Hacker News Reader

Comment by mrweasel | Hacker News Reader

AI is going to be what fiber was to the dotcom bubble. Someone spend a lot of money on a lot of infrastructure, some of which is going to be incredibly useful, but sold for much less than it cost to build. Hardware just depreciates much much faster than fiber networks.

View on HN

I'm not saying that data center buildouts can't overshoot demand but AI and compute is different than fiber buildout. The more compute you have, the smarter the AI. You can use the compute to let the AI think longer (maybe hours/days/weeks) on a solution. You can run multiple AI agents simultaneously and have them work together or check each other's work. You can train and inference better models with more compute.

So there is always use for more compute to solve problems.

Fiber installations can overshoot relatively easily. No matter how much fiber you have installed, that 4k movie isn't going to change. The 3 hours of watch time for consumers isn't going to change.

You can't really use compute more because power is already the bottleneck. Datacenter buildouts are now being measured in GW which tells you everything you need to know. Newer hardware will be a lot more power-efficient but also highly scarce for that reason.

Energy is also being scaled up. But the fundamental difference between compute and fiber buildup is different in my opinion.

Did you pay attention in computer science classes? There are problems you can't simply brute-force. You can throw all the computing power you want at them, but they won't terminate before the heat-death of the universe. An LLM can only output a convolution of its data set. That's its plateau. It can't solve problems, it can only output an existing solution. Compute power can make it faster to narrow down to that existing solution, but it can't make the LLM smarter.

Maybe LLMs can solve novel problems, maybe not. We don't know for sure. It's trending like it can.

There are still plenty of problems that having more tokens would allow them to be solved, and solved faster, better. There is no absolutely no way we've already met AI compute demands for the problems that LLMs can solve today.

Every solution generated by an AI for a novel problem was ultimately rescinded. There is no trend, there is only hope.

There is zero evidence that LLMs can do anything novel without a human in the loop. At most LLM is a hammer. Not exactly useless by any stretch of the imagination, but yes you need a human to swing it.

Not really. You can leverage randomness (and LLMs absolutely do) to generate bespoke solutions and then use known methods to verify them. I'm not saying LLMs are great at this, they are gimped by their inability to "save" what they learn, but we know that any kind of "new idea" is a function of random and deterministic processes mixed together in varying amounts.

Everything is either random, deterministic, or some shade of the two. Human brain "magic" included.

LLMs are considered Turing complete.

Only if you instantiate it once.

If you use it like an agent and stick it in a loop and run it until it achieves a specific outcome it's not.

current shortages are exactly the result of fabs not wanting to commit extra capex due to overbuild risk and inference demand seems to be growing 10x yoy; you've famously got 8 year old TPUs at google at 100% load.

    Hardware just depreciates much much faster than fiber

The manfucaturing capacity expanded to meet the demand for new hardware doesn't (as much)

But if the demand drops for six months, the manufacturers are going to scale back production.

If it drops for a year, they're likely to start shedding capacity, one way or another.

This is not an equivalent situation. The vast, vast majority of what's being produced for this bubble is going to be waste once it pops.

I guess you could at least mine the boards from defunct AI companies for memory chips ? Latest videos from Gamers Nexus showed it is apparently not that hard to transfer memory chips from board to board.

Then you can leach precious metals from the PCB itself.

Yes, but if they weren't overproduced and then either run into the ground or left waiting for demand that would never come, then wouldn't those memory chips and precious metals all still be more available than otherwise...?

I mean, sure, yes, we can try to recycle waste products—but it's still less wasteful not to produce them unnecessarily in the first place.

Agreed, but basic things like sanity seem to have left this hype cycle long ago.

Thankfully real world and physics has the final word, so we can at least plan what to do once the whole thing runs into a wall of reality & make the best of the wreckage.

This goes beyond profits. It will be important for national security.

Reply on news.ycombinator.com