aichannel

While GPT 4.5 is indeed better than 4, it's going to get much harder to notice significant improvements going forward. Had to interact a lot with it to get a feel of its strengths

Reasoning models still best for coding though

🍁 Founder of Frenglish.ai | Software Developer | AI & Web3 ❤️——

was gpt4.5 intended to be better than the reasoning models? i don't think so right

it was just essentially an expensive experiment in scaling compute to train The Next GPT but pretty much ended up being terrible compared to "agentic models" that incorporated chain of thought like deepseek, grok,sonnet 3.7

@operator & /operator | BCI, bioinformatics, search, chaos, decentralized AI

oh i see you're working with translations

i did see a benchmark recently that even tho 4.5 is better the improvement in translation is dramatically lower than the previous launch iterations

i wonder if the local tech companies are better. like i would imagine naver/kakaotalk would be way better at training a korean <-> english model

No it wasn't and yes 100% agree reasoning models are much better for coding and such. But was quite impressed with context understand of 4.5, particularly for translations

The ~15x higher cost for the API is a killer though