cross-posted from: https://lemmy.ml/post/2811405
"We view this moment of hype around generative AI as dangerous. There is a pack mentality in rushing to invest in these tools, while overlooking the fact that they threaten workers and impact consumers by creating lesser quality products and allowing more erroneous outputs. For example, earlier this year America’s National Eating Disorders Association fired helpline workers and attempted to replace them with a chatbot. The bot was then shut down after its responses actively encouraged disordered eating behaviors. "
We’ve been trying to do that approach for decades and progress has been slow and disappointing.
When we finally decided “screw it, just build a giant black box and throw terabytes of text at it to see what happens” we got GPT3 and now the world is about to be revolutionized.
Start revolutionizing, we’ve been waiting for months now…
Gosh, months.
If it’s supposed to be the labor extinguisher of the future, yes I expect something in the order of months
Your expectations are unrealistic. I am a programmer and I find tools like ChatGPT and copilot to be fantastic, but the company I work for has banned use of them until the legal department has figured out what the heck (and they won’t figure out what the heck until the judicial system figures out what the heck, and the legislative layer above that). It takes time for these sorts of massive shifts in well-established systems to happen.
I am too and it can write boilerplate. It can’t do anything at a systems level, and I can’t even trust it to write something that can handle edge cases. I still have to do all the real work, it just writes the boilerplate, which is something I almost never do anyway. The legal side of it is almost exclusively IP rights, and I can’t risk putting GPL3 code in my project, and I certainly can’t risk putting IP in that it will regurgitate somewhere else
The black box isn’t being done because it’s a new idea, it’s actually the other way around. The newer idea is actually the method for easier analysis. There’s a few reasons that they aren’t doing that though.
If doing it the “wrong way” is cheap and works well, then perhaps it’s not the “wrong way.”
There are many companies (and researchers and hobbyists now) who are doing this stuff other than OpenAI, at this point. They just broke the ice and showed what was possible.
I just explained that it’s not cheap. It costs far more to buy a cheap car and do constant maintenance than it is to buy the mid tier car without much maintenance. That’s what’s happening with AI right now, we’re buying the cheap car and paying for it in labor and development costs. I’m saying that the right way is to buy the more expensive one, which will be cheaper in the long run.
There is no agent on the planet who is intentionally choosing to make their models harder to analyze. This is a ridiculous idea that you could only believe if you didn’t understand where the complexity comes from in the first place. Creating ML models that can be efficiently and effectively trained and interpreted is an extremely hard and unsolved problem, and whomever could solve it would be rolling in cash.