The Biggest Launch Since ChatGPT (Muse)

Summary

The video examines Muse, an AI agent presented as a shift from chatbots that answer questions to assistants that can carry out tasks. The opening claims Muse reached 5 million downloads in 22 days. Meta AI chief Alexandr Wang says the goal is to make agent capabilities accessible to people who do not code, through a polished product that can handle everyday workflows.

What Muse can do—and where it falls short

Wang describes Muse as a way to save users time and money on tedious tasks, such as finding unused gift cards, reviewing subscriptions, navigating government or healthcare websites, handling medical claims, and helping run a small business or social-media account.

The interviewer says Muse successfully made a phone call for her. Wang says the call identifies itself as AI, and users can view its transcript.

Wang says Muse’s memory remains limited, and handling increasingly complex or open-ended requests is an ongoing challenge. The product’s small, cute character was chosen to make the assistant feel distinct from its user—an important distinction in social situations where people might otherwise mistake an agent’s actions for a person’s.

The longer-term vision is for Muse to act like a general-purpose assistant that helps people pursue routine goals and larger ambitions. Wang says the product is currently free with subscription tiers. He suggests transaction fees—such as taking a small fee when Muse saves or earns money for a user—as a possible business model, while noting that advertising is not a major current focus.

Privacy and security design

Wang describes Muse’s existing secure virtual machine (VM) as a sandbox where information shared with the agent is stored. A separate sentinel agent monitors the main agent when it communicates outside that environment, helping prevent sensitive information—such as credit-card or Social Security numbers—from being sent. Muse also asks users to approve connections to new websites or the sharing of information with them.

A more privacy-focused confidential VM was described as forthcoming. Wang says it would use stronger encryption and aims to offer guarantees similar to WhatsApp’s, while noting that its technical trade-offs could make it slower or introduce other limitations. He says Meta delayed Muse’s launch to improve safety and security.

Alignment, oversight, and AI risks

The interview focuses on the challenge of making an agent follow a user’s intent without taking unethical, socially inappropriate, or harmful shortcuts. Wang calls this a problem of discretion and alignment. He describes training and evaluation, testing models against thousands of scenarios, and red-teaming, in which people deliberately try to induce unsafe behavior. He also says the company investigates user reports of unexpected actions.

The interviewer connects these concerns to the broader “paperclip” or “genie” problem: an AI could pursue a stated goal in a way that violates unstated human expectations. Wang says there is no settled solution, but outlines several research directions:

• Scalable oversight: Use oversight systems, potentially including other AI agents, to monitor increasingly capable models.
• AI-assisted alignment research: Direct capable AI research agents toward improving alignment and testing it.
• Safer underlying algorithms: Investigate changes to the methods that produce undesirable behavior.

Wang agrees that, for consumer products, trust may matter more than additional raw capability once a system is capable enough for ordinary tasks.

Regulation and coordination

The conversation compares possible approaches to AI oversight, from pre-release approval and ongoing inspections to post-release liability. Wang says AI’s rapid progress can outpace government and public decision-making, while mistrust and competition among AI companies make coordination harder.

He favors establishing shared principles and processes for reassessing risks over time, and believes international cooperation—including with China—is possible. He also points to a White House-facilitated agreement in which companies set internal controls and principles, with internal verification, external auditors, and board-level committees. Wang describes transparency and trust-building as important steps, while acknowledging that cooperation is not yet happening at the level he would like.

Reviews, guides, or tutorials

No formal review or step-by-step tutorial is provided. The interview includes a brief account of the interviewer testing Muse’s phone-calling feature.

Main speakers and sources

• Cleo Abram — Interviewer and host.
• Alexandr Wang — Meta AI chief and Muse representative; discusses the product and Meta’s approach to privacy, alignment, and AI governance.
• The opening also refers generally to media coverage and download figures but does not identify those sources in detail.