Open Source AI Is Already Dead

0:00 This is a real page, open right now, and it holds the biggest model DeepMind has ever published. 865 GB of weights under an MIT license. No sign-up, no waitlist, no key. Now, scroll that file list to

0:15 the bottom, and something's missing. There's no training data in it. There's no training script, either. Just a 2 KB config file and 64 shards of finished numbers. So, you can run this model. You

0:27 can't rebuild it. Those are very different gifts. For 3 years, we've called that open source. The license says something narrower. Here's how I'll spend the next 12 minutes. First, what's

0:38 actually inside a weights release, then why a lab hands you one at all, and then what changed this summer. Because four things shifted between April and August, and they all lean the same way. Meta

0:49 shipped its best model ever with no weights at all. Alibaba pulled its flagship off the open shelf and put it behind a paid API. Two Chinese labs published enormous models with a dollar

1:01 figure written into the license. And two governments started treating a weights file as something you might need permission to export. Start with the object itself, because most of us have

1:11 only ever used one from the outside. A model is a very large pile of numbers, and those numbers are the weights. Training is the process that sets them, and training needs three ingredients:

1:22 the data, the code that learns from it, and an enormous amount of compute. When a lab publishes, it can hand you any mix of those three. What most of them hand you is the third result on its own.

1:35 The finished numbers, frozen. Think about a cake for a second. Open weights is the cake. Open source, done properly, is the cake plus the recipe, plus the shopping list. Either way, you get to

1:47 eat. Only one version lets you bake another one yourself. The Open Source Initiative has been deciding what open source means since 1998, and it now publishes a page on this with

1:58 a title that doesn't hedge. Open weights, it says, is not quite what you've been told. Its definition of open source AI asks for four freedoms: to use, to

2:09 study, to modify, to share. The one that keeps failing is study. You can't really study a system whose training data you haven't seen. You can poke it, you can measure it. What you can't do is trace a

2:21 bad answer back to the thing that caused it. If a benchmark score leaks into the training set, the weights won't tell you, and the paper won't either. Which brings us to a fairly quiet moment in

2:31 May. The G7 digital ministers stopped arguing about the word and built a ladder instead. Four rungs high. Worked out with the open source initiative over 3 months. At the bottom sits weights

2:44 available. Above it, open weights. Above that, open source AI. And at the very top, open source with open data, where you get the training set, too. And almost every model you've heard of is on

2:56 rung one or rung two. The top rung is close to empty, and the models parked up there, like Elmo from the Allen Institute, aren't the ones the benchmarks are about. So, that's the

3:07 first piece. The recipe was off the table from the start, which makes the interesting question not why they're stopping, but why they ever handed you the cake. A frontier training run in

3:17 2026 costs somewhere between 200 and 500 million dollars. That isn't a generosity budget. There has to be a reason that survives a board meeting. There's one, and it's 24 years old. It comes from an

3:31 essay Joel Spolsky wrote in 2002 titled Strategy Letter 5. He wrote that smart companies try to commoditize their complements. A complement is whatever people have to buy alongside your

3:43 product. If you push its price towards zero, demand for your thing goes up. Google gave Android away and kept every search that ran on it. You might know the shape as razors and blades. Give

3:54 away the handle, sell the refills. In AI, the model is the handle. If a model you can download for nothing is good enough, then no one can charge a premium for an average model, and the money

4:06 slides down to the layer underneath. Compute, storage, the cloud bill. Meta wrote this down without quite meaning to. In April, it shipped Muse Spark, its first frontier model, and its first with

4:18 no weights. 4 months later, it shipped Muse Glimmer, 30 billion parameters, Apache license, free to download. And the model card says exactly what Glimmer is, distilled from Muse Spark. Distilled

4:31 means trained to copy a bigger model's answers, so the free one is a compression of the paid one. That's the whole strategy in a single sentence on Meta's own page. You get the small copy,

4:43 they keep the original. Alibaba's version of the same trade turns up in a financial filing. It's AI cloud revenue grew 45% last quarter, and the AI product line inside it just posted a

4:55 12th straight quarter of triple-digit growth. Qwen is free. The cloud it runs best on isn't. There's a second reason, too, and it has less to do with strategy than with position. Chinese labs didn't

5:07 start out generous. They started out behind. When you're behind, giving the model away is the cheapest distribution there is. First, you skip the sales team, then you skip building out your

5:18 own serving capacity. And after that, the whole world test your work for free. Beijing wrote the same idea into national policy with published targets for how many open models the country

5:29 should have, which sets up the trap the strategy walks into. Distribution is a very good reason to open up while you're catching up. It stops being a reason the moment you've caught up. So, here's the

5:40 question the whole field has been walking around. What happens on the day the free model is as good as the paid one? We started finding out this summer. On the 26th of

5:49 July, Moonshot published Qwen 3 2.8 trillion parameters, and for about 15 lines, the license reads exactly like MIT. Then clause two arrives. If you run a model as a service business

6:03 and your revenue crosses $20 million across any 12 months, you have to sign a separate agreement with Moonshot before you can use it commercially at all. Then 17 days later, Alibaba published

6:15 Qwen 3.8 Max 2.4 trillion parameters. Same shape of license, bigger number. This time it's $50 million, and then you email their business team. That license goes one step further

6:28 because it defines the category it's defending. An AI work assistant, it says, is a product built mainly for AI-assisted coding or office productivity. Then it names two

6:39 examples, and both of them belong to Alibaba. So, the license protects its authors' own products by name. The model's free right up to the moment you start competing with the people who

6:50 wrote the terms. That wasn't the first signal, either. Qwen 3.7, the generation before, shipped without weights entirely. Max in May, Plus in June, both API only from the lab

7:04 that had been the loudest open-source story in the field. Not every lab moved, though. DeepSeek shipped a 1.6 trillion parameter model under plain MIT in August, and it wasn't alone. So, the

7:16 split isn't China against America. It runs between labs that sell a cloud and labs that are buying a reputation. So, that's the commercial half, and it's the quieter half. The louder half is that

7:27 the labs aren't the only ones deciding anymore. On the 2nd of June, the White House signed an order on frontier AI. It creates a category called a covered frontier model, and the benchmark that

7:39 decides membership is classified. And developers may hand the government access up to 30 days before they release. It's voluntary. The order states that in writing and creates no

7:50 licensing requirement. But once a classified list exists and a 30-day window exists, publishing a model has a procedure attached to it for the first time.

7:59 Five weeks later, Reuters reported that China's Ministry of Commerce had been sitting down with Alibaba, ByteDance, and Z.ai to talk about limiting overseas access to their strongest models. Quen

8:12 was named in the room. The shape floated in those meetings was three tiers. A simple filing for basic tools, a security review for stronger ones, and for the most sensitive models, domestic

8:24 use only. Nothing's been decided. There's no draft rule and no date, and Reuters said so plainly. This is a discussion, not an order, and whatever comes out of it may only apply to models

8:35 that don't exist yet. It's also happening while the official politics point the other way. In July, Xi Jinping told a conference that China must encourage open source and openness. His

8:46 Commerce Ministry was working out how to restrict it at the same time. In Europe, the idea's already law. The EU's AI Act hands free and open source models an exemption from some of the paperwork,

8:59 then takes that exemption back above 10 to the 25 floating point operations of training compute. In plain terms, small and open is fine. Big and open needs a conversation. So, we have three

9:11 governments, three different mechanisms, and one shape. Openness is welcome below a line, and the line gets drawn at capability. So, is it dying? This is where I'd push back on the panic because

9:23 the raw supply says the opposite. More open weights shipped in 2026 than in any year before it. Thinking Machines published a 975 billion parameter model under Apache.

9:35 Google shipped Gemma 4. Meta came back with Glimmer. Chinese labs crossed a trillion parameters under MIT more than once. And look at what people actually download,

9:46 because that number's the one that matters. On the model hub last month, Meta's 30 billion parameter Glimmer took 665,000 downloads. Alibaba's 2.4 trillion

9:57 flagship took 52,000. That gap makes sense the second you do the hardware arithmetic. Those 865 GB need seven data center cards just to hold the weights before you served a single token. Very

10:12 few teams have that. Most of us have one GPU and a deadline. So, here's my verdict. Open source AI isn't about to die in the next year. It already died in 2024 on the day the field agreed to call

10:25 a download open source. And the correction hasn't arrived since. What's dying now is frontier openness. And it's dying on a schedule set by revenue thresholds, rather than by a philosophy.

10:36 For most people building something, open weights still win. And it isn't close. The model you can actually run is a 30 billion parameter distillation. It costs nothing to license, and it gets

10:48 downloaded 12 times more than the frontier model. I'd take that trade at twice the price. The part it fails is real, though. If you need to audit a model, reproduce a result, or prove what

10:59 went into it, the thing you were promised hasn't shipped. OLMo is the honorable exception. And it comes from a research institute rather than a competitor. And the thing worth being

11:09 annoyed at isn't a company. It's the word. Open did the work the license wouldn't do for 3 years, and it bought a lot of trust it had not paid for. So, here's what I'd leave you sitting with.

11:21 Every one of these licenses now has a number written into it. 20 million, 50 million, 10 to the 25. And somebody picks those numbers again every year. What's your plan for the year they get

11:33 smaller?