Pinned Tweet
What's described in this thread is an incredibly uneconomical way to automate processes.
What is referred to as "multi agent systems and semantic layers" is the software doing 100x the work to get a marginally more reliable output, which introduces very un-software like variable costs to achieve the automation (on top of the base models already having extraordinary variable costs vs typical (deterministic) software).
This is primarily why token demand has exploded. Not from increased end-user demand, but from the systems requiring all of these extra layers and iterative processes just to disguise the probabilistic output in a deterministic-looking costume.
And hallucinations never go away, they just get buried under all the filters. There will always be edge cases, so even in domain specific silos humans will still be in the loop and you will never achieve actual automation.
On the cost side... using neural nets to automate is like Microsoft having to spend money every time a user enters data into a cell in excel. There are enormous variable costs to service these agents, and they're going much higher because they need to get much better to actually be useful.
Right now, venture capital and recycled megacap tech cash flows are allowing these products to be developed and delivered at a massive loss. And still, roughly 97% of ChatGPT and Copilot users alike are willing to pay exactly zero.
Heavily subsidized, and users still have little interest. The products are very very far away from where they need to be to change this. It may as well be considered science fiction. And in the event they do actually get there one day, they'll be copied by everyone else.
This is all a galactic misallocation of capital and brain power. There is no there there; another very long AI winter is coming.
