Meta's CTO explains Muse: Returning to human-centered interaction, the core of the next generation of control systems is "intent understanding."

On the 24th, Meta's Chief Technology Officer, Andrew Bosworth, gave a systematic explanation of the company's AI assistant, Muse, and its broader AI strategy in a podcast. He pointed out that the current AI industry is overly focused on benchmark scores, neglecting the essence of user experience—this is where Muse differentiates itself.

Muse received high praise from Bosworth himself from the time it was in beta testing, and he said he quickly migrated a large amount of workload to the platform. After the product was officially released, the reactions from external users were highly consistent with internal expectations, which greatly encouraged his team. Bosworth emphasized that Muse's success did not come from maximizing the model's original intelligence, but from a deep reflection on the question of "how to enable humans to use tools better".

Meanwhile, Bosworth outlined Meta's vision for the next generation of human-computer interaction: a shift from "direct manipulation" to "intent understanding," with English becoming the new programming language. Regarding AI security and privacy, Meta has taken several noteworthy steps: the company delayed the release of Muse to refine its privacy architecture and launched end-to-end encrypted "private AI processing"—communication between users and servers is fully encrypted, with no records left on the server side. Bosworth characterizes this mechanism as a trusted infrastructure for the most sensitive application scenarios.

Muse's differentiation: prioritizing user experience over benchmark scores.

Bosworth directly pointed out a core misconception in the current AI industry. "We've spent a lot of time over the past few years discussing the science of models, discussing who's 10% or 20% ahead on a certain benchmark, but for most people, that doesn't really matter," he said. "What matters is the quality of the conversation and whether you trust it."

This is where Muse's design philosophy comes from. Bosworth explained that after obtaining a powerful base model, the team didn't simply maximize its raw performance, but focused on refining the user experience—including dialogue rhythm, tone of voice, and even a "little character" with visual feedback. This character isn't merely decorative: for users outside the early adopter circle, it solves a crucial cognitive problem—"Is the AI working now? How do I interact with it?"

Bosworth also specifically mentioned a unique feature of Muse: it proactively offers counterarguments to users, presenting an "other side" after giving advice. He believes this feature is not yet seen in other mainstream models.

From "Direct Manipulation" to "Intent Understanding": A Historic Shift in Interface Paradigms

Bosworth positions Muse and AI assistants more broadly as a paradigm shift in the history of human-computer interaction. He traces it back to the graphical interface developed by Xerox PARC for the Alto computer—this "direct control" paradigm has been used by Mac and Windows ever since, forming the mainstream interaction logic for the past half-century: click, window focus, one application at a time.

"If you're closing a window on a PC and the mouse cursor is a pixel off-center from the X button, nothing happens. A baby looking at the screen knows you want to click the X, but a computer doesn't," Bosworth said. "That's the problem."

In his view, the core value of AI lies in breaking this one-way relationship of "humans adapting to machines." When a system can understand the user's intent rather than just responding to precise instructions, human-computer interaction enters a completely new stage. He likens this evolution to the continuous abstraction of programming languages from machine code to assembly language, and then to Python: "English is becoming the new programming language"—a trend already concretely reflected in Meta's AR/VR glasses product line: AI assistants are integrated into the system layer, and users can control the spatial interface through natural language without manually moving the panel.

Private AI processing: Meta's bet on privacy infrastructure

At the privacy architecture level, Bosworth detailed Meta's "private processing" mechanism. It works as follows: communication between the user's device and the Meta server is end-to-end encrypted. Data is decrypted and processed on the server side only within a "Trusted Execution Environment," meaning even server operators cannot access the processed content, and the server retains no records after processing. The user's context information is stored locally on the device and is under the user's control.

Bosworth acknowledged that this approach involves product-level trade-offs—it means AI cannot retain memories across sessions—but he believes this sacrifice is worthwhile for use cases involving highly sensitive content. Notably, Meta partnered with Moxie Marlinspike, the founder of the Signal protocol, in the development of this architecture, and pledged to publish a white paper for independent security researchers to verify and to establish a bug bounty program.

When discussing public trust, Bosworth cited Microsoft as an example: the company faced heavy criticism for security issues in the late 1990s, but a decade later became an industry benchmark for security. "Trust comes on foot and goes on horseback," he said. "Rebuilding trust requires consistent, long-term execution to consistently meet or exceed user expectations."

AI Extinction Theory and AGI: Bosworth's "Middle Ground"

In response to the recent widespread claims on social media that AI will cause human extinction, Bosworth clearly stated his position, but with careful wording. "I cannot assign zero probability to these people, but I myself do not see that possibility," he said.

"I work with this technology and have a fairly deep understanding of its underlying mechanisms, so I feel confident in it."

He also holds a conservative view on the timeline for achieving AGI, believing that there is a widespread misconception in the industry that "we're almost there," and pointing out that this misconception has repeatedly appeared in the history of AI—he cited the Dartmouth Conference in the 1970s as an example, when attendees believed that general artificial intelligence would be achieved within a decade. "I think we are still quite a long way from what we now call AGI," he said.

On the issue of consciousness and alignment, Bosworth's position is close to Yann LeCun's: he does not believe that current AI possesses consciousness and warns that anthropomorphizing AI, giving it emotions or consciousness attributes, could have negative consequences from an alignment perspective. He simplifies "alignment" to a pragmatic definition: whether a tool does what you want it to do. "If it's not aligned, you won't use it. This is a powerful constraint that every product in human history has experienced."

Open source and "empowering people": Meta's AI political philosophy

Bosworth explained Meta's open-source AI strategy within the company's longer historical context. He cited Meta's original mission statement—"Empowering people to make the world more open and connected"—arguing that "empowering people" is the more fundamental anchor point of this statement, rather than "open" or "connected" itself.

"Concentrating AI in the hands of a few companies is not a way to maintain society in the long run," he said, explicitly expressing his support for Mark Zuckerberg's recent declaration that "AI belongs to everyone." On the product front, Meta has released Muse Glimmer—a distilled version of Muse Spark—released as open source.

Bosworth also acknowledged the controversy surrounding the distillation training method: this technique uses the output of a large model to train a smaller model, bypassing the enormous cost of training on raw data, thus sparking discussions within the industry regarding intellectual property and resource fairness. However, he clearly believes that the value of open access outweighs this cost.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.