AI will always be tied to concentrated capital due to the cost of training state of the art models.

So will CPUs, trains and new medicines. Some sectors simply need a huge amount of capital.

It's unlikely that the market will open up any time soon so that you can get your LLM from someone who isn't a fascist billionaire tech bro.

But this I think is not just likely, but inevitable. As long as at least one company puts out open-source AI models, users will have the option of switching to it. So even if they don't switch, they can use this leverage to keep AI access sellers in check. Eventually, the whole market of selling access to AI will collapse, and the only AI trainers will be people who want to give it out for free (governments, universities) or companies who use it as a loss leader to sell an actual product.

This is mostly true, but you’re glossing over the fact that AI will always be tied to concentrated capital due to the cost of training state of the art models.

I think this is a false assumption.

We don't know what the ultimate state of AI will be, we only know that we're at the beginning of the journey.

Look at other technologies in similar positions. At one point computers were like this. Nobody but the military and the largest companies could afford them and now you can buy children's toys with more storage and processing power than was available to the entire US military at the dawn of computers.

We had no idea what computers would become, even the experts making guesses that sounded wild at the time were off by several orders of magnitude. Bill Gates once said "640K [of RAM] ought to be enough for anybody", modern PCs have around 100,000 times that.

The reason that AI seems like it's only affordable to concentrated capital is because models using a large amount of parameter require a large amount of RAM. The 5TB of RAM estimated for some of the largest frontier models is a lot, about 100x a workstation PC. That's a few orders of magnitude less than 100,000x. Given equal scaling speed, we're closer to running frontier models on our desktop than we are to the dawn of computers.

In addition, the research is leaning towards there being a ceiling on how useful it is to add more parameters (use more RAM). Doubling parameters doesn't result in doubling of capability so there is effectively a soft cap on the amount of parameters.

Hardware is growing near exponentially while the RAM demands of models have a ceiling.

Keep in mind that this only relates to the state of the largest LLMs. You can train generative models on consumer graphics cards. Object recognition, classification, etc are all tiny compared to LLMs. These are also AI and, I'd argue, that they actually produce more value than LLMs (which are failing to deliver on their promise of replacing white collar labor).

It’s unlikely that the market will open up any time soon so that you can get your LLM from someone who isn’t a fascist billionaire tech bro.

You can find open weight models (of all kinds, not just LLMs) freely available here: https://huggingface.co/

Many can be run on locally on gaming hardware. Some of the classification and object recognition models can be run on microcontrollers (a popular use it to use them to make aimbots in FPS games).

There are larger models that can be run on more expensive dedicated hardware (like NVIDIA DGX Spark, ~$5,000) if you want to keep your projects completely internal. Or you can run those open weight models on hardware hosted in datacenters ran by companies you research to be ethical.

You can use AI without using one of the AI companies.

You do realize LLMs are made outside the US as well. I have zero problem with Chinese companies making LLMs that are released in the open and you can even run on your own hardware.

there are already decomp efforts of proprietary games that are accelerated thanks to LLMs, so yes