[Premium] Why Markets Got DeepSeek All Wrong...

TLDR;

Technology:

🐳 DeepSeek Ships Again.

🤨 But… DeepSeek R1 isn’t as Smart as It Looks

🌚 Moon Data Centers, Closer than Ever?

Product:

🥰 Anthropic Adds Citations

😍 Luma Launches Ray2

Markets:

🦠 $25 Million to Cure Cancer with AI

🥵 Why NVIDIA’s Massive Crash is Stupid

MULTIMODALITY

DeepSeek Ships Again with Janus-Pro-7B

The whale continues to cook. DeepSeek, the hottest name in AI right now, has released yet another model. This time, it’s a multimodal model that beats Stable Diffusion 3 (Medium) and Dall-e3 in two crucial benchmarks for image generation.

TheWhiteBox’s takeaway:

This release is interesting because it indicates that DeepSeek is also interested in multimodality, which is great for the open-source community. More importantly, it shows that the model is any-to-any, an apparently irrelevant point that is actually very relevant.

Most multimodal models today aren’t really multimodal; they are ‘grafted’ models, text models that, through a linear projection, can also ‘understand’ images. In other words, we take pictures and ‘transform them into text tokens.’ In a way, you can see grafting as taking an image and sending the caption that describes the image to the model (it’s not quite like this; it’s just an analogy).

That works, but at the end of the day, the model is still thinking in ‘text land.’

Instead, any-to-any models aren’t text models, they are literally ‘any-data’ models. Of course, ‘any’ carries a lot of weight here, and what they really mean by ‘any’ is that the model is modality-agnostic; it doesn’t care whether you send it images or text and reasons over them in a multimodal manner indistinctively. Put another way, ‘any-to-any’ models don’t care if you send them a photo of a dog or a text describing the photo; they understand both independently of the form factor.

This is true multimodality, and it’s nice to see that we are focusing so much on improving models that can be run on consumer hardware. The whale is once again cooking… could we soon see a stellar multimodal model from them?

FRONTIER MODELS

But… R1 Isn’t as Smart as it Seems.

DeepSeek R1’s results aren’t what they seem.

As shown by this thread, Jenia Jitsev shows how these models are inferior to models like o1 when handling input variations, tests, or robustness. In other words, as the famous Apple reasoning paper showed, most LLMs/LRMs still rely on token memorization. They can only really reason when they’ve seen the pattern enough times.

As a way to test this, they introduced several variations to the prompt, with the required abstracted reasoning untouched, to trick the model (like changing the names of the people in a math problem, which doesn’t change the outcome).

And in these types of tests, DeepSeek r1 is very inferior to o1. And I mean very.

As seen in the image above, R1 is extremely sensitive to prompt variations, while o1-preview holds its own perfectly. This signals that R1 is overfitting the benchmarks, aka memorizing the original, unaffected sequence. When the…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.