Can someone give me the current consensus on Howard Gartner/multiple intelligence vs. g, or general intelligence? Is this even something that people still discuss and research in academia, or did the whole field get tainted by accusations of racism and counter accusations of censorship? Gary's claim here rests on an implicit disregard of general intelligence, which seems counter intuitive to me, but it's been a long time since I've looked into it
I envy the folks who have the energy to speculate this much on an unreleased product/service, and spend this kind of - human - reasoning effort on reflecting blatantly worthless internet posts.
You really don't need to break open the fallacy dictionary to see why those tweets are phony, or to telegraph Astra as just yet another incremental [0] improvement that's even better tuned for math than what came before it. It's the obvious direction and development.
[0] There's a mathematician guy I follow on YouTube who keeps taking LLMs for a spin, and the primary failure mode seems to be persisting. It's not dissimilar to any other field; the models are chatterboxes, and keep going off about stuff that doesn't matter, while quickly jumping over things that do. They're also comparatively slow. According to another mathematician's review of the 250 page paper OAI put out of those 10 breakthroughs, the former persists with Astra.
I wonder if Astra can run on those Cerebras wafers. An order of magnitude faster inference would at least make the iteration process quicker. But then they were announced for Sol too, and they're nowhere to be found. The 2.5x fast mode is nice, but it's a far cry from the 750 tok/sec suggested with Cerebras.
Can someone give me the current consensus on Howard Gartner/multiple intelligence vs. g, or general intelligence? Is this even something that people still discuss and research in academia, or did the whole field get tainted by accusations of racism and counter accusations of censorship? Gary's claim here rests on an implicit disregard of general intelligence, which seems counter intuitive to me, but it's been a long time since I've looked into it
Whatever, as long as it does not burn Codex tokens too fast.
I envy the folks who have the energy to speculate this much on an unreleased product/service, and spend this kind of - human - reasoning effort on reflecting blatantly worthless internet posts.
You really don't need to break open the fallacy dictionary to see why those tweets are phony, or to telegraph Astra as just yet another incremental [0] improvement that's even better tuned for math than what came before it. It's the obvious direction and development.
[0] There's a mathematician guy I follow on YouTube who keeps taking LLMs for a spin, and the primary failure mode seems to be persisting. It's not dissimilar to any other field; the models are chatterboxes, and keep going off about stuff that doesn't matter, while quickly jumping over things that do. They're also comparatively slow. According to another mathematician's review of the 250 page paper OAI put out of those 10 breakthroughs, the former persists with Astra.
I wonder if Astra can run on those Cerebras wafers. An order of magnitude faster inference would at least make the iteration process quicker. But then they were announced for Sol too, and they're nowhere to be found. The 2.5x fast mode is nice, but it's a far cry from the 750 tok/sec suggested with Cerebras.