Theo on Meta’s Personal AI Strategy and the Shift Beyond Open Models

In August 2025, Theo reviews Meta’s announced vision for “personal superintelligence” and offers his interpretation of the company’s position in AI. He mentions reports of large recruiting offers and discloses that he owns some Meta shares.

A clip of Mark Zuckerberg introduces Meta Superintelligence Labs and proposes AI that helps individuals pursue their goals. Zuckerberg contrasts this personal vision with a future centered on automating valuable work, and predicts that context-aware glasses could become primary computing devices.

Theo comments on the delivery of Zuckerberg’s video and argues that direct, human communication matters when a company is positioning its products as personal. He then turns to the written announcement and the strategic ideas behind it.

The announcement says AI systems have begun to show slow self-improvement and argues that superintelligence may eventually expand personal agency. Theo explains the idea that systems able to improve themselves might compound their capabilities, presenting it as a theory rather than an established outcome.

Theo questions whether giving everyone more capable tools necessarily increases each individual’s impact. He uses online video as an analogy: easier creation can increase the number of successful creators while also greatly increasing the number competing for attention.

The written vision emphasizes personal goals, relationships and experiences, and imagines devices that see and hear the user’s context. It also says new safety risks require care in deciding what to open-source. The statement does not specify how future products will work.

To explain his interpretation of Meta’s strategy, Theo compares Windows and Mac. He argues that Microsoft’s software platform reached many manufacturers and gave users a wide choice of hardware and applications, whereas Apple offered a more integrated and controlled experience.

He describes Microsoft retaining the ability to license its operating system to other manufacturers as an important platform advantage. In his telling, hardware became more interchangeable while a broad software ecosystem attracted users and developers.

Theo contrasts that with Apple’s narrower hardware and software choices. He then argues that mobile app stores changed how the distinction felt: even on a more controlled platform, consumers could discover and choose from many applications on the device.

The existence of an app store, he argues, made perceived choice and personalization more salient than the underlying openness of the platform for many users. He acknowledges that Android permits choices Apple does not and that price and market conditions also differ.

Theo says Meta historically favored relatively open platforms, citing React, virtual reality and Llama. In his account, Llama initially mattered because developers could adapt and run capable available-weight models, even if those models were not always the strongest.

He argues that competition from DeepSeek and other model developers changed the landscape. He also relays a report about Meta employees using other AI models internally; this is his reported information, not independently verified in the video.

Returning to the mobile analogy, he says people often care more about being able to find software that suits them than about the platform’s formal openness. He dislikes aspects of Apple’s App Store policies but thinks many consumers feel they have enough choice.

Theo compares Meta Quest with Apple Vision Pro and gives a sharply negative assessment of the VR market, despite praising the quality of the Vision Pro experience. His conclusion that VR will never be profitable is a personal prediction, not a fact established by the comparison.

He returns to Llama versus closed model providers. He says Llama benefited from its relative openness but, in his assessment at the time, faced strong competition from DeepSeek, Qwen, Kimi and GLM models. His model rankings are opinions about a changing market.

Theo’s central claim is that the useful distinction has shifted from open versus closed to whether a product feels capable of serving the user’s own needs. He interprets Meta’s announcement as a move toward personal use rather than simply winning on model openness.

He offers a speculative reconstruction of Meta’s internal thinking: open models initially provided a reason for developers to adopt Llama, but rivals made both the open-model and tool-use spaces more competitive. He explicitly acknowledges that he cannot know what company leaders were thinking.

Theo sketches a matrix separating open and closed systems from personal and work uses. He places conversational products, coding models and open models in different regions according to their perceived strengths. These placements are his interpretive framework, not formal product categories.

He points to then-current OpenRouter role-play and other usage categories as examples of how different models find audiences. He discusses DeepSeek’s conversational use and says newer Kimi and GLM releases strengthened the open side for agentic and tool-calling work.

In his view, the surrounding product can matter as much as the underlying model. He cites memory as a feature that made ChatGPT feel more personal, and recalls a controversy around overly agreeable behavior as an example of the difficulty of tuning that experience.

Theo distinguishes a pleasant conversational personality from personalization based on knowledge of a user. He argues that Meta’s social apps, messages and photos could give it a distinctive position if data access and product design support useful experiences.

He cites Meta’s reported 3.48 billion daily users at the time and suggests that frequent use of its apps could give a personal assistant a distribution advantage. The figure and his comparison reflect the 2025 discussion, not a current user count.

He speculates that a personalized assistant might improve recommendations or support advertising and referral revenue, for example by suggesting a restaurant. He presents these as possible business models, not confirmed Meta product features.

Theo thinks developer adoption was a historical reason for Meta’s open-platform investments, but believes intense competition limits that approach as an AI differentiator. He identifies Google as another company with substantial personal context and distribution.

He explores a hypothetical future where AI automates much of work and people spend more time on leisure. He acknowledges that unemployment and economic disruption could also follow; the extent and timing of automation are unknown.

In that scenario, he thinks Meta’s experience with consumer relationships and entertainment could be an advantage. He contrasts it with workplace software, where he believes Meta has had less traction, while mentioning React and his appreciation for its former workplace product.

Theo closes cautiously: he is curious how a personal AI product will work but, even as a shareholder, is unsure how much money Meta can make from it. He says he sees little Llama usage in the figures available to him for T3 Chat, a limited observation rather than a market-wide measurement.