Axios AI+

April 28, 2025
Phew. That was a lot of parties. Thanks to everyone who said hi during my first weekend of White House correspondents' dinner activities. Today's AI+ is 801 words, a 3-minute read.
1 big thing: Microsoft and Apple stumbles could boost OpenAI
Nearly a year after laying out a vision for how generative AI could fundamentally alter the way personal computers operate, both Microsoft and Apple have failed to deliver.
Why it matters: The fumbles leave an opening for an upstart β possibly OpenAI β to deliver a device that harnesses the power of generative AI in a way that makes computing meaningfully better.
Driving the news: After months of delays, Microsoft on Friday finally released Recall, the AI-aided search feature that had been delayed for privacy concerns.
- As part of last week's update, Microsoft also added an improved local search and "click-to-do" feature that both run locally on PCs.
- The features are nice-to-haves, hinting at the potential of on-device AI, but so far none are game-changers.
What they're saying: Microsoft's Windows and devices head Pavan Davuluri promised more to come, but was vague on what to expect β or when.
- "Copilot+ PCs are going to be materially different as we head forward," Davuluri said.
Apple has faced its share of delays as well, with Apple Intelligence bringing modest improvements to Macs, phones and tablets, but failing to deliver a more powerful Siri that would pull together highly personalized information from across email, contacts and other applications.
The big picture: These sputtering starts stand in contrast to the bolder vision both companies laid out last yearΒ β one in which local device-based AI does for personal information what ChatGPT does for broader knowledge.
- In one demo at last year's developer conference, Apple previewed Apple Intelligence answering "When does my mom's flight land?" by drawing upon various pieces of knowledge across the system within multiple apps.
- Microsoft promised a similarly bold AI overhaul when it unveiled the Copilot+ PC last May. "We're entering this new era where computers can not only understand us, but can actually anticipate what we want and our intent," CEO Satya Nadella told reporters.
Between the lines: With Microsoft and Apple thus far failing to deliver on their bigger visions, the most powerful AI app on Macs and PCs remains the ability to summon cloud-based assistants such as ChatGPT, Claude, Copilot or Gemini.
- OpenAI, meanwhile, is moving forward with its own secretive hardware effort. Sam Altman has been vague on just what he and Jony Ive are up to, but has dangled a few hints.
- "If there's something amazing to do, we'll do it," Altman said in 2023, before Ive confirmed some type of device was in the works last year. Altman wouldn't say what the company is building, but did say it won't be a phone.
Yes, but: OpenAI's device is still likely years off, potentially giving Microsoft and Apple some time to get things right.
- "It's a long way away," Altman said of the hardware product, during an on-stage interview with me last September. Altman noted it took OpenAI more than four and a half years to ship its first product, "and I thought that was fast."
2. Llama goes to space
Late last week, engineers on Earth beamed a custom version of Meta's Llama large language model to the International Space Station.
Why it matters: It's a dramatic example of why sometimes you need AI running on your own device rather than in the cloud across an internet connection.
- The model is designed to help astronauts gain quick access to information to maintain and fix their onboard computer without having to look through pages and pages of maintenance documents and manuals β or wait for messages to travel to and from an AI server on Earth.
How it works: Space Llama, based on Llama 3.2, runs locally on the large existing Hewlett Packard Enterprise computer aboard the space station β critical given the intermittent and limited connectivity available to the astronauts.
- Zane Price, the Booz Allen engineer who led the project, told Axios that the system is designed to be more reliable than the out-of-the-box Llama, sticking to the documents it's been trained on and being clear when it doesn't know an answer.
- "We are making sure it's not hallucinating," he told me in an interview Friday at the Washington AI Network's TGAIFriday Lunch.
3. Training data
- The Wall Street Journal said months of testing found Meta's AI chatbots, including those patterned after celebrities, willing to engage in sexual banter, even when users said they were teens.
- Elon Musk's xAI Holdings will reportedly raise nearly $20 billion in funding for the AI/social media company that combines X (formerly Twitter) and xAI. (Bloomberg)
- Hundreds of Google DeepMind workers in the U.K. are considering a move to unionize to help fight against the company's AI defense deals. (Financial Times)
- Prompt engineering is no longer a hot job, as AI models get better at figuring out what users are asking for. (Wall Street Journal)
4. + This
I'm still mesmerized by this buzzer-beating putback during this weekend's Nuggets-Clippers game.
Thanks to Scott Rosenberg and Megan Morrone for editing this newsletter and Matt Piper for copy editing.
Sign up for Axios AI+




