# Meta’s Complicated AI Context > Source: > Published: 2026-08-28 11:38:06+00:00 # Meta’s Complicated AI Context There's a lot of positive buzz around Meta's AI efforts at the moment – if nothing else, it's impressive the *speed* at which they've zoomed back at least [ near the frontier](https://spyglass.org/meta-muse-spark-ai/). Can the forthcoming 'Watermelon' model fully get them there? Or set some kind of new mark? We should see soon enough, but more important for Meta will be creating an actual *product*people want to use AI for. [Will 'Hatch'](https://www.businessinsider.com/meta-hatch-personal-ai-agent-capabilities-employees-memo-2026-8?ref=spyglass.org)be the consumer harness that matters? The "OpenClaw for Normies" that [I've been waiting for](https://spyglass.org/openclaw-versus-closed-claude/)? I remain skeptical simply because of [Meta's recent history](https://spyglass.org/meta-many-bets/). When they can't buy, they've had trouble *building* products that actually resonate of late. It *seemed* like they [found their buy-in here with Manus](https://spyglass.org/meta-manus-deal/), but, well, [China had other ideas](https://spyglass.org/inklings-googles-anthropic-money-openaiphone-xchat-elon-vs-openai-china-blocks-manus/#:~:text=%F0%9F%87%A8%F0%9F%87%B3%20China%20Blocks%20Meta%27s%20Manus%20Deal). So the question in my mind is if they controlled the company long enough to effectively clone the IP before they had to [give it back](https://spyglass.org/inklings-ai-overkill-anthropic-walks-back-siri-ai-just-works-elons-diamond-hands-apple-tv-is-hbo-openais-rsi-clause-siri-is-not-dtf/#:~:text=Meta%27s%20Manus%20shitshow%20continues%2C%20as%20the%20companies%20are%20being%20cleaved%20apart%20%E2%80%93%20appropriately%2C%20with%20firewalls%20put%20in%20place%20%E2%80%93%20as%20the%20buy%2Dback%20talks%20continue.%20%5BBloomberg%20%F0%9F%94%92%5D)? Anyway, this dive into Meta's history with AI by Harry McCracken gives some important context for the go-forward story. (And I swear I'm not just saying that because I'm quoted twice in the piece.) Mark Zuckerberg certainly gets credit for some early insight into AI being important back in 2013 when he hired Yann LeCun – but this was also because he had just lost the bidding to buy DeepMind. (And in [pretty embarrassing fashion](https://thoughts.spyglass.org/i/192334503/when-larry-met-demis-at-elons-birthday?ref=spyglass.org), according to [subsequent reporting](https://www.wsj.com/tech/ai/deepmind-google-demis-hassabis-5bd6de54?ref=spyglass.org).) This directly led to the creation of PyTorch, the open source deep learning library which Meta absolutely does deserve credit for and propelled the entire field forward. But as for the "open" strategy with their actual early Llama models, people forget, but that was decidedly more complicated: While Zuckerberg was selling the world on headsets and smart glasses, a twist of fate helped propel Meta near the forefront of AI labs—at least for a time. In February 2023, Meta announced a new LLM called Llama (for Large Language Meta AI). At first, it planned to share its creation with academic researchers on a case-by-case basis. But a week later, as the world was still wrapping its mind around the three-month-old ChatGPT, a leaked version showed up on the dark web site 4chan. Suddenly, Llama was available to anyone who could figure out how to download and install it. And anyone who did could modify the code to suit their own purposes. Llama’s unintentional release led to fears of it being misused by bad actors. Some of that played out, including the creation of a Llama-powered Discord bot that spewed hate speech. But there was also excitement about the freely available model’s potential to democratize the spread of useful forms of AI. Meta had been distributing the Llama weights to outside researchers, which is undoubtedly how they leaked, but there was no indication that this was *the actual strategy* for the general public. In fact, when it happened, everyone was scared shitless that Meta has just recklessly endangered the world! That was overblown, of course, and so it (retroactively – though undoubtedly guided by Meta's previous success with opening up some of their core technologies, including the aforementioned PyTorch) became the strategy: OpenAI, Anthropic, and Google tightly control their primary models. Zuckerberg realized he could drive improvements more quickly by sharing Meta’s work. The prospect rattled other companies. “Paradoxically, the one clear winner in all of this is Meta,” wrote Google engineer Luke Sernau in an internal memo published by SemiAnalysis. “Because the leaked model was theirs, they have effectively garnered an entire planet’s worth of free labor.” Less than five months later, Meta leaned into that advantage. Rather than vetting users, the company simply released Llama 2 to the public as an open-weight model. (AI purists use this term for models such as Llama, whose training data remains proprietary, making them less than fully open source.) Then came Llama 3, which was an even bigger success. Meta seemed to be off to the races with Zuckerberg [talking up "open"](https://spyglass.org/zuckerberg-llama-open-source-ai/) models non-stop on [every stage](https://spyglass.org/meta-open-ai/) (and [every podcast](https://spyglass.org/meta-llama-ai-zuckerberg-open-ai/)) that he could. Then the record scratch... Instead of cementing Meta (and "open" models) as the leader in AI, Llama 4 was an unmitigated disaster. The [gaming of leaderboards](https://www.theverge.com/meta/645012/meta-llama-4-maverick-benchmarks-gaming?ref=spyglass.org) was a symptom, not the cause. While it was easy to [blame "open"](https://spyglass.org/metas-open-source-ai-mistake/), the reality was that Meta got outflanked by other models [using mixture-of-experts (MoE) techniques](https://spyglass.org/metas-moe-ai-mistake/) to train. That was perhaps the biggest revelation of [the "DeepSeek Moment"](https://spyglass.org/deepseek-moment/), for Meta at least. Well that, and the [notion that distillation](https://spyglass.org/deepseek-is-seeking-you/) of models can work quite well – *perhaps especially if your model happens to have open weights*.1 And so ["open" was closed](https://spyglass.org/open-source-ai-was-the-path-forward/). And Llama was killed. [Scale was "hackquired"](https://spyglass.org/hackquisitions-hackquihires/). And billions were burned (on both [talent](https://spyglass.org/zucks-eleven/) and [compute](https://spyglass.org/metas-ai-relativity-theory/)). Again, the initial results look pretty good. But then came [the "Kimi K3 Moment"](https://spyglass.org/open-vs-closed-ai/) mixed with [the "AI Overkill" movement](https://spyglass.org/ai-overkill/) (and the [Anthropic heel turn](https://spyglass.org/anthropic-becomes-the-villain/), and ongoing OpenAI backlash), and [suddenly "open" was cool again](https://spyglass.org/open-weight-ai-models/)! As such, [Meta quickly pivoted back](https://spyglass.org/meta-open-ai-muse-glimmer/) – though [not fully this time](https://spyglass.org/meta-open-ish-ai-models/) – just in a way to try to take advantage of the marketing and framing as they try to [re-establish themselves](https://spyglass.org/meta-ai-cloud-business-model/) [externally](https://spyglass.org/you-are-the-tokens/) – and re-align themselves *internally*. Back in 2015, Zuckerberg had proudly told me that his success as a CEO had come from the effort he put into “building a culture where people think about the mission in the same way that I do.” Now that synchronicity seems to be gone, at least for the moment. According to a current MSL staffer, employees in the lab regard Zuckerberg’s “personal superintelligence” talk as sloganeering, not a road map for building useful products. One Silicon Valley insider says the richly compensated AI recruits he knows at the company are “mercenaries,” there “to ride this out for as long as Zuck’s interests hold and take as much money off the table as possible.” Morale among longer-serving staffers, who have seen friends and mentors depart by edict or choice, has cratered, says the MSL employee. "Sloganeering" is a great term. I still think if they nail 'Watermelon' (and can they come up with a better name than 'Muse Spark' for it?) and certainly 'Hatch' these problems will fade away. Success has a funny way of changing perception, fast and clearing up any internal morale issues. We'll see, it's nearly the end of [this particular race](https://spyglass.org/ai-frontier-race/)... **Previously, on** **Spyglass** **...**[ 1] Models which many other Big Tech companies were also [happy to use, but not pay for](https://spyglass.org/llama-milk/), which was obviously an issue for Meta as well given the costs. [It's almost like they needed a cloud...](https://spyglass.org/meta-cloud/) [↩](#)