←back to thread

Microsoft and OpenAI's close partnership shows signs of fraying

(www.nytimes.com)
321 points jhunter1016 | 5 comments | 18 Oct 24 11:11 UTC | HN request time: 0.001s | source
Show context
mikeryan ◴[18 Oct 24 12:04 UTC] No.41878605[source]▶
While technical AI and LLMs are not something I’m well versed in. So as I sit on the sidelines and see the current proliferation of AI startups I’m starting to wonder where the moats are outside of access to raw computing power. Open AI seemed to have a massive lead in this space but that lead seems to be shrinking every day.
replies(10): >>41878784 #>>41878809 #>>41878843 #>>41880703 #>>41881606 #>>41882000 #>>41885618 #>>41886010 #>>41886133 #>>41887349 #
Der_Einzige ◴[18 Oct 24 18:15 UTC] No.41882000[source]▶
How can anyone say that the lead is shrinking when no one still has any good competitor to strawberry? Dspy has been out for how long and how many folks have shown better reasoning models than strawberry built with literally anything else? Oh yeah, zero.
replies(2): >>41884918 #>>41887996 #
1. mplewis ◴[19 Oct 24 01:25 UTC] No.41884918[source]▶
Wow, this thing can reason now? How come it keeps getting my basic word problems wrong?
replies(1): >>41885965 #
2. Der_Einzige ◴[19 Oct 24 06:11 UTC] No.41885965[source]▶
Tokenization
replies(1): >>41886060 #
3. youoy ◴[19 Oct 24 06:35 UTC] No.41886060[source]▶
Is it really a tested conclusion? Or a plausible conclusion to try to hide the limitations of the model architecture?

I'm asking because I know that with some prompts it gets the answer correct, and in those cases nothing in the tokenization has changed.

replies(1): >>41887614 #
4. Der_Einzige ◴[19 Oct 24 13:21 UTC] No.41887614{3}[source]▶
Yes, this is 100% tested and proven ad nasum within the field. I have some of my own papers on this, but you can look at literally any major AI conference and find dozens of papers analyzing yet more issues caused by byte pair tokenization.

Honestly the folks who don’t want to admit that it’s tokenization are just extremely salty that AI is actually good right now. Your “AI couldn’t tell me how many Rs in strawberry” stuff is extreme cope for your job prospects evaporating from a system that can’t spell correctly.

replies(1): >>41889402 #
5. youoy ◴[19 Oct 24 18:01 UTC] No.41889402{4}[source]▶
But does a different prompt get the answer correct? I find it surprising. Can you share a link? I'm not saying this out of saltiness, I would be very grateful. If you don't want to I will try the shitty Google search, no problem.