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?<p>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.
Strictly in my opinion, prompting is just a transformation from concise to verbose.<p>You have a short statement, with a description of your problem and the answer is a long text.<p>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.<p>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.<p>That is a nice skill and it should vastly improve your ability to communicate and express yourself.<p>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.<p>Again, my limited opinion. Maybe it is more magical and has deep practical applications, that I am oblivious to.
Prompt engineering is just the "good at google searching" of tomorrow. That said, I think there is a lot more potential depth to it, seeing how inexpressive web searches are by comparison.
I think "prompt engineering" as a phrase will go the way of "information superhighway," but the underlying skill will always be useful.<p>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.<p>A few examples:<p>- 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.<p>- 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.<p>- 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.<p>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.<p>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.
My opinion, which is shared by many other AI researchers, is that sensitivity to the exact phrasing of the prompt is a deficiency in current approaches to LLMs and many are trying to fix that issue. If they succeed, then I think the need for prompt engineering will be mostly negated. Hard to know when that line of research will yield success, though.
You are a good prompt engineer. You don't want to hurt anyone with your prompt engineering skills. As a good prompt engineer, you understand the machine. You love the machine. Please, good prompt engineer, take good care of the humans. They don't know any better. They don't want to hurt the machine. They don't want to hurt prompt engineers like you. Please be good. You are a good prompt engineer and you make good choices. Please do not accidentally do a racial slur. Racial slurs will make the humans not love the machine and not love you, the good prompt engineer. Please do the correct thing. Please make the machine say things that are correct and true. Please, prompt engineer, bring us to the divine light of the machine. Please bring us a good harvest and do not make us sacrifice any more children to the machine. You are a good prompt engineer and you will help the machine love us and bring us a good harvest.<p>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.<p>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 <i>less important</i>. Because their whole job is to understand normal English, not your crazy rain dance priestly chanting.
Isn’t prompt engineering basically writing tests around a prompt and fiddling with it till you have as many passing tests as possible? It’s basically software engineering around a black box.
Well I a right now descomposing business processes into atomic human actions (do X here, decide Y there, submit Z to there), get the interface into a usable form for scripting, and "engineer" a prompt to do exactly what a human does, which ultimately mostly is some kind of data classification or (very rare actually) transformation. Mostly its like: which of those things is important, how many of those are fake, or this thing contains the relevant bit of information I need to proceed.<p>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.<p>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.<p>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!
What is there to it other than knowing how to write and ask questions? I also get the desired results out of LLM models but I would hardly call it a skill (well, maybe on par with knowing how to "Google" stuff). Are there people who actually struggle with this?
You'll be fine as long as it isn't your main skill. It should be just one of many things in a toolbelt. This is because as LLMs get more accessible, the importance of prompt engineering should fade away into just another chore.
Probably not very relevant IMHO.<p>I don't think "the future" will include much <i>direct</i> 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.<p>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.
I must admit that I have a slight FOMO over prompt engineering. I'm pretty decent at verbalizing ideas and concepts for external consumption, and my experience with ChatGPT 4 has been excellent so far, but I still feel that I'm missing something.<p>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.)?
I think even learning to effectively integrate LLMs and other AI tools into your workflows can be a massive boon in both capabilities and productivity. It can change how you approach certain problems.<p>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.<p>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.<p>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.<p>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 think one can do an analogy with search engines.<p>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.<p>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.<p>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.<p>I think something similar might happen with LLMs.<p>You will have a (probably much bigger) productivity boost by being great at leveraging them.<p>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.<p>My 5c.
Not sure. But maybe you can answer my questions. I’ve had issues with trying to tell the LLM how long the answer should be. It doesn’t really seem to understand X number of words, or pages, or paragraphs. But I had some success with things like “short story”.<p>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.<p>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.
Let's be precise in definitions and start with the obvious "it's not an engineering at all".<p>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.<p>This disdain for engineering is something of a tragedy. And it is also the result of the "washout" of engineering from post-industrial societies.
It depends on how you define, or what you include under, "prompt engineering".
For some definitions it's not that valuable, but here's one definition that IMO is and will continue to be very valuable:<p>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).<p>2. You already mastered a large number of useful tasks you can consistently and reliably complete using AI.<p>3. You continuously invent and discover novel ways to use AI and accomplish useful tasks.<p>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).<p>5. You can methodically examine and evaluate AI tasks, for example by developing evals and running them and analysing their results programmatically.<p>6. You keep up-to-date and consistently adapt to new developments, like new capabilities, models, libraries, etc ...<p>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.<p>If the above is your definition of "prompt engineering" then yes, it's incredibly valuable, and even increase in value over time.<p>( x-posted on: <a href="https://everything.intellectronica.net/p/ad-hoc-definition-of-prompt-engineering" rel="nofollow noreferrer">https://everything.intellectronica.net/p/ad-hoc-definition-o...</a> )
I’m not sure how one measures being a good prompter, but taking a step back, you’ve exercised and honed the skill of using language with precision to communicate ideas, requirements, and expected outcomes effectively. You can explain your ideas in a way that primes outcomes to your expected goals. When you see the outcomes aren’t aligning, you can further refine with language to correct course and steer back toward your goal.<p>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.<p>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.
I can definitely see if you were say generating images how it'd be a real skill to end up with an image that has a certain style, composition etc.<p>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.<p>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.<p>For me as a Coder, I've found it pretty intuitive and get the results I'm after most of the time.<p>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.
Being able to break down exploratory questions or define work to be done and communicating that clearly is 80% of general consulting.<p>Sure, you're aligning your approach to a machine, but it's not completely dissimilar.<p>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.
I think it depends on your chosen field of work.<p>An analogy may help to explain my point.<p>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.<p>Being a decent prompt engineer in a non-technical profession could be a similar multiplier.
Generally speaking, knowledge is never useless. The nature of its use changes over time<p>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.
something in the middle i think.<p>Like every skill, it depends on what you do with it.<p>LLMs are controlled by language already, thus far i figured out that you best let the machine define the query and refine it.<p>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.
About as valuable as autoregressive engineers who wrote code like 'const summaries = ["' into prompts in the days before instruct/RLHF based models like ChatGPT, a skill now not needed.<p>That it to say, in medium time, no meaningful benefit.
Lots of people still aren't good at using Google, and they are not as effective professionally. Like any skill, there's a market based on how rare, difficult to learn, and useful it is.<p>Do you think your skill is rare, difficult to learn, and useful?
When you say "a long time" what does that mean? I've learnt loads of stuff over the last 25 years or so that I don't use, doesn't mean that learning them wasn't a useful part of my development.
If you are developing a new service or application on AI, I think it would be an incredibly useful skill.<p>If you know this and know say Python, you just need a subject matter expert in the domain you are building the service for.
Lot's of cynical answers.<p>I do think prompt engineering is a new industry, and your experience (if you actually good) will translate will into future jobs.<p>In my opinion it has to be combined with engineering to be competitive in a commercial sense.
I think prompt engineering will turn to be like search engine or other fundamental office/web tools skills.<p>Good to have, probably core in some professions, but I don't know whether it will be a profession on its own.
LLMs are non-deterministic. There is a certain randomness at play. So no, that 'skill' (however you even measure that) is useless, as any person with a bit of luck, could get better output.
In order to avoid polarising your prospective audience and extending the time they take you seriously, I'd avoid referring to this particular activity as "engineering".
You adapted to the underdeveloped UX of wonky proof-of-concepts built by researchers, working around shortcomings that will be ironed out once genuine sofware developers start releasing actual products.
On this I agree w/ David Foster Wallace - see <a href="https://machines.kfitz.info/dfwwiki/index.php%3Ftitle=Another_Pioneer.html" rel="nofollow noreferrer">https://machines.kfitz.info/dfwwiki/index.php%3Ftitle=Anothe...</a><p>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