More seriously, the output quality of LLMs for code is pretty inconsistent. I think there's an analogy to be made with literature. For instance, a short story generated by an LLM can't really hold a candle to the work of a human author.
LLM-generated code can be a good starting point for avoiding tedious aspects of software development, like boilerplate or repetitive tasks. When it works, it saves a lot of time. For example, if I need to generate a bunch of similar functions, an LLM can sometimes act like an ad-hoc code generator, helping to skip the manual labor. I’ve also gotten some helpful suggestions on code style, though mostly for small snippets. It’s especially useful for refreshing things you already know—like quickly recalling "How do I do this with TypeScript?" without needing to search for documentation.
Anyway, literature writers and software engineers aren't going to be replaced anytime soon.
Imagine that current AI is already curating and generating datasets for the next generation.
Also consider that what we have now is only possible because hardware capability increased.
This was pretty much refuted by Meta with their LLama3 release. Two key points I got from a podcast with the lead data person, right after release:
a) Internet data is generally shit anyway. Previous generations of models are used to sift through, classify and clean up the data
b) post-processing (aka finetuning) uses mostly synthetic datasets. Reward models based on human annotators from previous runs were already outperforming said human annotators, so they just went with it.
This also invalidates a lot of the early "model collapse" findings when feeding the model's output to itself. It seems that many of the initial papers were either wrong, used toy models, or otherwise didn't use the proper techniques to avoid model collapse (or, perhaps they wanted to reach it...)
So yes we will see manual labor to finetune the data lair but this will only be necessary for a certain amount of time. And in parallel we also help by just using it: With the feedback we give these systems.
A feedback loop mechanism is fundamental part of AI ecosystems.
On the contrary, human annotation work is stepping up now because we create so many more prompts and want to test them.
There are crucial quality issues with Mechanical Turk, though, and when these really start damaging AI in obvious ways, the system (and the compensation, vetting procedures and oversight) seems likely to change.
The kind of apps that will be built in the next 5 years,are nowhere near what we have today.
Developers will need to update their skillset, though.
I do think barred lawyers will have a role for quite a while, but it is plausible it shrinks to oversight.
I am aware that remote robot surgeries have been a thing for quite a bit of time, but this is the first time ever I am hearing about unassisted robot surgeries being a thing at all.
A follow-up question: if an unassisted robot surgery goes wrong, who is liable? I know we have a similar dilemma with self-driving cars, but I was under the impression that things are way more regulated and strict in the realm of healthcare.
Now there's a chance that this unregulated wild west with a low barrier to entry that's benefited us for so long will come back to bite us in the ass. Kind of spooky to think about.
> making a large barrier to entry (which hurts a lot of us who started as kids/teens)
I doubt it. In my experience autodidacts are the best programmers I know.
I bet if you asked most programmers whether they'd like to have a professional guild similar to the writers who just went on strike, you'd probably be surprised, especially for gaming devs.
I wonder if this is at all related to NCEES re-releasing their controls licensure option?
Luckily ChatGPT and the rest have nothing to do with an AI not to mention AGI.
Unfortunately, I fear we will instead end up sweaty and dirty and bloodied, grappling in the parched terrain, trying to bash members of a neighboring clan with rocks and wooden clubs, while a skyline of crumbling skyscrapers looms in the distance.
- There's the biggest/most high quality training corpus that captures all aspects of dev work (code, changes, discussions about issues, etc.) out there with open source hosting sites like GitHub
- Synthetic data is easy to generate and verify, you can just run unit tests/debugger in a loop until you get it right. Try doing that with contracts or tax statements.
- Little to no regulatory protections.
This is all about labor and capital. When people toss hundreds of billions at something it almost always is.
The social relationship doesn't have to be this way. Technological improvement could help us instead of screw us over. But we'd first have to admit that profit exploitation isn't absolutely the best thing ever and we'll never do that. Soooo here we are.
Maybe once you're deep into APIs that talk to other APIs, but near the surface where the data is collected there's nothing but "grounding to reality".
As my professor of Software Engineering put it: when building a system for counting the number of people inside a room most people would put a turnstile and count the turns. Does this fulfill the requirement? No - people can turn the wheel multiple times, leave through a window, give birth inside the room, etc. Is it good enough? Only your client can say, and only after considering factors like "available technology" and "budget" that have nothing to do with software.
What about developers that code the AI systems? Well.. I am sure AGI will come from other "bootsrapping AIs" just like we see with compilers that compile themselves. When I see Altman and Sutskever talking about AGI being within reach, I feel they are talking about this bootstrapping AI being within reach.