The US and China's open-source AI showdown is coming

So let's start with a quick catch-up on how we got here.

The current wave of AI, generative AI, large language models, all that stuff, comes out of Silicon Valley.

And so for the last few years, generally speaking, people have had the idea that if you're talking about cutting-edge AI, you're talking about the U.S., Originally, there was an idea of making it an open technology that everyone could share around, but the economics of it, the vast amounts of investment that were needed to buy all the computers to train these AI models – in other words, to create them – That level of investment meant, no, investors didn't want to see something that could be freely shared around.

They wanted to be something that was closed source, that was tightly controlled, and that they could make a whole bunch of money out of.

Then, at the beginning of last year, China threw a spanner in the works when a company called DeepSeek, who hardly anybody had heard of at the time, released a model called R1.

And this model was capable.

It wasn't cutting edge.

It wasn't at the level of what you'd be getting out of OpenAI or Anthropic at the time.

It was a generation or two behind that.

But it was genuinely useful in a variety of use cases, and this astounded the AI industry and the financial markets for a few reasons.

The first of which was that DeepSeek was not an American company, so how dare they? Secondly, DeepSeek had seemingly trained this model at a much lower cost than what Anthropic and OpenAI were spending to train their cutting-edge models.

And even though we're talking about a less capable model, this had huge implications.

It called into question the vast amounts of money that these American firms were spending on training their models.

Was that really necessary when you could get something that's good enough for many use cases at a much lower cost? The third aspect to this, which is not unconnected with the second, is that DeepSeek made great use of a technique, quite new at the time, called distillation, which is when you train a small AI model on the output of a bigger one.

thereby transferring most of the smarts from the big model into the small one.

And this has many implications, not least of which is the fact that if you can have a capable and small enough AI model that you can actually install on your smartphone or your laptop, then when you need to call on your AI virtual assistant, you don't need to have a subscription.

You don't need to send your questions and your sensitive information off to some service that's hosted in somebody else's data center.

You don't need models that are of that size, that can only run on the giant computers they have in those data centers.

You've got something running locally that you can trust, that you can rely on, that's not sending your data anywhere, and that's good enough for certain tasks.

So when DeepSeek released R1, they also released small versions that they had distilled from R1 in this way.

But it also subsequently emerged, at least if you believe what OpenAI claims, that DeepSeek trained its large models on the output of ChatGPT as well.

And we've seen a lot of accusations in the last couple of months by American AI companies against Chinese AI companies going, hey, these people are stealing our AI.

by distilling it.

That's against our terms and conditions.

As we've discussed recently on this channel, well, that's against your terms and conditions, but there's nothing you can really do to stop it.