This week has been Deepseek all the way in the press (we wrote about it here for the first time in late December), but the releases last month have turned up the innovation and caused a press meltdown (quite a bit of it uninformed). We'll cover that, but other things also happened this week:
- U.S. Copyright Office says AI generated content can be copyrighted — if a human contributes to or edits it. Potentially one of the more important items in the news this week that received almost no coverage. New US guidance on copyright enforcement states that digital work that used AI in creation can be copyrighted as long as there was substantive human involvement. This potentially unleashes a far greater use of generative tools in creative workflows since it was previously unclear if such work could be copyrighted. It's important to note that there might be court challenges to this ruling and also that it does not necessarily absolve image creation tools themselves of violating copyright during training. It seems clear to me that this ruling gets things right in that we've been using "AI" tools for digital creation since the 1980s and before, many of them extremely sophisticated. It would be unclear what "is AI" and what is not.
- Sora rival Pika just dropped a new video AI model, and I can’t believe how good it is. On this note, the progress in video generation has been mindblowing. A year ago, generative AI for video seemed like it could be years away (it's a much harder problem than images), yet... here we are. For short clips, the video quality is now outstanding in models like the new Pika model (Pike 2.1). The models are now being trained to get some grasp of space and physics (with lots of tailored examples), and a lot of work has gone into realism. There's a side by side comparison with Veo 2 (Google's top model) over on the VeoInsights blog that gives some general insight into where we now are in generative video.
- ... and to Deepseek. The Chinese hedge fund has caused a storm by releasing a very well-performing series of models, including one that performs a chain of thought type reasoning (previously only found in the very top-end closed models). The processes they reported also suggested training (and, to some extent, inference) were very cheap compared to previous methods. The stock market reacted, OpenAI is upset some training data may have come from ChatGPT (to which some of the press has reacted with schadenfreude). There is also plenty of criticism that the model has clearly been trained to avoid topics sensitive to China. Yet, Deepseek is being deployed for inference (inc. by companies like Baseten and, most notably, Microsoft). The Deepseek app is also at the top of the Google and Apple app stores (except maybe in Italy). I won't repeat a lot of this coverage, but I'd say the following: 1) even if not all the details are 100% accurate,e there is clearly some excellent research in Deepseek's methods, and the fact that they are fully open means they will be replicated (Hugging face is already working on it), 2) this means likely "better models" for less, 3) it's a big victory for open source, closed models are now on the back foot, 4) it's a big loss for trying to use chip embargos to restrict non-US countries from gaining ground in AI, 5) It's not the "end" of OpenAI, NVIDIA, Google or anyone else building large models, the economics may change, but they will also gain from these techniques so with more compute they will likely be able to build even more powerful models. Seemingly, what is shocking to everybody is that the closed model of innovation has not held out long in the face of worldwide open innovation.
- Mistral Small V3 releases. Not wanting to get left out of the open source action, Mistral has released a strong 24B parameter model (which is very compact) that's been optimized for performance at that scale. The model is Apache 2.0 license and so widely re-usable. It combines a few of my favorite trends: full open source, small enough to run with little compute, and to run fast.
- China to host world's first half-marathon race between humans and robots. In the dystopian bucket this week, a half-marathon race that permits robot racers will be held in Beijing in April. Robot football and other sports competitions have been held for a long time and are great for advancing the state of the art, but this kind of mixed competition seems potentially problematic. At the moment, I'd guess there's little chance that robots will win. A few years in the future, there's probably no doubt that they would. The only question is when the crossover happens. As a demo, perhaps it generates interest and excitement, but if the competition really gets close, I wonder if we'll see foul play from either the humans or the robots, not to mention injuries (robots tend to be very heavy).
Wishing you a wonderful week and some robot free running while it lasts!
