Remember when they told us in CS class that it's better to design more efficient algorithms than to buy a faster CPU? Well here we are building nuclear reactors to run our brute force "scaled" LLMs. Really, really dumb.
> Remember when they told us in CS class that it's better to design more efficient algorithms than to buy a faster CPU?
No? The tradeoff is entirely one between the value of labour versus the value of industry. If dev hours are cheap and CPUs expensive. If it’s the other way, which it is in AI, you buy more CPUs and GPUs.
This makes sense if and only if you entirely ignore all secondary and tertiary effects of your choices.
Things like massively increased energy cost, strain on the grid, depriving local citizens of resources for your datacenter, and let's not forget ewaste, pollution from higher energy use, pollution caused by manufacturing more and more chips, pollution and cost of shipping more and more chips across the planet.
> Things like massively increased energy cost, strain on the grid
This is a peculiarly USA-localized problem. For a large number of reasons, datacenters are going up all over the world now, and proportionally more of them are outside the US than has been the case historically. And a lot of these places have easier access to cheaper, cleaner power with modernized grids capable of handling it.
> pollution from higher energy use
Somewhat coincidentally as well, energy costs in China and the EU are projected to go down significantly over the the next 10 years due to solar and renewables, where it's not so clear that's going to happen in the US.
As for the rest of the arguments around chip manufacturing and shipping and everything else, well, what do you expect? That we would just stop making chips? We only stopped using horses for transportation when we invented cars. I don't yet see what's going to replace our need for computing.
Both chips and developer time are expensive. Massively so, both in direct cost and secondary and tertiary elements. (If you think hiring more developers to optimise code has no knock-on effects, I have a bridge to sell you.)
There isn't an iron law about developer time being less valuable than chips. When chip progress stagnates, we tend towards optimising. When the developer pipeline is constrained, e.g. when a new frontier opens, we tend towards favouring exploration over optimisation.
If a CS programme is teaching someone to always try to optimise an algorithm versus consider whether hardware might be the limitation, it's not a very good one. In this case, when it comes to AI, there is massive investment going into trying to find more efficient training and inference algorithms. Research which, ironically enough, generally requires access to energy.
And you would have been mocked by your peers for being so concieted that you'd dare to look down on other people for not inventing an algorithm that doesn't exist and for which there's no evidence it's even possible for one to exist.
There's a ton of research into more efficient AI algorithms. We've also seen that GPT-5 has better performance despite being no bigger than previous models. GPU/ASIC vendors are also increasing energy efficiency every generation. More datacenters will be needed despite these improvements because we're probably only using 1% of the potential of AI so far.
You see this in other sectors where demand outstrips improvements in economy. Individual planes use substantially less fuel than they did 50 years ago, because there are now fewer engines on planes and the remaining engines are also more efficient; but the growth in air travel has substantially outpaced that.
You have a point. But it doesn't make sense to seek for the next unrealized breakthrough (low energy, brain-comparable power consumption) AI leap yet, when existing products are already so transformative.
There's no efficient algorithm for simulating a human brain, and you certainly haven't invented one so you've got absolutely no excuse to act smug about it. LLMs are already within an order of magnitude of the energy efficiency as the human brain, it's probably not possible to make them much more efficient algorithmically.
Your brain has a TDP of 15W while frontier LLMs require on the order of megawatts. That's 5-6 orders of magnitude difference, despite our semiconductors having a lithographic feature size that's also orders of magnitude smaller than our biological neurons. You should do some more research.
>Your brain has a TDP of 15W while frontier LLMs require on the order of megawatts.
You should do some basic maths; the megawatts are used for serving many LLM instances at once. The correct comparison is the cost of just a single LLM instance.
Yes, the cost figures published by LLM labs imply a power consumption measured in megawatts for each instance of top performance frontier models. Take the L.
