@Aron I just watched the first 25minutes of this new interview with Mark Zuckerberg which increases my confidence in their Ai strategy as they do fundamentally a lot of the right things in my opinion which should put them on a strong growth trajectory. (Like optimizing for what users like most based on user feedback and observed user behavior and interaction with their Ai products and building ai agents for ai research and development)
Overall I feel we should get better in spotting those kinds of interviews immediately and reporting on their most important insights as they are a very valuable and direct source when it comes to evaluating companies. (Very important to continuously optimize your YouTube feeds for getting the best possible recommendations by following the right channel, liking the right videos etc)
They did not. The event primarily focused on how Meta developed the Llama 4 Scout and Maverick models—incorporating feedback from Llama 3 users—along with how developers can access the models through the Llama API, and how Databricks customers are already leveraging Llama.
In an interview with Microsoft CEO Satya Nadella, Nadella noted that 20–30% of the code written inside Microsoft is now generated by AI, and that this percentage is steadily increasing. When Nadella asked Zuckerberg how much of Meta’s code is AI-generated, Zuckerberg said he didn’t know the exact figure off the top of his head, but added that Meta is building an AI model designed to help build future versions of the Llama models (minute 43:00).
Zuckerberg also said it’s unclear how the Behemoth model will ultimately be used, except that it will likely be distilled into more deployable versions. He confirmed that Behemoth’s pre-training is complete, and the team is now working on post-training (minute 55:00).
You are right. Closed-sourced models such as GPT and Gemini prohibit the use of their models to develop “competing” models.
Analysts largely impressed by the outcome of LlamaCon
Analyst Justin Post of Bofa said the event highlighted Llama’s flexibility and growing toolset for GenAI developers.
“The developer-focused event emphasized Llama’s flexibility and a growing set of tools and capabilities it offers for building and scaling Gen-AI applications,” Post wrote.
Brian Nowak of Morgan Stanley said Meta AI app still lags behind ChatGPT and Google Gemini but acknowledged it’s only the first version. He highlighted its strengths such as personalization via user data, free access vs. competitor paywalls, and promising digital assistant use cases (e.g., travel planning, shopping).
Analyst Brent Thill of Jefferies estimated Llama could be worth $80 billion and said Meta’s launch of its own Llama API shifts it from relying on third-party cloud APIs (e.g., AWS, Azure) to directly monetizing its own AI infrastructure.
“By announcing the Llama API, META positions itself to directly monetize Llama, marking its first foray into offering its own cloud infrastructure/platform for building and running Llama apps,” he wrote.
Citizens JMP analysts believe Meta could leverage its advertising products and distribution capabilities to drive the growth of Meta AI, enabling it to potentially gain a foothold in the search market over time.
Meta projected in 2024 that Gen AI could bring in between $460 billion and $1.4 trillion by 2035
According to unsealed court fillings, Meta projected in 2024 that its Gen AI driven revenue would be $2 billion in 2025 and between $460 billion and $1.4 trillion by 2035.
Bloomberg projected in 2024 that Gen AI could increase Meta’s ad revenue by $207 billion through 2032.
FAIR’s co-founder, Robert Fergus, who returned recently from Alphabet’s DeepMind, will lead the division
Meta appoints Robert Fergus, who recently joined the division from Alphabet’s DeepMind, to lead FAIR.
Fergus cofounded FAIR in 2014 and left Meta in 2020.
“Meta’s commitment to FAIR and long term research remains unwavering. We’re working towards building human-level experiences that transform the way we interact with technology and are dedicated to leading and advancing AI research,” announced in a LinkedIn post.
FAIR works on long-term projects that may not be tied to a consumer product and was responsible for building the original Llama.
Meta has further delayed the rollout of Llama 4 Behemoth and is reportedly considering restructuring its AI team
Meta is delaying the rollout of its Llama 4 Behemoth due to concerns over its capabilities, the Wall Street Journal reported citing people familiar with the matter.
According to sources, Meta engineers are struggling to improve the performance of Llama 4 Behemoth, which was initially scheduled for release in April, then postponed to June, and has now been delayed until the fall or later.
The sources said senior executives at the company are frustrated with the performance of the team building Llama 4 models and are contemplating making changes to the AI product group.
Meta divides its AI team into two to speed up execution, no senior official is leaving due to the changes
Meta divides its AI team into two; AI products team and AGI Foundations unit so as to speed up the rollout of new products and features, Axios reported citing an internal memo.
AI products team headed by Connor Hayes will be responsible for Meta AI assistant, Meta’s AI studio and features within Facebook, Instagram and WhatsApp.
AGI Foundations unit, co-headed by Ahmad Al-Dahle and Amir Frenkel will be responsible for the company’s Llama models and efforts to improve reasoning, multimedia and voice capabilities.
FAIR (Fundamental AI Research) will remain independent, though the team working on Multimedia has been moved to AGI.
Ahmad Al-Dahle has been the head of the original Gen AI team while Connor Hayes has been the head of product in Gen AI team.
Amir Frenkel was VP of engineering and product at Reality Labs.
The restructuring follows reports that Meta is delaying the launch of Llama 4 Behemoth due to concerns over its capabilities, April release of Llama 4 Scout and Maverick which failed to impress developers and recent departure of FAIR head Joelle Pineau.
Meta reiterated April 29 comments indicating that Meta AI has around 1 billion monthly actives.
“It may seem kind of funny that a billion monthly actives doesn’t seem like it’s at scale for us, but that’s where we’re at,” Zuckerberg told shareholders yesterday.
Meta signs a 20 year nuclear power deal with Constellation Energy that is worth “billions of dollars”
Meta signed a 20 year nuclear power deal with Constellation Energy that is worth “billions of dollars”.
“It’s billions of dollars of capital that you’re signing up for to run a plant for 20 more years,” said Constellation CEO Joe Dominguez.
The Power Purchase Agreement (PPA) will enable Meta to buy the clean attributes of the power generation to offset the less clean power elsewhere i.e. Meta won’t take power from the Clinton Clean Energy site.
According to Constellation, the plant produces 1,091 megawatts of electricity.
Jefferies estimated that Microsoft paid $100-$115/MWh for a similar deal in 2024.
Assessment
Assuming 93% capacity factor and $115/MWh electricity price, it would cost Meta Platforms around $20.5 billion over 20 years (1,091 MW×24 h/day×365 days/year x0.93x$115).
@Aron I just saw that there is a new interview with the Meta CTO which might have some good insights.
Overall i would say Ai developments remain my number 1 concern with Meta esp. after 11 out of the 14 original authors of Llama left the company and Behemoth is getting delayed due to capability concerns.
So I think it remains one top area to stay focused and on the cutting edge of developments on as it will influence my investment decisions a lot.
Good catch. I agree with your assessment. By the way, they recently divided the AI team into two to speed up execution.
Bosworth mainly talked about Reality Labs. He didn’t say anything on AI other than that it has boosted their conviction that having the devices wearable, with sensors and contacts has been important (min 12.28).
He said he was skeptical about the VR/AR due to the investments involved but Zuckerberg found a way to have him take up the role (min 3:06). This confirms past employee comments which said Zuckerberg tends to get excited with things and hard to advise.
He pointed out that they are on track to make 2025 a “make” year for Reality Labs (min 7:00). He had previously said 2025 would be a “make or break year” for Reality Labs.
He signaled that they are ahead of Apple Vision Pro, pointing out that Apple’s engineering team executed well but made some mistakes when it comes to weight (min 17:50)
Meta in talks to invest over $10 billion in Scale AI
Meta is in talks to make over $10 billion investment in Scale AI, a startup company that provides data for training AI, Bloomberg reported citing people familiar with the discussions.
Founded in 2016, Scale AI generated $870 million in revenue last year and expects to grow revenue to $2 billion in 2025.
Scale AI was valued at $14 billion last year.
Meta already partnered with Scale AI in its Defence Llama project, a Llama model intended for military use.
Zuckerberg is personally hiring 50 AI researchers and engineers to join a new AI team code-named Superintelligence Labs group
CEO Mark Zuckerberg is personally hiring a team of AI researchers and engineers following the underperformance of Meta’a AI team, Bloomberg reported citing people familiar with the plans.
Zuckerberg wants to hire 50 people for the new team code named Superintelligence Labs group, including new head of AI research.
Scale AI founder Alexandr Wang is expected to join the new team after the deal is done.
The sources told Bloomberg that Zuckerberg has gone into “founder mode” in the last two months and believes that Meta should outstrip other companies when it comes to artificial general intelligence (AGI).
It’s not yet clear how Superintelligence Labs group will work with the existing AI teams but some employees are expected to move to the new AI team.
This comes after Meta recently divided its AI team into two; AI products team and AGI Foundations unit so as to speed up execution.
Assessment
Meta’s AI teams definitely needs changes or improvements following underperformance of Llama models. However, I am not sure if it’s a good thing to have multiple teams working on almost the same thing. Perhaps employees working on existing teams will be fired thereafter, which could signal that morale at the existing teams is currently low.
According to employees interviewed by CNBC, Ahmad Al-Dahle new position as co-head of AGI Foundations was seen as a sign that Zuckerberg had lost confidence in him.
The people said Zuckerberg admires Wang (a 28 year old dropout from the Massachusetts Institute of Technology) and considers him capable of being an AI leader at Meta.
The people said Wang will be joined by some of his colleagues from Scale AI.
The Informationreported yesterday that Meta had agreed to pay $14.8 billion for 49% stake in Scale AI and Wang.
Zuckerberg taking researchers from leading AI firms by offering seven- to nine-figure compensation packages
Meta is offering seven- to nine-figure compensation packages to dozens of researchers from leading AI companies such as OpenAI and Google and some have agreed to join, the New York Times reported citing people familiar with the company’s plans.
Meta’s new AI model, V-JEPA 2 helps robots understand and respond to the physical world
Meta releases a new AI model called V-JEPA 2 that enables robots and other AI agents to understand the physical world and respond to it.
“These capabilities are essential to building AI agents that can think before they act, and V-JEPA 2 represents meaningful progress toward our ultimate goal of developing advanced machine intelligence,” Meta said.
“The models we use to develop this kind of intelligence in machines are called world models, and they enable three essential capabilities: understanding, predicting and planning,” it added.
Google DeepMind’s principal researcher and Sesame’s machine learning lead will join Meta’s “superintelligence” team
Jack Rae, a principal researcher at Google DeepMind, will joined Meta’s “superintelligence” team, Bloomberg reported citing people familiar with the matter.
Johan Schalkwyk, a machine learning lead at AI voice startup Sesame AI will also join the team.
Rae confirmed he’s joining Meta while Schalkwyk declined to respond to Bloomberg’s request for a comment.
