The ex-OpenAI CTO has poached John Schulman, the computer scientist and OpenAI founder. Just five months after joining rival Anthropic, Schulman will leap to Murati’s still-unnamed startup, as reported by Fortune’s Sharon Goldman.

Pro tip for all the OpenAI alumni out there: ChatGPT writes one hell of a resignation letter. It wasn’t an easy decision to prompt the AI chatbot to write one, but after reflecting on my personal and professional goals, it was time for me to pursue new challenges. —Andrew Nusca

Want to send thoughts or suggestions to Data Sheet? Drop a line here.

U.S. legislators move to ban DeepSeek from government devices

Two members of the U.S. House of Representatives have proposed legislation to ban the popular Chinese AI service DeepSeek from federal devices.

Reps. Josh Gottheimer, a Democrat from New Jersey, and Darin LaHood, a Republican from Illinois, cited concerns about Beijing’s ability to use the app for surveillance and misinformation as the driver for their “No DeepSeek on Government Devices Act.”

A ban of DeepSeek on federal electronics would be similar to a current policy for the Chinese-owned social media service TikTok (which is also working through the small matter of a nationwide ban that took effect last month).

“We simply can’t risk the CCP infiltrating the devices of our government officials and jeopardizing our national security,” Gottheimer said in a statement.

Earlier this week, the security research firm Feroot said DeepSeek uses code on the web login page of its service that could send user information to China Mobile, a Chinese state-owned telecom company that is banned from operating in the U.S. for having close ties to the PRC’s military. (DeepSeek has previously acknowledged storing data on servers in China.)

In the meantime, DeepSeek remains the most popular app in the Google Play store in the U.S. and a top five app in Apple’s App Store. 

At a time of heightened tension between the U.S. and China, the Chinese company’s R1 AI model shook the foundations of Silicon Valley last month after DeepSeek demonstrated that it was competitive with rival models from U.S. firms like Anthropic, Google, Meta, Microsoft, OpenAI, and others for a fraction of the cost. —AN

Amazon’s profit is suddenly going through the roof

For much of Amazon’s history, investors and competitors knocked the online retailer for its seeming inability to generate profits. As recently as 2014, the tech giant was unprofitable.

But after posting a record-breaking $20 billion in profits during the fourth quarter of 2024 alone, almost double its $10.6 billion from the same quarter the prior year, Amazon is once again looking like the version of itself that many of its rivals should fear. 

With that profit engine revved up and feeding an enormous cash buffer, Amazon CEO Andy Jassy is doubling down in the industry’s AI infrastructure arms race. Jassy told analysts on an earnings call Thursday that Amazon will spend around $100 billion or more on capital expenses in 2025, up at least 20% from 2024. 

The “vast majority” of this year’s capex will go toward supporting Amazon’s broad and deep investments into artificial intelligence infrastructure as the tech giant competes for what Jassy called “probably the biggest technology shift and opportunity since the internet.”

The massive capex comes as Amazon’s cloud and AI rivals up their spending. Google-parent company Alphabet, said on its most recent earnings call that it expects to spend $75 billion on capital expenses in 2025.

This comfortably profitable version of Amazon is powered by strong margins in the various businesses it has built alongside its online retail empire. The fast-growing AWS cloud computing business posted a healthy operating profit margin of 37% in the fourth quarter—or $10.6 billion—on revenue of $28.8 billion. Amazon’s ad business also helped bolster the company’s bottom line.

Keeping a tight lid on costs hasn’t hurt, either. Amazon’s operating expenses grew only 6% in the full year 2024, even as its total revenue increased 11%. —Jason Del Rey

DeepSeek users are playing checkers; DeepSeek replicators are playing chess

Speaking of DeepSeek: The company's innovations have sparked a massive effort by open-source AI researchers to replicate what the Chinese AI upstart achieved, and it’s already paying off.

Lewis Tunstall, a researcher at Hugging Face, told me he expects to see people using DeepSeek’s algorithmic advances to try to create reliable, “reasoning” AI agents in the coming months—and what’s more, he reckons the autonomous task-completing models will be small enough to run locally on people’s computers and phones.

Tunstall and his colleagues are spearheading a new project called Open-R1 that has already produced a script implementing DeepSeek’s novel reinforcement-learning algorithm.

“We’ve already seen lots of very nice examples of people taking this code and then showing that, if you apply it to a whole range of different models, it actually works,” said Tunstall. “You can take a model like [Meta’s open-source] Llama and show and teach it how to do mathematics almost from scratch.” Some people are now even taking models that are small enough to run in a browser and using Open-R1’s script to instill reasoning capabilities into them.

Now the Open-R1 team, and others in the open-source community, are working on creating datasets that AI engineers can use to train new reasoning models with DeepSeek’s techniques.

Bear in mind that it’s all of two weeks since DeepSeek made its big splash by releasing its R1 reasoning model. Open-source AI may be controversial for its inherent safety and security risks, but there’s no denying that the approach can get results fast. —David Meyer

More data

—Meta’s pirated-books-for-AI-training lawsuit deepens. Unsealed emails allege dozens of terabytes of torrented files.

—Apple to revamp iPhone SE. Imminent updates of the lower-cost model to attract users of competing devices.

—Trump tariffs could hurt gaming console sales, Take-Two Interactive CEO says. 

—Sprinklr cuts 500 people. The NYC customer service software company had already laid off about 200 over the previous two years.

—Pinterest records its first billion-dollar revenue quarter. Profitable growth? We love to see it.

—Gemini mulls IPO. The Winklevoss-backed crypto exchange could go public as early as this year.

—Affirm shares jump 15%. Q2 revenue was $866 million, up 47% year over year and handily beating estimates.

—Global tablet shipments grow 9% in 2024. Refresh cycles resume as IT budget restrictions ease.

Endstop triggered