I’m getting more and more convinced that we will end up running LLMs in our personal computers. Which makes me wonder where Anthropic/OpenAIs moats will come from.
1. in order to run LLMs, especially the best ones, you need complicated devices which are expensive
2. if you buy one for your personal use, you are probably not going to utilize it all the time and it will be idle a lot
It seems to me that it will always be more economical that the LLM-running devices are in a datacenter where it is easier to make sure they are always utilized
Local models today are fine for a lot of mundane tasks and will continue to be so. The use cases where paying for frontier models is worth it, will continue to shrink for folks not doing frontier work.
Or stall. Acceleration has been slowing significantly and gains seem to be tied to huge memory footprints.
Think of it like having a graphics card at home versus using a cloud gaming stream? Technically subscribing to GeForce is much cheaper up front than getting a card, but people still do that. So will the audience of people running agents at home be as large as PC gaming? I think that's kind of plausible.
That is not how LLMs are typically used though in my experience
> Think of it like having a graphics card at home versus using a cloud gaming stream?
Latency seems to be much more important in that use case
2. Eventually we'll get to where local models that don't have sycophancy and slot-machine mechanics trained into them will perform better.
I think consumers are primed for that type of behaviour though. I have an iPhone on my desk. It has something like 2-3tflops CPU+GPU, which is double that of the largest super computer on earth when Jurassic Park came out, and is probably more computing power than existed on earth when I was born in the 80s.
I use this device for around 1hr per day to write text messages.
There is no ceiling to the power of consumer hardware. If it's cheap enough, it will be bought.
SETI@Home is a very niche use case
and web browsing still happens by connecting to data centers and server farms, not by connecting to another laptop
Even two or three years people were pointing out "The ChatGPT subscriptions you can buy with $2000 give you much more compute than whatever home setup you come up with" on r/LocalLLM. I did my own elementary school maths and came to the same conclusion.
Yet till this day people still boast how their beefy M4 Pro/Max machine with 32+GB RAM (which is not at all a "normal person's setup" and costs $2000+) runs LLMs smoothly, and "that's the future".
Someone needs to re-learn basic maths and take a walk around Best Buy to understand what "consumer laptop" looks like.
AI vendors are really going to struggle to shift tokens far beyond the frontier of human capabilities. It's reasonable (not guaranteed) to assume that, if the trend of frontier models (doubling capabilities on benchmarks every n months) holds, then the same trend will hold for local models, and those local models will meet and exceed the perception frontier. This would mean a human cannot tell the difference between Mistral-Open-2030 and Claude Opus 2030.
That's a bunch of "ifs", but there's nothing exceptional about those "ifs". They're basically the scenario if nothing changes between now and ~2030 with regards to capabilities trend attainment.
The price of a mini-PC with Intel Panther Lake is at least double in comparison with the price of a mini-PC with Arrow Lake H having similar specifications, and I am talking about barebones, before adding DRAM and SSDs, whose prices have risen even more.
The rise in prices is somewhat obfuscated by the confusing names of CPUs, i.e. some old and new CPUs may seem to be at similar prices and they have similar names, but the new CPU actually corresponds to a lower segment of the market, by having e.g. a smaller GPU and a lower clock frequency, while the CPU model that really corresponds to the old is named such that it seems to belong to the class corresponding to its present price.
As a concrete example of this obfuscation, which may confuse the buyers of laptops or mini-PCs, I have an ASUS 15 Pro with "Core Ultra 5 225H". If I would buy an ASUS 16 Pro now, the corresponding CPU model, the cheapest which is not worse than what I have, would be "Core Ultra X7 358H".
The whole replacing people angle is just the short term use case the more ghoulish executives are thinking about. In practice, lots of lots of new use cases have been made possible by LLMs. A lot of which can be done locally. But whatever capacity you have locally, they can have more of and for cheaper, and they manage the model instead of you doing it yourself. I think you put it nicely though, their moat will be thinned, and I doubt they'll be as profitable as their funding suggests, but at the same time the demand for them won't go away either. I don't know if OpenAI and Anthropic will be viable, but I'm nearly certain Deepseek is.
The tipping point will be power usage, if a local llm can run the same workload for less power that would be a game changer. Nvidia might get decimated, but even Google and others have moved on from GPUs already, they have faster and more power efficient TPUs. Add to that network bandwidth and availability issues, their moat remains. Also consider that even for graphics capabilities, user devices just don't have a consistent spec to make things like widespread 3d graphics and webgl usage viable. Someone's cheap android phone will never run a local llm reliably,same as it won't a 3d game. even if they have a high-end iphone, network providers aren't always performant as they are in western countries, and then there are people that won't want to install your app or local software, and then browser based exposure of the capability to sites which will have similar hardware spec issues, OS instabilities, competing tabs,etc...
- bulk discounts - cheaper electricity - high utilisation to spread the costs among many users
I don't see how PCs could ever compete against it. Most users AI demands would probably result in >90% idle time on the GPU.