if you want a long-lasting career in AI you need to work on the actual AI stuff, not just using the AI stuff.
Back in the day, when those new fangled relationship database things came on the scene, do you think people would have been well advised to try and find work on the actual database engine itself, instead of the more frivolous work of using the new technology to, say, solve actual business problems?
I don’t think LLM offer standardized API like that ?
If you are not rigorous, then what you are doing is essentially "black art". It may work for some tasks ad-hoc, but with the rapid pace of model improvement your skill will likely become irrelevant/not needed quickly.
You need some amount of experimentation to get the best results but in my experience what works for one model does nothing or worsens the output in others. Adding loras and different types of images into the equation makes this so variable that I would never consider it useful besides keeping a few key words I used to get x good result on y model and experimenting with those when I start a new project.
Calling it "prompt engineering" seems odd.
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...
The latter latches on to the surface level of any new tech but proceeds with swagger befitting a Nobel laureate.
Prompt engineering is not the same as programming. In some regards, it is “better”. In other regards, it is “worse”. They’re two different yet similar disciplines, each with their own strengths and weaknesses, but both equally legitimate.
Programming is simply the skill of being able to communicate with a compiler/interpreter effectively.
Prompt engineering is simply the skill of being able to communicate with an LLM effectively.
I wish we weren’t so divided over an incredible new technology.
https://www.reddit.com/r/datascience/comments/14nbwfv/where_...
But, even today prompts and rules of writing are not transferable even between existing LLMs. Future LLMs will have different architectures and requirements. I suspect today's prompt will be split into data and prompt. Probably just references, keywords to actual data.
In other words current state is transitory, next will be very different.
I mean, you think those graphics optimizations we pour millions of dollars into before releasing AAA games will matter when the GTX 8020 outperforms a 4090?
Hint: Delivering value for actual people is rarely the result of sitting on your hands and waiting for the next big platform, or even rolling up your sleeves and trying to learn how to build the next big thing.
You have this idea that by investing energy in something that will be obsoleted you're losing out, but spoiler, that's how 99% of software that delivers actual value works.
The cutting edge of tech almost always ends up being PaaS/SaaS serving itself:
Your mail gets to you because someone is working on software with limitations we solved decades ago.
Your paycheck ends up in your bank account because people invest a ton of time in codebases subject to problems we solved long ago.
Your anti-lock brakes aren't built on a Rust codebase, but some horrible memory unsafe mess running on a processor that's a decade out of date.
—
The reality is: 99% of the effort that goes into trying to build the next big thing goes nowhere. The expected value of you trying to learn "the actual AI stuff" to the greater world is near 0 compared to you "just prompt engineering" and putting out something that solves a pain-point nicely with GPT 3.5.
At the end of the day most of the value that gets delivered to actual users comes from engineers who went deeper into extracting value from the current thing.
We still need people to work on the next big thing so that the 1% of effort that isn't wasted can actually materialize... but in my experience the most successful engineers in that regard are still able to realize the delusion it requires, without being paralyzed by the cognitive dissonance that realization invites.
a different analogy that gets at my original concern: becoming an expert prompt engineer for an particular LLM is like becoming a power user for a piece of proprietary software that isn't getting any more updates.
In tech we take it for granted that just because there's some new hotness everyone wants to jump on it day 1. GPT 5 could drop tomorrow and if your tool delivers value using 3.5, it's not going to magically stop delivering value, and in most verticals people will prefer your battle tested 3.5 to some brand new 5.
And if 5 does simplify prompts for your use case and there'll still only be two options:
- making a the same thing as what you made with 3.5 is now trivial... in which case you still have the mindshare and the distribution solved to a degree your newly enabled clones don't.
- making a better version of what you made is now trivial... in which case you can just as trivially improve your version and already have the mindshare and distribution solved.
At the end of the day software developers often struggle to fit software into the larger ecosystem it slots into before it becomes something valuable, and to be GPT has been an amazing case study in the fact.
I think the use of the word engineering did half the damage, and I think people confusing the twitter memes with the interesting attempts at prompt engineering (via ReACT, Gorilla, etc) did the other half, but at the end of the day I think a lot of people will be left kicking themselves when "misguided prompt engineers" end up solving real useful problems in ways they didn't think were possible well before we reach the arbitrary goalposts for foundation models that people keep setting up.
- a C developer that writes modem drivers at a telecom company in the 90s about html/perl developers building first interactive websites
Is probably what someone said many decades ago.
Building computers, operating systems and compilers was always a niche.
So if you're a computer scientist or an electrical engineer, sure.
But if you're not, there was, is and will be "power users" who make the most of it.
- Teach, fix shit and bridge the gap for mortals
- Enjoy technology with a deep but not foundational understanding.this is different from LLMs because they don't have explicitly constructed abilities that we can compare across models. so we need to step back and approach every model as brand new and figure out what they're capable of. just because you have a prompt that works amazingly in one model, there's no guarantee that that prompt will continue to work in bigger "better" models.
so rather than working to devise way to trick a particular model into doing your task, it would probably be a better use of your time to learn how to train/modify models to explicitly solve the problem you care about.