Meta held talks to buy Thinking Machines, Perplexity, and Safe Superintelligence

At this point, it’s becoming easier to say which AI startups Mark Zuckerberg hasn’t looked at acquiring.

In addition to Ilya Sutskever’s Safe Superintelligence (SSI), sources tell me the Meta CEO recently discussed buying ex-OpenAI CTO Mira Murati’s Thinking Machines Lab and Perplexity, the AI-native Google rival. None of these talks progressed to the formal offer stage for various reasons, including disagreements over deal prices and strategy, but together they illustrate how aggressively Zuckerberg has been canvassing the industry to reboot his AI efforts.

Now, details about the team Zuckerberg is assembling are starting to come into view: SSI co-founder and CEO Daniel Gross, along with ex-Github CEO Nat Friedman, are poised to co-lead the Meta AI assistant. Both men will report to Alexandr Wang, the former Scale CEO Zuckerberg just paid over $14 billion to quickly hire. Wang told his Scale team goodbye last Friday and was in the Meta office on Monday. This week, he has been meeting with top Meta leaders (more on that below) and continuing to recruit for the new AI team Zuckerberg has tasked him with building. I expect the team to be unveiled as soon as next week.

Rather than join Meta, Sutskever, Murati, and Perplexity CEO Aravind Srinivas have all gone on to raise more money at higher valuations. Sutskever, a titan of the AI research community who co-founded OpenAI, recently raised a couple of billion dollars for SSI. Both Meta and Google are investors in his company, I’m told. Murati also just raised a couple of billion dollars. Neither she nor Sutskever is close to releasing a product. Srinivas, meanwhile, is in the process of raising around $500 million for Perplexity.

Spokespeople for all the companies involved either declined to comment or didn’t respond in time for publication. The Information and CNBC first reported Zuckerberg’s talks with Safe Superintelligence, while Bloomberg first reported the Perplexity talks.

While Zuckerberg’s recruiting drive is motivated by the urgency he feels to fix Meta’s AI strategy, the situation also highlights the fierce competition for top AI talent these days. In my conversations this week, those on the inside of the industry aren’t surprised by Zuckerberg making nine-figure — or even, yes, 10-figure — compensation offers for the best AI talent. There are certain senior people at OpenAI, for example, who are already compensated in that ballpark, thanks to the company’s meteoric increase in valuation over the last few years.

Speaking of OpenAI, it’s clear that CEO Sam Altman is at least a bit rattled by Zuckerberg’s hiring spree. His decision to appear on his brother’s podcast this week and say that “none of our best people” are leaving for Meta was probably meant to convey a position of strength, but in reality, it looks like he is throwing his former colleagues under the bus. I was confused by Altman’s suggestion that Meta paying a lot upfront for talent won’t “set up a great culture.” After all, didn’t OpenAI just pay $6.5 billion to hire Jony Ive and his small hardware team?

“We think that glasses are the best form factor for AI”

When I joined a Zoom call with Alex Himel, Meta’s VP of wearables, this week, he had just gotten off a call with Zuckerberg’s new AI chief, Alexandr Wang.

“There’s an increasing number of Alexes that I talk to on a regular basis,” Himel joked as we started our conversation about Meta’s new glasses release with Oakley. “I was just in my first meeting with him. There were like three people in a room with the camera real far away, and I was like, ‘Who is talking right now?’ And then I was like, ‘Oh, hey, it’s Alex.’”

The following Q&A has been edited for length and clarity:

How did your meeting with Alex just now go?

The meeting was about how to make AI as awesome as it can be for glasses. Obviously, there are some unique use cases in the glasses that aren’t stuff you do on a phone. The thing we’re trying to figure out is how to balance it all, because AI can be everything to everyone or it could be amazing for more specific use cases.

We’re trying to figure out how to strike the right balance because there’s a ton of stuff in the underlying Llama models and that whole pipeline that we don’t care about on glasses. Then there’s stuff we really, really care about, like egocentric view and trying to feed video into the models to help with some of the really aspirational use cases that we wouldn’t build otherwise.

You are referring to this new lineup with Oakley as “AI glasses.” Is that the new branding for this category? They are AI glasses, not smart glasses?

We refer to the category as AI glasses. You saw Orion. You used it for longer than anyone else in the demo, which I commend you for. We used to think that’s what you needed to hit scale for this new category. You needed the big field of view and display to overlay virtual content. Our opinion of that has definitely changed. We think we can hit scale faster, and AI is the reason we think that’s possible.

Right now, the top two use cases for the glasses are audio — phone calls, music, podcasts — and taking photos and videos. We look at participation rates of our active users, and those have been one and two since launch. Audio is one. A very close second is photos and videos.

AI has been number three from the start. As we’ve been launching more markets — we’re now in 18 — and we’ve been adding more features, AI is creeping up. Our biggest investment by a mile on the software side is AI functionality, because we think that glasses are the best form factor for AI. They are something you’re already wearing all the time. They can see what you see. They can hear what you hear. They’re super accessible.

Is your goal to have AI supersede audio and photo to be the most used feature for glasses, or is that not how you think about it?

From a math standpoint, at best, you could tie. We do want AI to be something that’s increasingly used by more people more frequently. We think there’s definitely room for the audio to get better. There’s definitely room for image quality to get better. The AI stuff has much more headroom.

How much of the AI is onboard the glasses versus the cloud? I imagine you have lots of physical constraints with this kind of device.

We’ve now got one billion-parameter models that can run on the frame. So, increasingly, there’s stuff there. Then we have stuff running on the phone.

If you were watching WWDC, Apple made a couple of announcements that we haven’t had a chance to test yet, but we’re excited about. One is the Wi-Fi Aware APIs. We should be able to transfer photos and videos without having people tap that annoying dialogue box every time. That’d be great. The second one was processor background access, which should allow us to do image processing when you transfer the media over. Syncing would work just like it does on Android.

Do you think the market for these new Oakley glasses will be as big as the Ray-Bans? Or is it more niche because they are more outdoors and athlete-focused?

We work with EssilorLuxottica, which is a great partner. Ray-Ban is their largest brand. Within that, the most popular style is Wayfair. When we launched the original Ray-Ban Meta glasses, we went with the most popular style for the most popular brand.

Their second biggest brand is Oakley. A lot of people wear them. The Holbrook is really popular. The HSTN, which is what we’re launching, is a really popular analog frame. We increasingly see people using the Ray-Ban Meta glasses for active use cases. This is our first step into the performance category. There’s more to come.

What’s your reaction to Google’s announcements at I/O for their XR glasses platform and eyewear partnerships?

We’ve been working with EssilorLuxottica for like five years now. That’s a long time for a partnership. It takes a while to get really in sync. I feel very good about the state of our partnership. We’re able to work quickly. The Oakley Meta glasses are the fastest program we’ve had by quite a bit. It took less than nine months.

I thought the demos they [Google] did were pretty good. I thought some of those were pretty compelling. They didn’t announce a product, so I can’t react specifically to what they’re doing. It’s flattering that people see the traction we’re getting and want to jump in as well.

On the AR glasses front, what have you been learning from Orion now that you’ve been showing it to the outside world?

We’ve been going full speed on that. We’ve actually hit some pretty good internal milestones for the next version of it, which is the one we plan to sell. The biggest learning from using them is that we feel increasingly good about the input and interaction model with eye tracking and the neural band. I wore mine during March Madness in the office. I was literally watching the games. Picture yourself sitting at a table with a virtual TV just above people’s heads. It was amazing.

If you haven’t already, don’t forget to subscribe to The Verge, which includes unlimited access to Command Line and all of our reporting.

As always, I welcome your feedback, especially if you’ve also turned down Zuck. You can respond here or ping me securely on Signal.