So it seems like DeepSeek might have totally revolutionized how AI models are worked, and to be honest, it's not looking good for OpenAI...

Researchers at Stanford and UW (go huskies!) managed to create a model that rivals o1 preview! o1 preview is the cutting edge chain-of-thought (thinking) model that OpenAI put out a few months back. They did this using the same technique DeepSeek did, starting out with Alibaba's open source Qwen model, and using a carefully curated sample of just 1000 prompts from Google's latest thinking model!

If you remember last week's drop, this model just came out alongside some other new tech, and it has put Google firmly back in the AI race due to its speed and low cost. Speaking of low cost, since these responses from Gemini were free, and due to the immense efficiency of this technique, the total cost to train this model was under $50. They used just 30 minutes of compute time on 16 H100's (Nvidia's cutting edge AI superchip) to create something that rivals the best of what OpenAI can offer.

So what does this mean for the AI sector as a whole? Well... This technique cannot really be used to push forward the bleeding edge of AI performance, unfortunately, due to its reliance on output from another model, it is almost impossible to improve beyond the original source. What this does mean is that anyone who builds a future cutting edge model will likely be unable to charge exorbitant prices, due to the fact that the open source community would likely be able to recreate it for a fraction of the cost.

OpenAI:

These guys are tired of paying Nvidia billions to build their chips, so they've decided to take a crack at designing their own! While Microsoft and Amazon have both been working on this for a while, Google has actually been using their custom Tensor chips for a while now, which have allowed them to slash training costs and has significantly contributed to the speed at which Gemini is able to respond to messages.

Additionally, as OpenAI continues its quest to become a for profit company, abandoning its original mission and values, Elon has been continuing to fight back against the transition. While Musk was an early backer of OpenAI, he left the organization a short while ago, and is now upset that they're restructuring without him. This move will likely make Sam Altman extremely wealthy, due to the high share of ownership he is scheduled to receive. Musk has been attempting to slow this process down as much as possible, recently offering to buy OpenAI for $100 Billion. OpenAI has an extremely complicated structure with a non-profit controlling a for profit organization, but Sam Altman and the board have flatly refused any talk of a purchase.

Thomson Reuters (of Reuters fame) has won the lawsuit he filed against Ross Intelligence! This lawsuit was filed way back in 2020, before ChatGPT was even released. Basically Ross scraped Reuters' news sites for training data to feed to their AI systems, and then attempted to claim fair use. The judge has ruled in favor of Reuters, imposing a fine on Ross Intelligence, but, while this lawsuit is significant, it is important to realize that the AI Ross Intelligence created was not generative, so how this result will be applied to ongoing lawsuits directed at OpenAI is uncertain.

DeepSeek:

While the underlying model remains open source and offline only, it has recently been revealed that there are many security issues with the iOS app for DeepSeek, to the point where I would recommend against using it. The app lacks basic security measures, and has a tendency to send user data unencrypted to DeepSeek's servers, which causes a whole host of privacy concerns. This is likely due to the fact the DeepSeek R1 was mostly a side project by a VC firm who may not have a ton of experience with software.

Enjoy your break everyone!