What I read yesterday
21/11/2024
Fortune Magazine
We recently wrote about Meta and that note can be found here. We discussed how Meta is doubling down on it AI Bet. An integral part of their strategy is to release their LLMs (known as Llama) as open source so they can benefit from systems and code developers build using Llama. Yesterday we read a quite long Fortune article which outlined how Llama has re-formed Meta. We present an edited version of it below.
How Mark Zuckerberg has fully rebuilt Meta around Llama
It was the summer of 2023, and the question at hand was whether to release a Llama into the wild.
The Llama in question wasn’t an animal: Llama 2 was the follow-up release of Meta’s generative AI mode—a would-be challenger to OpenAI’s GPT-4. The first Llama had come out a few months earlier. It had originally been intended only for researchers, but after it leaked online, it caught on with developers, who loved that it was free—unlike the large language models (LLMs) from OpenAI, Google and Anthropic—as well as state-of-the-art. Also unlike those rivals, it was open source, which meant researchers, developers, and other users could access the underlying code and its “weights” (which determine how the model processes information) to use, modify, or improve it.
Yann LeCun, Meta’s chief AI scientist, and Joelle Pineau, VP of AI research and head of Meta’s FAIR (Fundamental AI Research) team, wanted to give Llama 2 a wide open-source release. They felt strongly that open-sourcing Llama 2 would enable the model to become more powerful more quickly, at a lower cost. It could help the company catch up in a generative AI race in which it was seen as lagging badly behind its rivals, even as the company struggled to recover from a pivot to the metaverse whose meagre offerings and cheesy, legless avatars had underwhelmed investors and customers.
But there were also weighty reasons not to take that path. Once customers got accustomed to a free product, how could you ever monetize it? And as other execs pointed out in debates on the topic, the legal repercussions were potentially ugly: What if someone hijacked the model to go on a hacking spree? It didn’t help that two earlier releases of Meta open-source AI products had backfired badly, earning the company tongue-lashings from everyone from scientists to U.S. senators.
Zuckerberg has long touted open-source technology (Facebook itself was built on open-source software), but he likes to gather all opinions; he spoke to “everybody who was either for, anti, or in the middle” on the open-source question, recalls Ahmad Al-Dahle, Meta’s head of generative AI. But in the end it was Zuckerberg himself, LeCun says, who made the final decision to release Llama 2 as an open-source model: “He said, ‘Okay, we’re just going to do it.’”
On July 18, 2023, Meta released Llama 2 “free for research and commercial use.”
..in hindsight, the release of Llama 2 marked a crucial crossroads for Meta and Zuckerberg—the beginning of a remarkable comeback, all thanks to tech named after a furry camelid. By the time llama 3 models were released in April and July 2024, Llama had mostly caught up to its closed-source rivals in speed and accuracy. On several benchmarks, the largest Llama 3 model matched or outperformed the best proprietary models from OpenAI and Anthropic. One advantage in Llama’s favour: Meta uses publicly shared data from billions of Facebook and Instagram accounts to train its AI models.
Llama is at the core of a complete strategic pivot on the part of Meta to go all in on generative AI. Zuckerberg is now seen as a champion of “democratizing tech” among Silicon Valley developers—just two years after he and his company were being questioned, and sometimes mocked, for going all in on the metaverse, and vilified for having contributed to political polarization, extremism, and harming the mental health of teenagers.
While ChatGPT remains the dominant gen AI tool in the popular imagination, Llama models now power many, if not most, of the Meta products that billions of consumers encounter every day.
Meta’s AI assistant, which reaches across Facebook, Instagram, WhatsApp, and Messenger, is built with Llama, while users can create their own AI chatbot with AI Studio. Text-generation tools for advertisers are built on Llama. Llama helps power the conversational assistant that is part of Meta’s hit Ray-Ban glasses, and the feature in the Quest headset that lets users ask questions about their surroundings.
The company is said to be developing its own AI-powered search engine.
Llama models have been downloaded over 600 million times on sites like open-source AI community Hugging Face.
Still, the pivot has perplexed many Meta watchers. The company has spent billions to build the Llama models. Meanwhile, it’s giving Llama away for free to thousands of companies, including giants like Goldman Sachs, AT&T, and Accenture. Some investors are struggling to understand where and when, exactly, Meta’s revenue would start to justify the eye-watering spend.
Nonetheless, Llama’s contrarian success has allowed Zuckerberg to shrug off the lukewarm response to his metaverse ambitions and the company’s painful “year of efficiency” in late 2022 and early 2023.
The rise of Llama has also given Zuckerberg a chance to address a long-simmering sore point in his otherwise meteoric career: the fact that Facebook, and now Meta, have so often seen their services and products constrained by rules imposed by Apple and Google—the rival giants whose app stores are Meta’s primary points of distribution in the mobile device era. As he wrote in a July blog post: “We must ensure that we always have access to the best technology, and that we’re not locking into a competitor’s closed ecosystem where they can restrict what we build.”
With Llama, Meta and Zuckerberg have the chance to set a new industry standard. In the long run, OpenAI’s ChatGPT may be seen as the fiery spark that ignited the generative AI boom. But for now, at least, Llama’s own future’s so bright, Zuckerberg has gotta wear AI-powered Ray-Ban shades.
Prior to the generative AI boom, Meta’s use of AI was mostly behind the scenes—either research focused or integrated under the hood of its recommendation algorithms and content moderation.
Generative AI began to take off with OpenAI’s release of ChatGPT, just as the Meta pivot was looking particularly unwise. With metaverse spending through the roof and consumers utterly uninterested, Meta’s stock hit a seven-year low, inspiring headlines like, “How Much Trouble Is Mark Zuckerberg In?” The company began laying off thousands of employees.
Despite restrictions, the original full (Llama) model leaked online within weeks,
Meta leaders were taken aback by the sheer demand for the leaked Llama model from researchers and developers. These would-be users wanted the flexibility and control that would come with open access to a profoundly powerful LLM.
· A law firm, for example, could use it to train a specialized model for legal use—and own the intellectual property.
· A health care company could audit and manage the data behind the model, ensuring HIPAA compliance.
· Researchers could experiment and examine the inner workings of the model. “We got incoming requests from people who said, ‘You have to open-source that stuff. It’s so valuable that you could create an entire industry, like a new internet,’” LeCun says.
Given this response to the leaked Llama1, they decided to make Llama 2 open -source.
Llama 2 was not entirely open. Meta did not disclose the datasets—including all that Facebook and Instagram material—used to train the model, which are widely regarded as its key competitive advantage.
It also restricted usage by companies with more than 700 million monthly active users, primarily meant to deter Meta’s Big Tech competitors.
But the source code and model weights could be downloaded, and Meta encouraged users to contribute improvements, bug fixes, and refinements of results to a collaborative community.
Meta's strategy can seem counterintuitive, coming from a company with $135 billion in annual revenue. Open-source software has typically been seen as a way to democratize technology to the advantage of small startups or under-resourced teams— the kinds scrambling to compete with giants like Meta.
In a July 2024 blog post called “Open Source Is the Path Forward,” Zuckerberg made it clear that giving away Llama is not an altruistic move. Open-sourcing, he said, would give Meta a competitive edge in the AI race—and could eventually make Llama the go-to platform for generative AI. Just as important, he wrote: “Openly releasing Llama doesn’t undercut our revenue, sustainability, or ability to invest in research like it does for closed providers” like OpenAI or Google.
Shweta Khajuria, an analyst at Wolfe Research who covers Meta, calls releasing Llama as open-source “a stroke of genius” that will enable Meta to
· attract top talent,
· accelerate innovation on its own platform,
· develop new revenue sources, and
· extend its longevity.
Khajuria believes there will be plenty of new monetization opportunities for Meta down the line, such as subscription and advertising options for current Meta AI features based on Llama, as well as AI-powered in-app business messaging. “Meta benefits from having billions of users where Perplexity and Claude and ChatGPT don’t necessarily have that base,” she says. “Once they have a critical mass of users and usage around the world, they can monetize.”
Patrick Wendell is cofounder and VP of engineering at data and AI company Databricks..He sees Meta’s move as much more far-reaching. If the internet was the first big wave of technology, which enabled Facebook’s creation, and mobile was the second, dominated by Apple and Google, “I think [Zuckerberg’s] calculus is the third big wave is coming, and he does not want to have one or two companies completely control all access to AI,” Wendell says. “One way you can avoid that is by basically commoditizing the market, giving away the core IP for free…so no one gains a monopoly.”
Current versions of Llama still have restrictions that traditional open-source software doesn’t (including lack of access to datasets). In October, the Open Source Initiative, which coined the term, criticized Meta for “confusing” users and "polluting” the nomenclature, and noted that Google and Microsoft had dropped their use of the term (using the phrase “open weights” instead).
..Meta is winning where it matters. Nathan Lambert, a research scientist at the nonprofit Allen Institute for AI, says that more than 90% of the open-source AI models currently in use are based on Llama. Open-source coders accept that Zuckerberg “has some corporate realities that will distort his messaging,” he says. “At the end of the day, the community needs Llama models.
Meanwhile the company is hard at work on Llama 4 models currently being trained on a cluster of over 100,000 pricey Nvidia GPUs, a cluster that Zuckerberg recently said was “bigger than anything that I’ve seen reported for what others are doing.”
Not everyone loves the idea of a bigger-than-anything Llama. -Some critics fear that an open-source model like Llama is dangerous in the hands of malicious actors, precisely because it’s too open. Those concerns may grow in today’s tense geopolitical atmosphere. On Nov. 1, Reuters reported that China’s army had built AI applications for military use on the back of an early version of Llama.
In a co-written essay in August, Zuckerberg and Spotify cofounder Daniel Ek noted that open-source development is “the best shot at harnessing AI to drive progress and create economic opportunity and security for everyone.”
Whatever the outcome of Meta’s increasingly loud open-source activism, many argue that Zuckerberg is exactly the right messenger. His personal involvement in promoting Llama and open-source, insiders agree, is the key reason Meta has been able to move with such speed and focus. “He’s one of a few founder leaders left at these big tech companies,” says Clegg. “One of the great advantages of that means you have a very short line of command.”
Zuckerberg’s renewal, and Meta’s transformation, are sure to test investor patience due to skyrocketing capital expenditures. Khajuria, the Wolfe analyst, says investors will tolerate it for now “because Meta has laid the groundwork of telling folks what the opportunity is.” That said, if revenue does not begin accelerating, exiting 2025 into 2026, “I think investors will start losing patience,” she warns. (Zuckerberg is somewhat insulated from investor discontent; he controls about 61% of voting shares at Meta.)
One thing is clear, LeCun says: The kind of gamble Meta is taking, with its massive investment in GPUs and all things generative AI, requires a leader willing to take big swings. And Meta has not only that leader, but a massively profitable core business to fund the vision.
AI Smart Glasses
Meta is the leader in the market for AI Smart Glasses. However, they will have a new competitor soon. According to an article in Digiwatch, Samsung has teamed up with Google And Qualcomm to develop a rival product.
Samsung has teamed up with Google and Qualcomm to develop advanced AI-powered smart glasses, set for release in Q3 2025. Initial production will feature 500,000 units, targeting a competitive edge over existing options like Meta’s and Ray-Ban’s smart glasses. Equipped with AI and augmented reality (AR) technologies, the glasses promise enhanced interactivity and user experiences.
The device boasts Qualcomm’s AR1 chip for performance and NXP’s auxiliary processor for added computing. High-resolution imaging is ensured with a 12MP Sony IMX681 camera, supporting superior video and image capture. Lightweight at 50 grams, it offers features like gesture and human recognition, QR-based payments, and extended use powered by a 155mAh battery.
Google’s Gemini large language model will integrate into the software, delivering smarter user interactions and contextual understanding. Samsung disclosed the development during its earnings report, with analysts expecting a possible showcase at the January Galaxy Unpacked event, alongside the Galaxy S25.
Market excitement grows as Samsung enters the smart glass market, setting a high standard for innovation and functionality. Observers anticipate a significant shift in wearable technology driven by AI and AR advancements.
The poorest US state, Mississippi has higher GDP than France or Italy, the second poorest, West Virginia, is higher than Canada or Germany.
21/11/2024