5 ms·
I think the problem is that GPT-4 is not advanced enough yet. They need to train on more parameters, exaflops, and data size in the right proportions and then t
by ftxbro 3y ago
I think the problem is that GPT-4 is not advanced enough yet. They need to train on more parameters, exaflops, and data size in the right proportions and then try again.
- ethanbond 3y agoAlso build more integrations with more APIs. If it can sort of spawn other GPT-4 sessions to improve its working memory and also use their plethora of APIs as well (without human confirmation but on behalf of a human) then I imagine this problem will just solve itself.
- famouswaffles 3y agoIn the end the issue it created did lead to the problem being solved. I don't think "not advanced enough" is really the issue here. I think the main thing is that when you give GPT-4 access to tools and ask it to help with a problem, you are essentially outsourcing cognition. That means the machine possibly taking actions you didn't originally envision.
- londons_explore 3y agoIt's like hiring an employee, and then handing them your username and password to do their work. Smart people/companies will hire an employee, and then give them a new login, so that at least the employee only embarasses themselves.