🫂 Not-so-free hugs

While many people who dabble in AI tend to stick to the mainstream chatbots — ChatGPT and Claude known best among them, Gemini has its fans, but perhaps you rock with Grok, nobody’s judging — this particular moment in AI model development is in fact incredibly rich and varied, with hundreds if not thousands of AI labs trying to make their mark with new training data, a unique spin on how their LLMs are constructed, or a distinct point of view in what a model ought to do. 

If you want to obtain one of these models, or the weights used to train them, or one of the adaptations of an existing open-sourced model, you go to Hugging Face.  

• Much like GitHub hosted an ocean of code, Hugging Face has long served as the central clearinghouse of AI model development, footing the extensive hosting bill for models that can become enormous and facilitating the easy transfer of those models. 

• Nvidia just bought it for $12.93 billion. There are plenty of reasons Nvidia might want to do that — starting, and perhaps ending with, heck, they’ve got the money — but there are a few strategic advantages in particular for the chip manufacturer in getting some skin in the game on AI models. 

• Meta was in many ways instrumental for kicking off the open source model boom, as one of the early large companies to release open-weight models. Chinese AI labs in particular — Alibaba’s Qwen models, Z.ai’s GLM models among them — have carried the torch even as Meta has reconsidered precisely how much to release in the open. 

• These open models have been appealing because, first of all, if Anthropic and OpenAI are charging you by the token, for a lot of uses it’s considerably cheaper in the long run for a company to simply buy its own hosting or chips and then run its own models, turning a recurring, pay-as-you-think cost into a (functionally) one-off capex investment.

Hugging Face likely was surveying the world of the hyperscalers of the AI boom and not seeing a surfeit of buyers who would want to keep its offerings intact. Why would an AI lab or a hyperscaler buy a company that specializes in the distribution of free, self-hosted alternatives to the proprietary models that for which they charge $25 per megatoken to access?

The Takeaway

This brings us back to Nvidia. 

• First, Nvidia sells chips, and if you’re a company that’s offloading the simple work to a self-hosted Qwen model you downloaded off of HuggingFace, you’re buying Nvidia’s chips.

• Second, OpenAI and Google have caught on to the fact that chips are lucrative, and are making their own to serve their specific businesses. A world in which, say Gemini or ChatGPT becomes a solitary winner of the AI competition is a bad world for Nvidia, as it becomes a supplier to a functional monopoly, which is a terrible business. Nvidia wants a vibrant ecosystem where there are lots and lots of models that can compete with the big labs, keeping those potential rivals in check. 

• Lastly, and this is long-term thinking, but the hyperscalers probably won’t be hyperscaling indefinitely. Not that a build-out will ever be complete, but the work will turn to maintenance and upgrades. When we built the transcontinental railroads, at a certain point we had enough railroads, and the work became maintenance. What of Nvidia then? Well, with Hugging Face they’ll have skin in the game on whatever comes next.

— Walt Hickey

“Once an oil nation, Ecuador has become a pesco-state” – The Economist 

Read more

A darkweb identity theft service called Nexus was found selling digital scans of 153 million driver’s licenses from people in the United States and Canada, the source of which appears to be an active breach of a major identity verification company. 

• August non-farm payrolls report scheduled for release at 8:30 a.m. ET.