Fuck AI

To the extent that LLM does stuff, they exist in a more middle ground of "they work... maybe". Problem being that in the ways that matter, "maybe" is really hard to take out of things.

Normally, taking the technology as it comes for what it is is reasonable, but there's a whole investment angle here around expectations. Investment is coming in as if they are expecting an emergent AGI to come out of it. The drive to build everything around executing LLM models and only LLM models is driving the cost of everything in tech up. Money that might have been spent on more general purpose friendly compute is being redirected at Grace Blackwell infrastructure connected by NVL72, which is not particularly interesting for almost all other applications. It's just sucking up RAM and storage and starving everything else.

Problem is that confusion has caused an unreasonable level of investment, with expectations that have no significant indications of being met. So we need credible experts reiterating what those in the thick of it consider 'obvious' to try to fight that irrational behavior.

Currently, they have seen a rapidly evolving tech industry with lots of things that didn't quite work right but were quickly iterated into something useful. The dot-com being an example of generally solid principles attempted too soon, and the tech companies 'fixed' a lot of the problems in short order. So they see this thing that almost seems like a conversational human and assume that 'almost' will be addressed by the tech geniuses before we know it, and whatever naysaying those same experts might be saying is just misplaced humility. The 'experts' may have nerdy reasons why they view it as limited, but the investors "common sense" experience are inconsistent with that feedback.

Of course there's survivorship bias and plenty of 'big tech' that retreated after an optimistic push if you went looking to see that tech can actually fail to close that gap, but none of those examples are anywhere near the scale of the AI push.