Your assumption of the intention behind GPT-5 seems wrong, and thus your conclusion on future AI development based on the scaling hypothesis too. GPT-5 was not actually an attempt at making a new more powerful flagship model but making a much cheaper and more reliable model of similar performance, and that was achieved. OpenAI just made the wildly misleading choice to release it under this name, which had previously been associated with order of magnitude increases in training compute - but which wasn't the case for this release. https://open.substack.com/pub/transformernews/p/gpt-5-underwhelming-launch-pace-ai-development
It may be true that GPT-5 was not trained to be a powerful new flagship model, but I think they called it GPT-5 and initially hyped it like one because after two years they weren’t getting an exciting new qualitative advance with scale. If they had they would have released that model as GPT-5 instead. Maybe they will eventually, but it’s not like they’re not trying.
Passing on to the new college student:)
Great! Reading my substack is what the experience of going to college is all about.
Garbage in... garbage out. The tough part IMHO is who is deciding who defines garbage. For example, in the case you describe of Grok.
Training an LLM on Twitter and on Musk’s tweets in particular really doesn’t seem like a recipe for producing great content!
Your assumption of the intention behind GPT-5 seems wrong, and thus your conclusion on future AI development based on the scaling hypothesis too. GPT-5 was not actually an attempt at making a new more powerful flagship model but making a much cheaper and more reliable model of similar performance, and that was achieved. OpenAI just made the wildly misleading choice to release it under this name, which had previously been associated with order of magnitude increases in training compute - but which wasn't the case for this release. https://open.substack.com/pub/transformernews/p/gpt-5-underwhelming-launch-pace-ai-development
It may be true that GPT-5 was not trained to be a powerful new flagship model, but I think they called it GPT-5 and initially hyped it like one because after two years they weren’t getting an exciting new qualitative advance with scale. If they had they would have released that model as GPT-5 instead. Maybe they will eventually, but it’s not like they’re not trying.