Thoughts on "Open" Models

Everyone remembers Steve Jobs’ 'Thoughts on Flash' open letter, written in 2010, because it absolutely eviscerated Adobe and effectively started the long, slow death of one of their key technologies at the time. But a few years before that, Steve Jobs had other “thoughts” — on music, and digital rights management specifically.

The result of 'Thoughts on Music' was a much faster outcome.¹ Within two years, DRM vanished from iTunes and the industry at large. It was a masterstroke by Jobs with Apple facing growing pressure — especially in Europe (sound familiar?) – to open up iTunes. But with his essay, Jobs shifted the blame and focus to the record labels (many of which were European-controlled!), knowing it was unlikely to dent the all-important (at the time) iPod sales.

Anyway, this is the framing that pops into my head when thinking about the open letter from many of the biggest players in tech calling for 'open weight' models to underpin America’s AI leadership. The situation is far different — obviously, given the geopolitics at play here if nothing else — but this feels a bit like the same type of rallying cry that’s going to be hard for the industry to ignore.

In fact, two of the three big players in AI that weren’t initial signatories, OpenAI and Google, now are both seemingly on board. Both would point to the fact that they offer “open” models already — in fact, they’re both pointing to that very fact. Still, it’s impossible to ignore their initial absence. Either the other players kept them out of the loop to send a message, or they kept themselves out of the loop to try to avoid the outcome.

There are a few other elephants that seem absent from this particular room — notably, Amazon — but the big one is obviously Anthropic.

While it’s easy to see everyone just getting in line and saying, “yeah of course the US should be leading in open models”, it’s hard to see Anthropic saying the same. Because from Dario Amodei on down, they’ve railed against the notion in the past. Why? Security, of course.

Critics will say it’s a fig leaf to cover their obvious conflict as the de-facto leader in closed AI models at the moment. But it’s still a pretty decent sized fig leaf! Each day seemingly brings new major security risks uncovered (or created) by AI.

Of course, others would use the “offense is the best defense” argument here and they might be right too, if for nothing else than it feels like this genie may already be out of this bottle. So perhaps that’s the notion that causes Anthropic to bend here, if they do. But they’re undoubtedly going to hold out as long as they can.

And they now find themselves with an unlikely ally in this regard: the US government!

Yes, their old nemesis is now the one that seems most aligned with locking down the models rather than going the other way. They have their own reasons, of course. Security is still one of them, but geopolitics and negotiating leverage loom large as well.

Everyone is conflicted here. NVIDIA has been leading the US open weight charge because they undoubtedly believe it will help remove the political pressure on their hardware. Is this the path to fully unlocking the Chinese market? Probably not, but if the US fully locks down US models, NVIDIA is unlikely to ever grow their business there. It is worth pointing out that NVIDIA is not open sourcing CUDA, even though that’s exactly what Alibaba is doing in China with their would-be competitor.

Microsoft now seems all-in on the idea of being a “Switzerland” for models, and Satya Nadella clearly doesn’t want to see OpenAI and Anthropic simply run away with the market — which is more than mildly awkward given the ownership stake Microsoft has in each!

Speaking of awkward, how weird is it that no sooner does Meta spend hundreds of billions of dollars to shift their focus away from open weight models, do such models become ground zero for this debate? They’re still signed on here, but their actual actions are mainly in the closed camp for now. Was this another big mistake for Meta?

Perhaps a key thing worth pointing out: the letter doesn’t suggest open weights is the only path forward, simply that the US shouldn’t ban them or even just dissuade the building and use of them. In this world, the closed models still have their place, but the question left unanswered is what that place is relative to the open models and vice versa.

In some ways, this has been the debate in AI from the get-go. Again, Meta bet on open before they switched to closed. So did Alibaba in China, before they just shifted back to open in light of President Xi Jinping’s outlining of China’s stance. That stance is obviously that stance because they’re not at the forefront of the frontier at the moment, otherwise it would probably be a very different stance! So yeah, conflicts all around.

Still, the shift towards open has a groundswell, at least right now. But the main reason is decidedly capitalist: cost.

The price of frontier AI keeps growing more untenable — both to use and to build, as Google can attest this week with the hit to their stock price on the news of yet another jack up in CapEx. Meanwhile, everyone from the biggest enterprises like Microsoft to the smallest startups are feeling the token burn. If someone can come up with the “good enough” open model… unfortunately, it seems like it was a Chinese company, Moonshot, that just did it with Kimi K3.

But even that is not so straightforward! Because these models are only open weight and not fully open source, what exactly goes into making them, and what they’re going to output for certain queries is… largely unknown. And potentially problematic!

Further, there’s the fundamental question of just how much they relied on distillation from the closed frontier models. And if those go away, you have a real chicken-and-egg problem.

This is all angling towards a world where frontier models remain closed but are perhaps used to distill open variants that are closer to the cutting edge after some set period of time. This is essentially what OpenAI and Google have been doing, but if it’s more formalized, everyone might feel better. Especially if those models can then legally be used to distill others – at least in the US. It would keep some power in the hands of the frontier, while trickling down more flexibility with some regularity.

But who knows, that’s just a guess this week. After one hell of a week of news.

Still, the vibes right now are clear. Everyone seems to be falling in line quickly behind this notion of not only protecting "open" models, but pushing for the US to combat China to take the lead in their build and spread. Well, everyone except Anthropic.² And their strange bedfellow here, the US government. And that obviously matters – especially when the stakes are higher than, say, DRM.³

¹ It's really weird/sad/annoying that Apple no longer hosts these pivotal posts on their site. You'd think they would embrace their history and importance?! ↩

² And it's still not entirely clear that OpenAI is fully on board with this given that their apparent actions behind the scenes suggest otherwise! ↩

³ Man, does the industry miss Steve Jobs' voice and gravitas right now... ↩