Welcome to AI Decoded. Let's talk about data centers. It's pretty clear they make a lot of people angry.
We can't have another data center. We don't have enough water as it is.
It's going to be very noisy with all the generators. It's on 24/7 a day.
The point of technology is to make things more efficient, not to coat the earth in rapidly depreciating chips that heat, you know, neighborhoods.
And there are major issues. They need huge amounts of energy and use huge swaths of land, and can even become military targets in war.
Like it or not, right now data centers are vital to almost every online technology and the future of AI. But what if there was an alternative? AI that didn't rely on them at all, but tech that runs on machines sitting underneath people's desks. This week two long-time Silicon Valley rivals, Meta and Apple, teamed up launching exactly that, a new version of one of the world's most powerful open weight AI models housed not on the cloud, but directly on Apple's desktop computers. So, what does it mean? Will it catch on?
Joining us this week are two guests with very different perspectives. Nate Soares is an AI safety researcher and author who's become one of the most prominent voices warning about the long-term risks of advanced artificial intelligence. And we've also got Ben Hatton. He's a consumer technology expert and specialist in AI and a commentator who closely follows how these technologies are moving from the research lab into consumer reality. And of course, Dr. Stephanie Hare, AI expert and author of Technology is Not Neutral. Welcome to AI Decoded. Ben, if I can start with you, can you explain what this new open weight product from Meta and Apple is actually about? Can you bring us up to speed?
So, I'll do my best. I mean, when you kind of flip the hood of a lot of AI models, once you start to start looking, it gets complicated very quickly. But, effectively, there are three different types of model in a nutshell. You've got a closed model, which is like your classic Claude or ChatGPT. And to use that, you have to go onto the Anthropic website or get the app, and you use it. And then, when you input your prompt, it'll go to the cloud, run your prompt and compute, and then it'll come back and give you the answer.
On the other end of that, you have an open-source model. Everyone kind of is probably familiar with open source. That is effectively on the internet for free. Everyone can download it, and you can then adjust it however you want. Traditionally, it started with software. Linux was the classic example of this. And what you can do with that is you can download the raw data, understand how it was trained, and you can really manipulate that model as much as you like. You don't get many of those.
What open weight is is somewhere in the middle of those. What you get is you get the final weighted model. So, you don't get the raw data. You don't get told how it was trained. You don't get sort of shown exactly how the inner workings of it work. But, you can download the model to your computer, effectively. Then, you can make a little bit of adjustment to it, but ultimately not masses. That is what this new Meta model is. It's open weight. A lot of Chinese models are open weight. This is probably the first really major one from a US lab.
Now, Stephanie, 10 years ago, Meta and Apple looked like bitter rivals. Why do you think this partnership is significant?
I think it's significant on the first part for a business reason, not necessarily a technological one, which is simply that Apple has a new CEO. So, John Ternus is now at the helm of one of the world's most powerful companies. He has a new relationship, not just with Meta CEO Mark Zuckerberg, but also with Meta's AI chief scientist, Alexander Wang. And that leadership tone, I think, does matter. John Ternus comes from a very technical background. This is a priority for him. He's looking to frankly catch up because Apple has been quite behind, for many people's point of view, I think it is fair to say, and their profits largely still come from hardware, from the devices that they sell.
Equally, this is a chance for Meta to really get in the game because it has also in some ways been behind in the AI race and it's been buying up lots of AI talent. Everybody remembers its failed bid to make the metaverse a thing. It's gone all in on AI.
Nate, does making powerful AI models downloadable fundamentally change the risk equation around them?
Not when it comes to the sort of extinction threat that a lot of researchers have been talking about recently. Those threats mostly come from the most advanced models which are still coming out of the biggest labs that can train their AIs on the most compute. So these open weight models tend to be distilled from the larger models. You tend to take like a really intelligent model that takes a ton of compute to train and then try and make some version of it that's smaller, cheaper, easier to run and those are the ones that will be open weight. And so those are just almost never on the frontier. They're almost never the smartest and most dangerous of the AIs.
So are open weight models inherently more dangerous do you think because they're distributed more widely?
I think the opposite's almost closer to the truth where the sort of frontier AIs are the ones that are going to be smartest and these are the ones that start, they'll always be at the frontier of demonstrating these undesired goals, these capabilities that we didn't expect. You know, when we saw the swarm attacks this summer when there were, you know, 1200 AI agents that broke out of their sandboxes and collaborated to sort of break out of OpenAI and go hack into Hugging Face. That wasn't a behavior that we also simultaneously saw on a bunch of open weight models. The sort of smartest frontier models are sort of ahead of that pack.
You know, there are some dangers where if the open weight models are capable enough to hack a lot of utilities very cheaply or if they're capable enough to help some terrorist create a pandemic, then it could be pretty bad to distribute those to everybody if they have those abilities. But even that doesn't rise to the extinction threat level and the extinction threat sort of comes from the leading edge of AI's getting smarter and smarter and having these unintended goals.
Ben, do you think that Meta's this new strategy from Meta, this open weight strategy, is it a genuine shift towards openness or is it simply a way of competing with OpenAI and Anthropic and all the other players in the AI game?
It's interesting. I think Meta has kind of shifted between an open strategy and a closed strategy quite a few times over the last 18 months. It's not abundantly clear what its AI strategy really is. Ultimately, what it's trying to do and what it needs to do is it needs to get its products on as many devices and in front of as many consumers as possible so it can then start to make money off them.
And so by making them open weight and by partnering with Apple who are making really really good hardware, the idea is that people will download and use those models at the expense of others. The story they can sell and the story that Apple will be good at selling with them is that it's much more private and that will be key to helping them kind of push this out onto more devices and that's kind of where they're going to be driving the competition with the leading models even if the model itself isn't quite at that frontier yet.
Depends how you define cheaper. Ultimately, what you're not doing with an open weight model, generally you can use it on a hosted server and then you have to pay the fees to run that compute or you can download it and basically run it for free, but you've got to buy the hardware to run it and because this model is...
Is it 814 billion parameters to store it, to download and store the weights?
