LlamaCon 2025 live: Updates from Meta's first generative AI developer conference keynote

That's it for our liveblog today everyone, thank you for reading. Thanks so much too, to Karissa and Igor for working so hard on such a long day, and Devindra (and everyone else on the team) for the help. We'll see you lovely folks again on another liveblog.

I have boarded the LlamaCon shuttle and am finally headed home. Thanks to everyone who joined us today!

As we wrap, we got so many Little Llama teases and yet no concrete news on the new model.

Sounds like we're wrapping up here. Zuckerberg asks Nadella what he's looking forward to and Satya has a very Satya answer: software's ability to solve problems.

"The world needs a new factor of production to meet new challenges," Nadella says of the promise AI, adding how it took 50 years for industries to adapt to the advent of electricity. "We're all investing [in AI] like it won't take 50 years," Zuck responds, laughing.

We're getting a bit philosophical now. Nadella is saying that lines between apps, websites and documents are starting to blur, thanks to AI. Zuck breezes over his point "interesting, makes sense."

As the conversation turns to agentic systems, Zuck says he doesn't know "off the top of his head" the amount of Meta's code is being written by AI. For context, late last year Google CEO Sundar Pichai said a quarter of the search giant's new code was AI generated.

Nadella deftly dodges the topic of Microsoft's relationship with OpenAI by looking back to the company's history and his own history with open source projects.

Zuckerberg mentions Microsoft's partnership with OPenAI but points out that the company has also tried to support open-source models. Nadella says both closed and open-source models are needed: "I'm not dogmatic about it," he says.

Wow, Zuck is laying it on thick. He referred to Microsoft as the "greatest technology company" and said Nadella has been a solid ally in the open-source space. Nadella jokes that Zuckerberg once lectured him about Bing not being social, which gets a big laugh from the audience.

Maybe the AI can't count?

In case anyone is wondering, 10 minutes ago they told us it would start in 10 minutes. But they just now gave us a 5 minute warning (and a countdown) so hopefully we're *actually* starting soon.

I honestly don't know what to expect. My first instinct is to say "no." On the other hand, this is an odd matchup to say the least. As I mentioned earlier, we typically think of Meta and Microsoft as competitors in the AI space, and I can't think of a time when we've seen Zuck and Satya share a stage before so it definitely feels like they want to signal... something. I'm just not sure what.

Honestly shocked there still appears to be a sizable audience in front of that stage. I'm chowing down on some Korean stew at home. Karissa — do we expect Zuckerberg and Nadella to say anything of note?

They seem to be changing the stage over for a fireside, so hopefully we will be starting soon.

If you are all still with us at this point, welcome back and thanks for joining us this late in the day! It's 7pm here in the New York area. Definitely a choice for Meta to host this session at 4pm where they are.

Well, the last developer session just wrapped up and they announced a “networking break,” so I guess we will be waiting at least a few more minutes for Zuckerberg and Nadella.

I’m back in the main keynote area, where the AC is apparently set to arctic. The prior session is still happening, so unclear if Zuckerberg’s fireside will be starting on time.

I've certainly never had an ice cream taco and I've attended many Google, Apple, Amazon and Microsoft events (not to mention the smaller tech companies too). So kudos, Meta? And no, I didn't mean to brag :)

Excellent idea, Cherlynn. These were definitely above average, I'd say. Also, my first time getting an ice cream taco at a tech event.

Hi, I'm back just to campaign for us to have a review section dedicated to the Snacks of Tech Events. (SoTE?) How would you rate these, Karissa?

As we wait for Zuckerberg to return to the LlamaCon stage with Microsoft CEO Satya Nadella, I have to say I'm very curious to hear what they discuss. Microsoft is a major backer of Open AI, one of Meta's biggest competitors on the AI front so it was somewhat surprising to see him on the agenda. Interestingly, The Wall Street Journal published a story just yesterday reporting that Nadella and Open AI CEO Sam Altman have been "drifting apart" of late. The report notes several areas of tension between the two, including disagreements over OpenAI's access to Microsoft's compute resources. Given those reported tensions, I'll be paying extra close attention to Nadella and Zuckerberg's interactions.

One thing I'll say about Meta: they always keep visitors well-stocked when it comes to snacks. The afternoon break featured a soft pretzel bar (with a variety of dipping sauces to choose from) and mini ice cream tacos in four flavors.

In another bit of interesting timing, Meta is reporting earnings tomorrow for Q1 2025. Expect Zuck to talk more about all things Meta AI and Llama. (He's also likely to get some questions about whats happening over at Reality Labs, which was just hit with another round of layoffs.)

A lot of today’s updates have been pretty far in the weeds, but today is also highlighting how ubiquitous Llama already is. Earlier, we heard some examples from a Spotify engineer about how the company uses Meta’s models for features like AI playlists and its AI DJ. Karim Liman-Tinguiri, senior staff machine learning engineer at Spotify, explained that using a fine-tuned version of Llama helps the company keep costs and latency down. It also allows them to continually update their recommendations for users. For example, he said that if you create an AI playlist, like “jazz music for a roadtrip,” Spotify tracks how you interact with those songs and then feeds that data back into its model for better personalization in the future. (Spotify, by thew way, didn't always use Meta's AI models. When AI DJ first launched in 2023, the company used models from Open AI, but apparently has since switched over to Llama.)

I've been covering Meta for more than a decade, and been to many, many events they've hosted over the years. One thing that's stood out to me today is how low-key LlamaCon is compared to other big events like Connect (or F8 back in the day). LlamaCon is smaller than either of those ever were, but it also has a much more relaxed vibe overall. There are just a handful of sessions for developers this afternoon, and lots of space for them to kick back and relax. Of course, Meta has really big ambitions when it comes to AI and Llama, which does make me wonder if I'll be attending a much bigger LlamaCon at some point in the future.

Meta also made some other announcements at its inaugural LlamaCon, and you can read the details in the company's own words in the blog post it published today.

I'm in a session with Meta Director of Product, Dom Divakaruni, and he dropped another hint about the lightweight Llama 4 model Zuckerberg teased earlier today (the one he called "Little Llama"). Divakaruni said there would be a successor to the popular lightweight Llama 3 model coming "soon."

One piece of news that didn't get mentioned during the keynote today was something coming to WhatsApp called "Private Processing." According Meta, this is an "optional capability" that will let you use AI tools like summarizing messages while still keeping your conversations private.

Karissa writes:

"Buried in its LlamaCon updates, the company shared that it’s working on something called “Private Processing,” which will allow users to take advantage of generative AI capabilities within WhatsApp without eroding its privacy features.

WhatsApp, of course, is known for its strong privacy protections and end-to-end encryption. That would seem incompatible with cloud-based AI features like Meta AI. But Private Processing will essentially allow Meta to do both."

Read more: Meta's "Private Processing" could bring AI to WhatsApp chats without breaking privacy

During the keynote, Meta announced the new Llama API would be available as a limited free preview from today. It's meant to give developers a way to experiment with Meta's AI models.

Per Igor:

"The initial release of the Llama API includes tools devs can use to fine-tune and evaluate their apps.

Additionally, Meta notes it won't use user prompts and model responses to train its own models. "When you’re ready, the models you build on the Llama API are yours to take with you wherever you want to host them, and we don’t keep them locked on our servers," the company said. Meta expects to roll out the tool to more users in coming weeks and months."

Read more: Meta is making it easier to use Llama models for app development

In case you missed it, Meta did share some news during LlamaCon. Actually, it made some announcements right before the LlamaCon keynote started today. The company unveiled the Meta AI app this morning, calling it an "assistant that gets to know your preferences, remembers context and is personalized to you." It has already taken over as the companion app for Meta's AI glasses and includes a Discover feed where people can share how they're using the assistant.

According to our Will Shanklin:

"For users in the US and Canada, Meta AI can personalize its answers based on data you've shared with Meta products. This includes info like your social profile and content you like or engage with. The company says linking your Facebook and Instagram accounts to the same Meta AI account will provide "an even stronger personalized experience." If you don't want that, this might be a good time to check your privacy settings."

Read more: The new Meta AI app is a chatGPT competitor

I hate to bring us back down this rabbit hole but my reverse image search confirmed it is definitely not an Omega Seamaster Planet Ocean nor is it any number of Breitling, Citizen or Blancpain models that look similar. The most expensive of these I found was $12,000 (pre-owned), though. I will say my first thought was that it might be a Louis Vuitton Tambour Horizon. There's also a Tambour Street Diver model that looks very similar.

Oh we don't talk about the dot-com crash when we're in another bubble, Karissa

Alright, that's it for now. We're breaking for lunch, but will be back with more updates from LlamaCon, including another fireside with Microsoft CEO Satya Nadella to close things out.

A (probably inadvertently) amusing moment just now. Ghodsi is talking about how exciting things are in the AI world right now, and compares it to the early days of the internet around the dot com boom in 2000. He says we're at a similar point now in AI. What he doesn't mention, though, is that there was also a major crash in 2000 when a lot of early internet businesses went under!

OK... maybe I underestimated how much he spends on watches. I do not, for the tape, make this much money per year. And yes, he's a true everyman.

Ah yes, that makes sense Aaron. A man of the people!

Since we're in a bit of a lull, jumping in to contribute that Zuck has an obscene and well-documented (among the fashion blogs, at least) collection of Swiss watches. He seems to spend roughly my annual salary on a new watch each week.

This could have been a podcast, Zuck.

Zuck keeps dancing around reasoning models. Meta has yet to release their own chain-of-thought model. It doesn't seem like the company will announce anything new on that front today, though Ghodsi just said "We'd love to see it. Nudge, nudge."

This conversation is trending like it will be longer than the actual keynote. It's also ... a little dry. Zuckerberg and Ghodsi are obviously chummy, but we're not getting a ton of new information here (Zuckerberg's accessories notwithstanding.)

I did my best "zoom and enhance" and my best guess is that Zuckerberg is wearing a "dumb" watch. If any aficionados are reading this, let us know what you think it might be! There's at least one sub-dial that I can make out. Yep, I'm here for the most important reporting.

Interesting, they are now discussing DeepSeek. Zuckerberg notes that developers are able to "mix and match" models, while Ghodsi notes that Databricks saw a lot of people using Llama on top of DeepSeek.

For what it's worth, even the viewers commenting on the YouTube livestream are wondering if Zuckerberg is going to announce any actual new models here.

I unfortunately don't have a great angle here, but you can see a bright orange band.

I need Llama AI to analyze and detect what massive watch Zuck is wearing right now. Karissa points out he wouldn't wear an Apple Watch, so...

Not much to say here other than the fact that Zuck teased the company is working on a "Little Llama" model.

Zuckerberg and Ghodsi are talking about how the smaller Llama models have been the most popular among developers. Zuck mentions there is a smaller Llama 4 model in the works that is apparently known as "Little Llama" internally. No word on when that might drop.

Databricks' CEO is talking about how great it is that Llama is open-source, giving some examples of enterprise uses that wouldn't be possible otherwise, he also says that Llama helps them keep prices down for customers.

Databricks, for the uninitiated, provides software tooling companies can use to build their own machine learning applications.