The thing is, that's not true at all. AI is great for some tasks, and poor for other tasks. That's the reason to break it down like this, because people are trying to explain where AI will and won't revolutionise things, instead of following along with the already-popping AI bubble uncritically
For example: AI's smash translation. They won't ever beat out humans, but as an automated solution? They rock. Natural language processing in general is great. If you want to smush in a large amount of text, and smush out a large amount of other text that's 98% equivalent but in a different structure, that's what AI is good for. Same for audio, or picture manipulation. It works because it has tonnes of training data to match your input against
What AI cannot do, and will never be able to do, is take in a small amount of text (ie a prompt), and generate a large novel output with 100% accuracy. It simply doesn't have the training data to do this. AI excels in tasks where it is given large amounts of context and asked to perform a mechanistic operation, because its a tool which is designed to extract context and perform conversions based on that context due to its large amounts of training data. This is why in this article the author was able to get this to work: they could paste in a bunch of examples of similar mechanical conversions, and ask the AI to repeat the same process. It has trained on these kinds of conversions, so it works reasonably well
Its great at this, because its not a novel problem, and you're giving it its exact high quality use case: take a large amount of text in, and perform some kind of structural conversion on it
Where AI fails is when being asked to invent whole cloth solutions to new problems. This is where its very bad. So for example, if you ask an AI tool to solve your business problem via code, its going to suck. Because unless your business problem is something where there are literally 1000s examples of how to solve it, the AI simply lacks the training data to do what you ask it, it'll make gibberish
It isn't the nature of the power of the AI, its that its inherently good for solving certain kinds of problems, vs other kinds of problems. It can't be solved with more training. The OPs problem is a decent use case for it. Most coding problems aren't. That's not that it isn't useful - people have already been successfully using them for tonnes of stuff - but its important to point out that its only done so well because of the specific nature of the use case
Its become clear that AI requires someone of equivalent skill as the original use case to manage its output if 100% accuracy is required, which means that it can only ever function as an assistant for coders. Again, that's not to say it isn't wildly cool, its just acknowledging what its actually useful for instead of 'waiting to have my mind blown'
The difference is though there isn't a whole lot of "whole cloth novel solutions" being written in software today so much as a "write me this CRUD app to do ABC" which current generations are exceedingly good at.
There are probably 10% of truly novel problems out there, the rest are just already solved problems with slightly different constraints of resources ($), quality (read: reliability) and time. If LLMs get good enough at generating a field of solutions that minimize those three for any given problem, it will naturally tend to change the nature of most software being written today.
I think there's a gap of problems between CRUD and novel. I imagine novel to be very difficult, unsolved problems that would take some of the best in the industry to figure out. CRUD problems are really basic reading/writing data to a database with occasional business logic.
But there's also bespoke problems. They aren't quite novel, yet are complicated and require a lot of inside knowledge on business edge cases that aren't possible to sum up in a word document. Having worked with a lot of companies, I can tell you most businesses literally cannot sum up their requirements, and I'm usually teaching them how their business works. These bespoke problems also have big implications on how the app is deployed and run, which is a whole different thing.
Then you have LLMs, which seem allergic to requirements. If you tell an LLM "make this app, but don't do these 4 things," it's very different from saying "don't do these 12 things." It's more likely to hallucinate, and when you tell it to please remember requirement #3, it forgets requirement #7.
Well, my job is doing things with lots of restraints. And until I can get AI to read those things without hallucinating, it won't be helpful to me.
You need to substitute "AI" with "LLMs" or "current transformer architecture" or something. AI means something completely new every few years so speaking of what AI can't do or can never do doesn't make any sense.
I just wrote up a very similar comment. It’s really nice to see that there are other people who understand the limits of LLM in this hype cycle.
Like all the people surprised by Deepseek when it has been clear for the last 2 years there is no moat in foundation models and all the value is in 1) high quality data that becomes more valuable as the internet fills with AI junk 2) building the UX on top that will make specific tasks faster.
For example: AI's smash translation. They won't ever beat out humans, but as an automated solution? They rock. Natural language processing in general is great. If you want to smush in a large amount of text, and smush out a large amount of other text that's 98% equivalent but in a different structure, that's what AI is good for. Same for audio, or picture manipulation. It works because it has tonnes of training data to match your input against
What AI cannot do, and will never be able to do, is take in a small amount of text (ie a prompt), and generate a large novel output with 100% accuracy. It simply doesn't have the training data to do this. AI excels in tasks where it is given large amounts of context and asked to perform a mechanistic operation, because its a tool which is designed to extract context and perform conversions based on that context due to its large amounts of training data. This is why in this article the author was able to get this to work: they could paste in a bunch of examples of similar mechanical conversions, and ask the AI to repeat the same process. It has trained on these kinds of conversions, so it works reasonably well
Its great at this, because its not a novel problem, and you're giving it its exact high quality use case: take a large amount of text in, and perform some kind of structural conversion on it
Where AI fails is when being asked to invent whole cloth solutions to new problems. This is where its very bad. So for example, if you ask an AI tool to solve your business problem via code, its going to suck. Because unless your business problem is something where there are literally 1000s examples of how to solve it, the AI simply lacks the training data to do what you ask it, it'll make gibberish
It isn't the nature of the power of the AI, its that its inherently good for solving certain kinds of problems, vs other kinds of problems. It can't be solved with more training. The OPs problem is a decent use case for it. Most coding problems aren't. That's not that it isn't useful - people have already been successfully using them for tonnes of stuff - but its important to point out that its only done so well because of the specific nature of the use case
Its become clear that AI requires someone of equivalent skill as the original use case to manage its output if 100% accuracy is required, which means that it can only ever function as an assistant for coders. Again, that's not to say it isn't wildly cool, its just acknowledging what its actually useful for instead of 'waiting to have my mind blown'