Plain wireguard. Or maybe Pangolin?

https://github.com/fosrl/pangolin

I think they’re a dead-end mostly because of the exponential cost vs. performance. The decreasing returns are obvious, and the companies are trying to adapt by raising token prices, but that will not be enough with the current user numbers (or even double or triple, if we believe the analysts). I think that, at least with these large LLM companies, we’re actually beyond the point of economic equilibrium with this technology, at current energy and water prices.

And yes, training is more expensive than usage. That’s probably the reason why Anthropic suggested a pause in LLM development (training), supposedly because of the fear that AI could become Skynet, but really because they are getting an IPO soon and if people see their current balance numbers, the IPO would fail and the bubble would probably pop. Which really proves my point a little: the economics of these companies “improving their LLMs” (training) don’t make sense at current energy prices.

I think the author is mostly right about the current state of AI, but his future predictions (or worries) are based on a false premise: that the massive LLMs will keep improving in the future.

As far as I have seen the improvements have clearly slowed down, while the energy consumption is rising linearly (or worse). It’s like the energy (money) vs. performance graph is logarithmic, and the companies are expending double the energy to get a 10% improvement. Something like that is not sustainable, and the money seems to indicate so.

I really think that LLMs are a dead-end for AI. A really useful dead-end, once the bubble pops and with time, we get a useful working model for them, probably based mostly on local LLMs, maybe using specialized training data.