> OpenAI conveniently forgot to include this comparison (ARC-AGI-2) in their livestream recital of benchmark progress, which left the livestream looking like marketing rather than science.
Yeah but it was _supposed_ to be marketing, right? Like, of course a product video isn't science in the same way a "hot take" post also isn't science.
> Pricing is good, but profits may continue to be elusive; still no clear technical moat.
It's quite ironic that the technology that is displacing so many people from so many industries, has yet to make a profit. I fear the "creative" part of their destruction will take longer to achieve than they advertise.
Plenty of people in creative fields - especially illustrators - did not get jobs they would have gotten were it not for the fact that their job can now be done by a generative model.
Why is Grok so surprisingly decent? Does lack of mainstream liberal-left censorship (replaced with Musky censorship) result in some sort of a weird performance boost?
Fewer fingers on the scale means the LLM gets to actually do its thing. GPT-4 with zero filtering was scary smart according to the red teams that were testing it. The version the public got had a lobe tied behind it's back.
Having only Grok 3 to compare, and toying around with GPT-5... GPT-5 is pretty good.
There's nothing weird about a model performing better when it is built to more closely relate to reality instead of an ideologically tainted version of such. I don't know how much Musk & Co. interfere with the fine tuning of the models but it is clear that this interference is far less heavy-handed than what the other actors do to their models.
9 comments
[ 3.2 ms ] story [ 37.3 ms ] threadYeah but it was _supposed_ to be marketing, right? Like, of course a product video isn't science in the same way a "hot take" post also isn't science.
It's quite ironic that the technology that is displacing so many people from so many industries, has yet to make a profit. I fear the "creative" part of their destruction will take longer to achieve than they advertise.
Fewer fingers on the scale means the LLM gets to actually do its thing. GPT-4 with zero filtering was scary smart according to the red teams that were testing it. The version the public got had a lobe tied behind it's back.
Having only Grok 3 to compare, and toying around with GPT-5... GPT-5 is pretty good.