In December 2025, shortly before Manus was reportedly acquired by Meta, Peak Ji (English name: Peak; Manus co-founder and Chief Scientist) did a long interview. It was a dense 3.5-hour talk—tens of thousands of words. Even at 1.25x speed I struggled to keep up. His clarity and speed of thinking are impressive, and his English is solid too.

I took extensive notes and extracted the points below, focusing on technology and product rather than politics. Many lines are paraphrased, with some quotes preserved.

1. Background and Three Major Startup Phases

1) Early background

• Grew up between two styles: a scientist father (PKU physics professor) and an entrepreneurial mother (Zhongguancun serial entrepreneur).
• Not self-described as a "genius student"; more of a tinkerer with strong interests and uneven subjects.
• 2009 App Store was a turning point: it provided a way to prove that "tinkering" could create real economic value.

2) Startup #1: Mammoth Browser

• Built a third-party iOS browser in high school.
• A "copy" sales model: simple fixed price per sale; low maintenance.
• Reportedly earned ~$300k+ total, which is meaningful for a high schooler.
• Ended due to iOS compatibility changes and a desire to pursue more interesting work (NLP).
• View: browsers are rarely a good disruption startup category; they fit incumbents with distribution.

3) Startup #2: Magi (semantic search + knowledge graph)

Motivation: traditional search UX with "10 blue links" would not work for small screens and voice interfaces.

Key ideas and work:

• Word2vec (Mikolov, 2013) was described as a mind-blowing inflection: dense vectors enabled applying ML/DL methods to NLP.
• Built Open IE-like extraction to continuously extract entities/relations and build a knowledge graph.
• Tackled long-context issues early; reportedly reached ~16K context in 2018, ahead of the era.
• Product name Magi (from EVA), aimed at lifelong/continuous learning knowledge graphs.

Why it failed:

• Underestimated the data moat of search engines and the flywheel of data sources.
• Wearables and new interfaces arrived too late; ChatGPT was the real inflection.
• Engineering execution was strong (crawler, indexing, infra), but non-technical barriers were decisive.
• GPT-3 early access was a shock: prompt-based generality competed with many vertical pipelines.

4) Startup #3: Manus (general-purpose agent)

After an interim period (including building LLM work inside a unicorn and time as EIR), he joined Manus with the CEO Xiao Hong and other co-founders. Monica (an existing product) provided cash flow and business stability.

2. What Manus Is (and What It Is Not)

1) The "AI-native browser" attempt

They tried to build an AI-native browser:

• On-device models for cost/privacy, with AI controlling automation flows.
• They abandoned it after seeing:
  • browsers are inherently online, so offline-on-device obsession is misguided
  • small models (around 3B) lose too much quality vs. flagship cloud models
  • "AI takes over your computer" creates weird control conflicts with users
  • short tasks: humans are faster; long tasks: require keeping the machine awake
  • browser switching has huge migration friction (Arc example)

2) The Manus form factor

The core insight came from seeing non-engineers use Cursor:

• They were not "coding"; they were conversing with AI and letting code be a medium to solve non-coding tasks.
• Programming is not just a vertical skill; it is a general-purpose medium.

So they "moved the browser to the cloud":

• run agents asynchronously and concurrently in cloud sandboxes
• hide code and complexity for prosumers (knowledge workers, not necessarily programmers)
• name Manus from Latin "hand" and MIT motto "mens et manus"

3. Architecture and Key Bets

• Model strategy: do not build a foundation model too early; focus on context engineering and the agent framework, and leverage strengths across multiple models (coding, multimodality, reasoning).
• Sandboxing: lightweight full virtualization (Firecracker), isolated disposable sandboxes per session, supporting both Linux and Windows.
• Long context: context window length is not the only answer; compression awareness and offloading to file systems matter.
• MCP stance: cautious; MCP can pollute action space and increase costs. Prefer ways that keep the native action space clean.
• "Pure" agents: avoid over-constraining with rigid rule workflows; let intelligence solve tasks, including visual self-checking (layout, fonts, overlap) via observations.

4. Product Strategy and Metrics

• General vs. vertical: general agents can cover more scenarios; vertical is "adding constraints".
• Long tail: "long tail" can be daily work for some users; Manus can handle obscure formats by pulling tools from GitHub, etc.
• Cross-scenario synergy: in one session, an agent can research, build a site, set up a database, analyze traffic, generate slides, draft outreach emails—creating internal network effects.
• Data flywheel: user corrections teach the system; subjective evaluation is critical (and can be a moat).

Commercial notes (as described):

• Pricing moved from subscription + top-up complexity toward flexible subscription levels.
• Agent workloads consume far more tokens than chatbot workloads (orders of magnitude).

5. Industry Views

• The boundary between "model companies" and "app companies" may blur; top apps will build model capabilities and vice versa.
• SOTA half-life is short; differentiation comes from system-level execution and taste, not only model weights.
• Agent adoption will expand; operating systems will grow agentic capabilities rather than a separate "agent OS" winning alone.
• AI businesses behave more like manufacturing: higher fixed cost, more linear cost with usage, demanding operational discipline.

6. A Few Memorable Lines

• "For every complex problem there is an answer that is simple, clear and wrong."
• "Everything added dilutes everything else."
• "The monopoly of AI products is the monopoly of a mindshare, not the whole market."
• "Building agents is about doing 1000 small things right, not 3 big things."

This is still only a partial summary of a very dense talk, but it captures the themes I found most valuable: the shift from chatbots to agents, context engineering as the real craft, and the importance of system-level taste and evaluation.

Previous pageHow Does a Philosophy Professor Think About AI?

Next page2025 - A Year of "First Years"