China’s DeepSeek throws US tech into a panic

In biblical times, little David’s slingshot toppled the giant Goliath. Last week, a hitherto-unknown Chinese hedge fund released in open source an AI side project known as DeepSeek R1, which is an advanced reasoning model that rivals the performance of OpenAI’s most advanced o1 model. Moreover, so the DeepSeek team claimed, they had done all this with a small team of engineers, in only 2 months, with old slow Nvidia chips because of the US chip embargo, and for only $6 million. Panic gripped Silicon Valley, because US AI leaders such as OpenAI, Google, Anthropic, and Facebook, were spending over $100 million per model in training costs, with armies of engineers, over many months, with state-of-the-art Nvidia chips. After all that, could China actually be ahead? Two days ago, the panic hit Wall Street, with investors panic-selling AI darlings, especially chip giant Nvidia, at the prospect that cheap Chinese AI might cripple the ability of the US giants to make a decent return on the vast amounts of capital invested in them. This turn of events offers a rich basket of lessons for all of us who follow AI news. 1. These are still early days, lots will change before things settle out. 2. When hardware is scarce or expensive, competitors will try to compensate with efficient software design. DeepSeek’s architecture includes a number of clever new features that all together radically decrease the computation needed to train the model. 3. Never count China out. They graduate almost as many engineers each year as all the working engineers in the US.

China’s DeepSeek AI panics Silicon Valley and Wall Street

Anthropic CEO predicts “better-than-human” AI in 2027

Like other CEOs of AI startups, Anthropic CEO Dario Amodei is bullish on the near-term prospects for the invention of human-level artificial intelligence. Here, he takes time out from rubbing shoulders with other rich and powerful people at Davos to go on record that he believes that “we are on track for AI systems that are better than almost all humans at almost all tasks… in the next 2or 3 years.” Including, one supposes, being CEO of an AI startup.

Anthropic CEO predicts “better than human” AI in 2027

Zuck promises $65 bn on AI this year, data center the size of Manhattan

Facebook/Meta CEO Mark Zuckerberg is a major force in AI, even though he has no pressing business need to be one. Facebook and Instagram are minting cash through ads, and AI for now is only a “nice to have” feature. Yet Zuck is all in, spending billions on developing state of the art AI models, nicknamed Llama, that he open-sources and gives away for free. Part of this no doubt is just to scratch the techno-geek’s itch to be in on the coolest new technology. The cold-hearted business case for open sourcing his AI is that it limits the prices that AI vendors can charge for access to their models, weakening them as potential competitors. This is the same rationale for China’s embrace of open source, as in this week’s top story about DeepSeek. In any case, Zuck has recently posted on Facebook (natch) that Meta will be spending $65 billion this year on AI data centers, including one in Louisiana that is almost the size of Manhattan.

Zuck promises $65 bn for AI this year, and a data center the size of Manhattan

Apple taps 37-year company veteran to fix its AI and Siri projects

In the 14 years since the loss of Steve Jobs, successor Tim Cook has made Apple one of the most valuable companies on the planet, valued of over $3 trillion, by reaping the crops that Steve planted (iPhones, iPads, and iMacs.) But now AI calls for a new leap in innovation, and Tim the careful steward may be out of his depth. The future of the company’s flagship product, the iPhone, is clearly as the vehicle for a voice-activated AI assistant, yet Apple’s AI and Siri upgrade projects have floundered. Now Tim has called for a reorg, and is sending in Kim Vorrath, who has worked at Apple since starting as an intern when Sam Altman was in diapers. Maybe it’s just me, but sending in a middle-aged company lifer to shake things up does not reek of bold fresh starts. I would genuinely love to be wrong about this, but with iPhone sales tanking, Apple appears to need more than a bit of freshening up.

Apple taps a 37-year company veteran to fix its floundering AI and Siri projects

X Games adds AI judge for snowboarding events

Last week’s Winter X games in Aspen had a new participant - an AI judge for the Superpipe snowboarding events. The AI judge is still experimental, and its decisions did not affect scoring directly, but the X Games served as a tryout of its abilities and usefulness in a real world event. The system uses multiple high-definition cameras that stream the entire run of each athlete, and analyzes the performance in real time, based on having been trained on countless of hours of video of superpipe events. AI judging is being tried in multiple other sports, including Major League Baseball, soccer, and gymnastics. This particular system was born of a conversation between X Games CEO Jeremy Bloom, a former world champion skier, and Google co-founder Sergey Brin. The system is based on Google’s AI technology. Proving that it’s nice to have smart friends, who are also insanely rich.

X Games adds AI judge for snowboarding events

Bill Gates’ nuclear energy startup inks data center power deal

AI data centers use massive amounts of electricity, and the surging growth of AI threatens to outrun the available electricity supply. Increasingly, data center builders are turning to nuclear energy as a low-emissions source of 24/7 electricity, unaffected by the vagaries of wind or solar power. Now Sabey Data Centers, one of the largest developers of data centers in the US, has signed a deal with TerraPower, a startup funded by Bill Gates that designs and builds next-generation nuclear power plants that are smaller and theoretically safer than previous nuclear technology.

Bill Gates’ nuclear energy startup inks data center power deal

OpenAI releases Operator agent app to high-paying users

Finally! After what seems like untold eons of hype about “AI Agents” - i.e., AI that can actually do stuff, not just talk about it - OpenAI releases Operator, an app that will browse the web and perform simple tasks that you request, like booking a flight. Of course, you have to be one of the moneybags that is forking over $200 a month for the Pro version of ChatGPT to get access, but the pressures of the market on AI vendors are likely to result in a steep and rapid drop in price for agent functions. Google and Anthropic are both working on agent apps, and are not far behind.

OpenAI releases Operator agent app to Pro users

Journalist Rowan Cheung demos some of Operator’s capabilities

UK uses AI tool for “vibe check” on how parliament will react to proposed policies

The UK government is developing a suite of AI tools for civil servants that it calls - and I am not making this up - Humphrey, after the “Yes Minister” character. One part of Humphrey, known as Parlex, is used - and again, I am not making this up - to give ministers a “vibe check” on likely Parliamentary reaction to any proposed policy. Parlex allows bureaucrats to flag topics that may cause them difficulties with parliamentary backbenchers, and identify members of Parliament that may feel passionately about, either for or against, a proposal, based on past voting and speeches. The UK government has made AI a priority for itself and for the nation, and currently has 11 AI projects in development, with a host of others incubating. Most are mundane, with much less of the “Gen Z meets The Manchurian Candidate” vibe of Humphrey/Parlex.

UK uses AI for “vibe check” on how Parliament will react to policies

An updated view of which AI models to use by Wharton’s Ethan Mollick

Ethan Mollick is a professor at UPenn’s Wharton School of Management, and a prominent popularizer of knowledge about AI. (He has written a best-selling book on the subject, Co-Intelligence.) Every few months, Mollick updates an article in his Substack blog in which he reviews the current leading AI models available to the public, and rates their usefulness for various tasks. His latest update was released last week, and it is highly recommended for anyone who wants an overview of what is publicly available (it’s a lot) and how the various tools stack up for various uses. Tl;dr - ChatGPT and its various models are still the Swiss army knife of AI tools - not necessarily the best at anything, but plenty good enough at almost everything.

Which AI To Use Now: An Updated Opinionated Guide

Physiologic data from wearables predict IBD flares 7 weeks in advance

Researchers at Mount Sinai and the Icahn School of Medicine in New York developed machine learning models that used data from commercial fitness wearables (Apple Watch, Fitbit, and Oura ring) to predict flares of Inflammatory Bowel Disease (Crohn’s or Ulcerative Colitis) up to 7 weeks ahead. They found that measures commonly tracked by these devices, such as resting heart rate and heart rate variability correlated with inflammation, and they were able to detect this inflammation building weeks before an overt flare of the underlying disease. Given that 1 in 3 Americans wear such a device, fitness trackers are a mostly-untapped source of health data useful for individuals and medical researchers alike.

Physiologic data from wearables predict IBD flares up to 7 weeks in advance

Open-source AI model may help underserved hospitals

Researchers from a variety of institutions including Harvard, Stanford, and the Mayo Clinic are developing an open-sourced multimodal AI that can interpret medical images, answer questions about them, and generate complex interpretative reports.

One use for such a model is envisioned as serving as an assistant to physicians in underserved, resource-poor, and understaffed hospitals, to help them accurately triage patients who may need specialist care or transfer to a larger regional institution.

Open-source AI model may help underserved hospitals

Nature article