GPT 3.5

“I would say we got real conviction with GPT-IV. Not even 3.5.”

“In fact, it was originally gonna be called chat with GPT 3.5. And we didn't plan to be product. Didn't think it'd be a huge ship, but did think it would get people The world to like catch up with this and realize something was going on. And we mercifully renam…”

“There's something about wrapping the text in a UI that feels familiar, that feels more human, because you're used to reading, like, you're, it's like the medium is the message, maybe, like, you're seeing this LLM generated text in the Reddit UI, and that feels…”

“GPT-ONE really was Sora-ONE for this modality. And the progress we've made kind of in the last 18 months getting to this 3.5 moment, right, is really compressed compared to how long it took to go from GPT-ONE to 3.5 in the language domain.”

“We were only six weeks into the launch of GPT. 3.5 when we actually had a beta version of fin.”

“The GPT. 3.5 model was in our API for, you know, at least six months prior to that. I think anyone could have built something like this.”

“The only models that could run at real time speed, like we're talking like under two seconds of, you know, to response time were like GPT, 3.5. Or like really shitty text-to-speech models.”

“We would see, like, 70% hallucination rates, like, 3.5 at the time”

“You want to build at the right moment where, like, kind of, like, how we built perplexity around the GPT 3.5 time, not GPT four time. It would have been too late then.”

“The original, I think the original, I don't know, what was it, GPT 3.5 or something, was like a hundred x the cost of GPT-Foro Mini today. In, in the API. So a couple years, you've gone down two orders of magnitude in in cost. For much more intelligence.”

“Qualitatively, what we see is performance on par with GPT 3.5, which is to say, you know, in 2023, when open AI launched GPT, there were there were a bunch of like zero day, really simple jailbreaks and deep seek is essentially susceptible to all of those.”

“It signifies an order of magnitude improvement over the last models, kind of commensurate with the jump from 3.5 to four.”

“So I think, like a lot of people, I was fooled by how quickly after GPT-III.V, GPT-IV came out. It was just, you know, three months or so, but it had been in training for 18 months.”

“You start and validate with the most reasonable And the most intelligent, the frontier model, in our case, GPT-FORO. But once you've validated that the use case is possible, then you try and cost optimize and look for if you can still accomplish that with the …”

“The token context window was more or less, I think, 4000 tokens. So you kind of give it like a blog post to analyze just in the past, you know, week at Google IO Sundar announced a two million token window model or version of Gemini. So think about that. That'…”

“It took two years, for example, for researchers to figure out that GPT-III.V, the predecessor to GPT-IV, had research-grade chemistry knowledge, meaning you could ask it questions about, you know, how to synthesize dangerous chemicals, and it would tell you ho…”

“So mixed trial is as similar performance to GPT, 3.5.”

“Internally We have stronger models that are in between 3.5 and four that are basically second or third, the second or third best model in the world.”

“In order to 10 X or even a hundred X use cases and usages for AI outside of that, there's things that could just be done on existing models today. So you don't need to wait for GPT seven or whatever. You could start with GPT four or GPT 3.5 and add these thing…”

“Straight out of the bat, if you just use Lama tool directly, I don't think you could get like comparable performance, you know, with GPD, 3.5 or four today. But like with fine tuning, if you make that investment in curating your data set in running that fine t…”

“Like if I were a betting woman and I am, I bet there's a three, five level model in the open source ecosystem within a year.”

“I mean, this is kind of, you know, like the kind of the, like the first five minute reaction to GPT, 3.5 is this is a catastrophe for search advertising, which you kind of pull back and think, well, it's already not actually, you know, if I get, I can go to ch…”

“GPT four, if you subscribe, GPT 3.5, if you Don't, which is pretty much the same thing as far as I can tell so far”

“The total size of the software package that runs the model is about a hundred gigabytes.”