All the major chatbots were able to fill out an NCAA bracket this year, a sharp contrast from a year ago when AI struggled to decipher who would match up in later rounds.

Why it matters: It's an object lesson for a broader AI reality: What is impossible for AI at one point is often a trivial task just months later.

Zoom in: I gave the bots the same task as a year ago — look at a PDF bracket and make picks for each matchup.

• This year I explicitly encouraged the bots to search the web, check for injuries and explore which teams have a history of a Cinderella run.
• OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini all came up with plausible brackets that included a few upset picks, usually with a rationale to justify the choice.

Though they differed on who would win the early rounds, all three picked UConn to win the women's title. On the men's side, Gemini went with Arizona, while ChatGPT and Claude picked Duke.

• Among their upset predictions, Claude sees Akron beating Texas Tech and BYU defeating March Madness powerhouse Gonzaga, noting a key injury. ChatGPT predicted St. John's and VCU to play spoilers, while Gemini picked my alma mater, Miami (Ohio), to beat Tennessee.
• There were some tiny hiccups. For example, Gemini had the wrong potential matchups for the Final Four but was able to correct its picks after a follow-up prompt explaining the error.

This year I entered each bot's picks manually into the bracket challenge at ESPN.com.

• I'm pretty sure at least some of the bots could have entered the brackets themselves, given sufficient time and permissions.
• But I'm hesitant to hand over credentials or agentic powers to my computer.

Zoom out: The last two years of using AI for March Madness reflects my overall experience with AI.

• The most important lesson is to not assume that because the technology can't do something today, it won't be able to in a few months.
• And filling out an entire bracket may not be the best way to use AI, anyway. As my colleague Maxwell Millington points out, why not use the AI to help with research rather than giving it all the fun and glory of making the picks.
• For those so inclined, Yahoo Sports has teamed up with Google's NotebookLM, creating separate notebooks filled with data on the men's and women's tournaments that fans can use to ask all manner of college hoops queries.

Between the lines: Two of the hottest trends in tech are both at play during March Madness: AI and prediction markets.

Keep reading.

Tech companies aren't pulling back from Anthropic even after the Pentagon labeled the AI lab a supply chain risk, with major partners and enterprise customers telling Axios they aren't changing their contracts.

Why it matters: The Trump administration tried to kneecap one of the world's most powerful AI companies. So far, it's just giving it a leg up.

The big picture: Anthropic is raking in more revenue and attention than it did before President Trump and the Pentagon went after the company.

• The designation looked like it could hit Anthropic where it hurts most: enterprise contracts, its core revenue driver. The opposite happened.
• "I've seen enough. Anthropic is the new default for businesses," Ara Kharazian, lead economist at Ramp, said in a post with new data indicating a surge in the share of enterprises choosing Anthropic over OpenAI for their first AI contracts.
• Anthropic's enterprise revenue has continued to grow since the designation was applied, according to Pitchbook. Meanwhile, Claude has surged to the top spot on U.S. app downloads.

What they're saying: "We continue to work really, really closely with them, and we have no plans, in terms of the responsibility I have, to change anything," said Brian Delahunty, VP of engineering at Google.

• Delahunty told Axios that he recently spoke with Anthropic's CTO about continued partnership.
• Multiple Fortune 500 enterprises have told Axios they are not currently making any changes to their Anthropic contracts.
• Tech industry groups and employees are rallying around Anthropic, filing amicus briefs calling for a pause to the designation.

Read more.

The Commerce Department announced yesterday that the Trump administration's effort to export "full-stack" U.S. AI packages will enter its next phase on April 1.

Why it matters: The American AI exports program is meant to bundle the infrastructure, tools and models into ready-to-deploy AI systems for allies and partners, and has been touted as a key part of the White House's AI policy goals.

• But its rollout has been delayed and sometimes marred by confusion.

Driving the news: Starting April 1, industry-led groups will have 90 days to submit proposals for the White House's ambitious new AI exports program.

Keep reading.

• OpenAI is reportedly making a major strategic shift to focus on business customers, a pivot from its consumer-focused approach. (Wall Street Journal)
• Nvidia CEO Jensen Huang said yesterday that he expects the company to reap "at least" $1 trillion in revenue from its newest AI chips through 2027. (Axios)
• There's been endless debate about how AI will impact the labor market. Now, you can see the odds it'll impact your job — and even bet on it. (Axios)

There's a new open-source tool called Arnis that lets people turn any place on earth into a Minecraft world.