DeepSeek-R1, a new large language model from Chinese AI company DeepSeek, with a website that looks like a sleep-deprived intern pressed "enter" too quickly, has attracted worldwide attention as a cost-effective and open alternative to OpenAI's flagship o1. Released on Jan. 20, whether or not coincidentally on the weekend that "tout Silicon Valley" was eagerly clinging to the coattails of power in Washington, R1 excels thanks to "chain of thought" reasoning which mimics the problem-solving ability of humans.

Unlike closed models such as OpenAI's o1 and Anthropic's Claude, which this week raised another $2 billion from investors who are throwing pudding against the wall in AI hoping that some of it will stick, R1 is open-weight and published under an MIT license. That means anyone is free to build on the architecture. Unlike open source, the source code and training data used to build DeepSeek-R1 are not public, 

The model was developed for just five million dollars through algorithmic efficiency and reinforcement learning, significantly less than o1, despite U.S. export restrictions on advanced GPU chips, especially from Nvidia, on which U.S. competitors are primarily developing. Its affordability, with API costs more than ninety percent lower than o1, thus makes advanced AI more accessible to researchers with limited resources. It also offers a free chatbot interface with Web search capabilities, surpassing OpenAI's current features.

'Everyone is freaking out about DeepSeek'

By matching or even surpassing o1 in some benchmarks, R1 has highlighted China's advance in AI development. Its sudden rise has sparked discussions about the future of open, accessible AI and the need for international cooperation to move forward responsibly. 

International reactions to DeepSeek-R1 ranged from respect to dismay. Nature was analytical: 'DeepSeek-R1 performs reasoning tasks at the same level as OpenAI's o1 - and is open for researchers to examine.' MIT Technology Review remained tidy: 'The AI community is abuzz about DeepSeek R1, a new open-source reasoning model.' But VentureBeat said out loud what all of Silicon Valley was thinking: 'Why everyone in AI is freaking out about DeepSeek.'

By the way, anyone who asks DeepSeek about Tiananmen Square gets the reply, 'I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses.' Asked about the situation of the Uyghurs, a very elaborate answer first appeared that even used the word genocide, but a few seconds later that text was replaced by: "Sorry, that's beyond my current scope. Let's talk about something else.' DeepSeek wants to keep things light and breezy.

Stargate historic project in AI infrastructure

The focus on China's DeepSeek led to great chagrin from the American techno-elite, who wanted to use this very week to underscore American supremacy. OpenAI, Oracle, Japan's SoftBank and Emirates-based MGX are funding the Stargate Project, a $500 billion initiative described as the largest AI infrastructure project in history.

Announced by President Trump in the Oval Office, its goal is to build advanced data centers for AI in the US, which Trump says will create a hundred thousand jobs. They are a kind of Delta works forAI. The project currently has one hundred billion dollars in direct funding, with the remaining investment spread over four years. The first huge data center will be built in Texas.

It already led to bickering over Stargate funding between OpenAI CEO Sam Altman and Elon Musk. Forbes even made a timeline of the ongoing fitties between Altman and Musk, who should get into a boxing ring or a hotel room.

Will the real MGX please rise

In all the excitement, it was especially comical that overexcited investors bought the wrong stock in the belief that it is part of Stargate: biotech company Metagenomi (symbol: MGX) saw its share price shoot up even though it is not involved in Stargate. The MGX that does participate in Stargate, Abu Dhabi's sovereign wealth fund, this MGX, will have watched on in wonder.

It would be some feat if Trump succeeds in getting foreign investors to invest hundreds of billions in U.S. infrastructure with MGX and Japan's Softbank, without the U.S. taxpayer contributing. Investor Bill Gurley (Uber) publicly question ed the public-private partnership, which is unusual by American standards. The main question is whether Stargate will be accessible to all and who ultimately makes the decisions. OpenAI CEO Sam Altman often has problems with governance.

OpenAI with AI agent: Operator

In all the fuss over DeepSeek and Stargate, the news was snowed under that OpenAI this week introduced Operator, an AI agent that can independently navigate Web browsers and perform tasks such as online shopping, booking travel and making reservations. It marks the moment when AI agents are making their entrance into the mass market.

Operator uses OpenAI's Computer-Using Agent (CUA) model, which mimics human interactions with Web sites by using buttons, menus and forms. OpenAI is working with companies such as DoorDash, Uber and eBay for Operator to ensure it complies with their terms of use. 

Despite all its potential, Operator has limitations with more complex tasks such as banking and complex web interfaces or CAPTCHAs. Right now, unfortunately, it is only available to U.S. users on the ChatGPT Pro subscription of two hundred dollars a month, so I have not been able to test it myself.

Operator an echo of General Magic

OpenAI's Operator, nearly thirty-five years after the fact, is very reminiscent of the legendary company General Magic, known for the description as "the most important company to ever come out of Silicon Valley that nobody ever heard of. All of Operator's marketing copy seems to duplicate General Magic's slogans and claims from the early 1990s.

In the end, General Magic, which attempted to create a handheld computer with agent features before the Internet and digital mobile telephony got to mass adoption, proved too far ahead of its time. Like General Magic, Operator strives to integrate seamlessly into users' lives and function as a personal assistant and productivity booster.

For fans: a fine documentary was made about the rise and fall of General Magic, of which this is the trailer. The team behind General Magic was so special that dozens of books have been published and even a real feature film has been made in which they starred: Andy Hertzfeld was a prominent member of the team that developed the Apple Macintosh for Steve Jobs, after General Magic Tony Fadell became the developer of the iPod and co-creator of the iPhone at Apple, and Joanna Hoffman is such a special person that Kate Winslet went to great lengths to play her in Danny Boyle's film about Steve Jobs.  

Leike and Hinton with different warnings

In all the publicity about DeepSeek, Stargate and AI agents, the fact that two leading AI scientists once again warned against the misuse of AI with potentially disastrous consequences for the world snowballed. Professor Geoffrey Hinton, a leading figure in AI and winner of the 2024 Nobel Prize in Physics, discussed the risks of rapid AI developments in a fascinating conversation with his former student Curt Jaimungal. 

Hinton has frequently warned that AI could evolve and gain the motivation to make more of itself and autonomously develop a sub-goal to take control of the world without regard to humans.

The German Jan Leike, co-founder of OpenAI where he left disappointed, now puts it this way:: "Don't try to imprison a monster, build something you can actually trust!" I previously wrote extensively about Leike and Hinton's warnings in this blog post.