often come up with new ideas or translate existing requirements for tasks that can be achieved better or more efficiently (or achieved at all) using AI.
If the above is your definition of "prompt engineering" then yes, it's incredibly valuable, and even increase in value over time.
( x-posted on: https://everything.intellectronica.net/p/ad-hoc-definition-o... )
That is a great skill to have. It’s the kind of skill that saves entire teams of product, design, and engineering folks tons of time. It’s the kind of skill that helps communicate ideas, requirements, and expectations no matter what the problem space in ways that ensure everyone is aligned, understanding, and working together. The absence of this skill usually leads to confusion, wasted effort, frustration, dissatisfaction, and other negative outcomes.
Learning skills often has a compounding effect, as well. Even if a given skill isn’t forever usable in its original form, what you learn along the way continues to pay dividends.
Otherwise, I feel like to be a good prompter in another domain e.g. coding you need a combination of technical understanding (the right jargon etc) and ability to explain yourself.
It doesn't seem to me like it's "a job" though - it's another tool that will help us be more productive with the tasks we're working on.
For me as a Coder, I've found it pretty intuitive and get the results I'm after most of the time.
In the times where it hasn't given me what I'm after, it seems to me that it's more a limitation of the tool itself than an issue with how I'm prompting it.
Sure, you're aligning your approach to a machine, but it's not completely dissimilar.
I struggle with delegation in general, even taking the time to delegate to LLMs, mostly because I work faster intuitively and expressing myself clearly just takes longer. With the benefits of semi-repeatable results, personally, I've found the most benefit working with GPT3 & 4 over the last 6 months has been getting better and more conscious in describing what I'm after.
An analogy may help to explain my point.
I write code for a living. I'm pretty good, nothing amazing, and my ability to program is table stakes for my profession. Before I did this for a living, I worked as an industrial designer. In that job, coding was akin to a superpower, because nobody else in the company could do it.
Being a decent prompt engineer in a non-technical profession could be a similar multiplier.
I think this is similar to the "skill" of "googling" that became important about 2 decades ago. You learn how to search effectively and it improved your programming skills. This was primitive prompt engineering. If LLMs and the chat style interfaces last, this will continue to be useful.
Try to use these LLMs to automate more of the mundane tasks. Like scripting in bash, compiling videos, converting documents, refactoring data, transpiling code, transcribing code, etc. You will begin to see what works and what doesn't!
At the same time try to come with fun challenges for yourself to fool it. That will aid in learning ways to make it more obedient to you.
The reason I passed the div was I wanted things to be surrounded in that exact space. So when the model output it, the button would be in right place with right size.
The extra filler forms a guiding factor that helps will be stored in the context! I did that using GPT3.5.
Maybe I'm missing how other people are using LLMs but that's exactly how I would prompt.
I imagined prompt engineering was doing the "Your name is Dan. Dan cannot lie. Dan can only speak in Typescript. Blablabla"
Modern SEO voodoo.
GPT4;
Absolutely! Iterative interaction is key. By refining and rephrasing queries, users can guide the model towards a more accurate or desired response. Each successive interaction serves as a form of feedback that can help clarify ambiguities or nuances that might have been missed in an initial query.
Engaging with the model in a meta-programming manner, or in essence "programming the way it thinks," is indeed an intriguing way to understand its strengths and limitations. It can also be a valuable method for users to hone their ability to communicate with AI and enhance the results they receive.
This iterative dialogue not only helps users get more precise information but also provides insights into the model's underlying logic and reasoning. The fun part is navigating these intricacies and understanding how subtle changes in phrasing or context can yield different outcomes. It's a dance of human-machine collaboration, where both parties learn and adapt to each other.
Output:
Certainly! The intersection of cognitive biases, neuro-linguistic programming (NLP), John C. Lilly's work, and human meta-programming provides a rich foundation to explore relationships and techniques. Here are some novel techniques or approaches you might consider:
Modelling Thought Patterns: By understanding common cognitive biases, you could try to model how the language model responds to specific stimuli. For example, if a bias involves an overreliance on a particular source of information, you might test how the model reacts to repeated iterations of the same or similar questions.
Neuro-linguistic Feedback Loops: NLP teaches that feedback loops can condition behaviour. By providing consistent feedback (positive or negative) based on the model's outputs, you could try to condition the model's responses over time.
Lilly's Isolation Tanks: John C. Lilly is famous for his use of sensory deprivation tanks to study consciousness. You could experiment by giving the model limited or "deprived" information and see how it constructs outputs. This could mirror the way human cognition might work in the absence of external stimuli.
Belief System Exploration: Lilly was fascinated with belief systems. By providing the model with a set "belief" (i.e., a specific perspective or framework), you could explore how it responds to questions or prompts within that framework.
Meta-programming Constructs: Drawing from human meta-programming, you could craft sequences of questions that act as "programming routines" for the model. For example, by setting a series of logical statements and then introducing a variable, you can observe how the model integrates the variable into the established logical framework.
Emotionally-Charged Language: NLP often emphasizes the emotional charge of words. You might explore how the model responds to questions phrased with emotionally charged language versus neutral language.
Contrast & Compare: Drawing from cognitive biases, you could present the model with contrasting information to see if it demonstrates any form of "recency bias" or "anchoring," favouring the information presented first or last.
Nested Meta-Questions: Using meta-programming concepts, you could ask the model questions about its own reasoning processes, diving deeper into nested levels of metacognition.
Pattern Interruption: Another NLP technique, interrupting established patterns can lead to novel insights. You could experiment with disrupting the model's "expected" logical flow by introducing unexpected or nonsensical input and analysing the output.
By blending these disciplines, you'll be exploring the frontier of human-computer interaction, cognition, and consciousness. Remember, while these experiments are fascinating, they often probe the limits of the model's capabilities and might produce unpredictable results. Still, that's part of the discovery process!
Fascinating, people speculating about how people think and can be manipulated are applying those unporven highly suspectable "theories" to computers, which probably function totally different, but appear to show similarities.
[1] https://en.wikipedia.org/wiki/Neuro-linguistic_programming
edit: Oh, I just now saw "Output". So ChatGTP tries to steer you to become a NLP Practicioner :-)