FFFFFFFFFFFFFFFFFFFFF
this Ai doesn't even know it's own company's policies
pull down to refresh
30 sats \ 3 replies \ @optimism 1h
You did 1.5 million requests in a day?!?!! Whoa!!
reply
100 sats \ 2 replies \ @SimpleStacker OP 1h
Yes I'm pretty much a power user at this point. It's for a research project
reply
30 sats \ 1 reply \ @optimism 1h
Just know that if you're using mini, you can probably run something better on a g4dn or g5 on AWS spot, for lots cheaper than the $300 in one day implied above. My current g4dn costs me approx $1.20 an hour or so and if it gets out of hand I just use a smaller one.
reply
100 sats \ 0 replies \ @SimpleStacker OP 1h
Hmm, good point. If I had to redo this I might've started on AWS, but I think I'm too committed to my current setup/workflow now. Good thing is I should only have to do this many requests once or twice to establish my dataset, and more times only if I screwed up or need to change the prompt or something. It's not a live service that I'm running or anything.
reply
30 sats \ 1 reply \ @Luxas 40m
The mistake is using shitty ChatGPT. It's legit been the worse performing model I have used. They seriously dumbed it down post-launch. I used it on day one of release. It's really sad to see how ill-performant the model has become. Laughable at this point.
reply
0 sats \ 0 replies \ @SimpleStacker OP 38m
You're right. I asked Claude the same question and Claude gave me the right answer. Man, GPT-5 sucks.
reply
30 sats \ 4 replies \ @kepford 1h
The models don't even know what they are in my experience. It's been a while since I tested it but maybe 6 months ago I would ask the chatbot what model it was and it could not answer.
Why? They don't know anything. It's how they work. They guess. Most of the newer ones search the web. A key to using them effectively is knowing how they work, their weak points and strong points.
IMO most of the issues people have with using them are more skill issues than critical flaws with the models. If you use them enough you are gonna hit weird stuff. It's wild sometimes. Love reading stuff like this btw.
reply
0 sats \ 3 replies \ @SimpleStacker OP 1h
Exactly, which is why I was so ready to push back. I'm actually a bit surprised that GPT-5 didn't automatically search the web for this question though, from the very start.
reply
0 sats \ 2 replies \ @kepford 57m
These things are dumb. They need to be hammered over and over to do and not do stuff. Even then they still do weird things. It's early and I can't take people seriously that think it's game over because of generative AI.
reply
0 sats \ 1 reply \ @SimpleStacker OP 50m
Seriously. Anyone who's used them for any serious work sees their limitations quite clearly.
On the other hand, the progress in AI in the last 4 years truly has been astonishing, so maybe it's that pace of change projected forward that people are worried about.
And the other thing I'll say is, the AI is better than 90% of undergraduates at knowledge based tasks. Factoring in cost and speed, there's no way I'd hire an undergrad fresh out of college to do things like research stuff for me, or even annotate data for me, now that I have AI in my pocket.
reply
0 sats \ 0 replies \ @kepford 2m
With my work it's better than most junior devs when you consider the cost and hand holding one must do with a new Dev. Basically it can allow someone that knows what they are doing be more productive. In theory at least.
reply