President Trump signed a narrowed executive order on artificial intelligence and cybersecurity yesterday.

Why it matters: The new order lets the White House kick the can down the road while it considers new rules for cutting-edge AI models and what to do about AI's advanced cybersecurity capabilities.

• The order is an attempt to shore up the country's cyber defenses as models like Mythos reveal shocking cybersecurity vulnerabilities, but it doesn't compel AI companies to share information about their latest models.

Driving the news: The surprise move comes more than a week after Trump canceled the release of another version of the order with stricter requirements, saying it could have hurt American competitiveness.

What they're saying: "Advanced AI capabilities make our Nation stronger, but also introduce new national security considerations that require coordinated action," the executive order states.  

• "As these capabilities evolve, my Administration will continue to work closely with industry to ensure that the best and most secure technology is deployed rapidly to confront any and all threats to our country."

What's inside: Per the executive order, national security agencies will be required to bolster cybersecurity abilities and create a "cybersecurity clearinghouse."

• Within 60 days, the Treasury Department, the National Security Agency, the Cybersecurity and Infrastructure Security Agency, the National Institute of Standards and Technology and White House officials must "develop and maintain a classified benchmarking process to assess the advanced cyber capabilities of AI models" and decide when a model should be treated as a "covered frontier model."

The intrigue: Former White House AI czar and current adviser David Sacks and National Economic Council deputy director Ryan Baasch pushed for language prohibiting the creation of mandatory government licensing, according to a source familiar.

Behind the scenes: Sacks was able to secure a shorter window for pre-deployment testing — 30 days — and a voluntary framework as some pushed to make it mandatory, according to a source familiar.

Keep reading.

OpenAI CEO Sam Altman said the company's top token user is going through 100 billion tokens per month.

Why it matters: As people inside and out of AI labs are spending more and more on tokens, Altman acknowledges that cost concerns have become a "huge issue."

What they're saying: "The token leader at OpenAI uses about 100 billion tokens a month. To my embarrassment, that's not the token leader in the world. We found someone that used even more," Altman said on a livestream about OpenAI's enterprise adoption.

• The number is staggering when put into context: Over six years ago, the top token user at OpenAI blew through about 100,000 tokens per month, Altman said.
• That's a 1 million-fold increase in token usage.

Zoom out: Altman has talked about the AI costs problem several times this week.

• He said cost concerns are the second-most common issue he hears about from customers behind simplifying AI workflows. Cost never came up before but is now "all of a sudden a huge issue."
• "We want you all to be able to use AI and never worry about it being great and affordable," he added.

Follow the money: Figuring out how to make AI cheaper will be especially important given what Altman said is coming next from OpenAI: "constant running proactive AI."

Keep reading.

Microsoft used its Build developer conference yesterday to debut the company's first internally developed reasoning model, along with a personal AI agent dubbed Scout that is built on top of OpenClaw.

Why it matters: Microsoft is seeking to show it is a serious player in AI beyond its relationship with OpenAI.

Driving the news: The new reasoning model, dubbed MAI-Thinking-1, is a midsized model (35 billion active parameters) and designed to compete more on cost than by rivaling the most powerful frontier models.

• Microsoft also noted that it was not distilled from any other models and is trained only on commercially licensed data, rather than simply the kinds of "publicly available" data typically used to train large language models.
• As for Scout, Microsoft didn't announce the full details of the agent, but said it will work on its own within tools such as Outlook and Teams and can help with tasks such as preparing for meetings.
• That's all on top of the announcement earlier this week that a new class of Windows PCs powered by Nvidia chips is coming later this year.
• Microsoft also announced Project Solara, an Android-based operating system designed to run AI agents on a range of small devices such as earbuds and speakers.

Gopuff, an instant delivery platform, is launching an agentic personal shopping assistant that runs on Elon Musk's SpaceXAI chatbot Grok, turning it into a tool that fills your cart for you.

The big picture: AI labs and retailers are looking to AI-powered commerce as a mutually beneficial revenue driver.

Driving the news: Gopuff is launching "Go," an AI shopping assistant that works via voice and text and can automatically add items to your cart based on its memory and context about your previous purchases.

• Rather than searching for specific items, users can describe a situation, like a game-day party they're hosting or the desire for a healthy breakfast, and the AI agent will assemble a cart automatically based on that prompt.
• "Go" can also predict when shoppers are running low on items like coffee or paper towels, which it says it can pack and deliver from its own warehouses in as little as 15 minutes.
• It also takes into account weather or what's trending in your city, to suggest a "snow day" shopping cart, for example.

Zoom in: Gopuff drivers primarily use their own cars, bikes, e-bikes and mopeds. The new deal does not, unfortunately, allow for delivery with SpaceX rockets.

Keep reading.

• Uber put a $1,500 per month limit on AI coding tools after multiple executives had viral comments about token usage. (Bloomberg)
• Five Anthropic IPO storylines to watch. (Axios)
• Meta is reining in its plans to track employee mouse movements to train its AI models, due to concerns from staff. (Reuters)

ChatGPT hit 1 billion monthly active users in just three years, setting a new record for the fastest app to reach the milestone, per Sensor Tower data.