In this conversation, Xiaojun sits down again with Yang Zhilin, founder and CEO of Moonshot AI. A year after their last in-depth report likening his startup journey to “advancing toward an austere and unknown snowy mountain,” Yang still feels he’s on that mountainside. Some things are clearer now, but new questions keep emerging. He describes AI development as an infinite climb: each solved technical problem reveals the next, and that’s exactly what makes the journey exhilarating.
The most tangible leap in the past year is models moving from simple Q&A to sustained, multi-hour work. Two years ago, generating a coherent article was a struggle; today, K2 can clone a code repo, translate it into another language, fix all bugs, and run tests—all without human intervention. He attributes this to reinforcement learning and “test-time scaling.” Instead of outputting a single answer, models now produce long chains of thought, internal revision, and self‑verification. There are two flavors: the “brain in a fishbowl” approach (pure reasoning, no external interaction, like OpenAI’s o1) and the “agent” approach (interacting with tools, browsing, executing code, getting environmental feedback). Both scale up test-time compute, converting a single‑shot answer into a linear or parallel problem‑solving process. Yang stresses these are not strictly sequential; a model can advance both dimensions in parallel.
This duality also shapes product strategy. Third‑party developers once built scaffolding around base models, reverse‑engineering the training‑setup cues. Now first‑party agents (e.g., Claude Code, GPT Agent) are emerging, where model creators design tools and training environments in tandem—models are natively good at using those tools. Moonshot still focuses mostly on the model layer, but Yang sees the trend toward first‑party products as a major shift. He also reflects on OpenAI’s L1–L5 taxonomy (chatbot, reasoner, agent, innovator, organization). He sees reasoning and agent as mutually reinforcing, with the eventual goal of models participating in their own R&D (L4) and forming multi‑agent systems (L5). Yet he insists these aren’t discrete steps; all capabilities keep rising together.
On the engineering side, Yang highlights the pursuit of token efficiency. With high‑quality text data growth slowing, making the most of every token becomes crucial. He cites the Muon optimizer: unlike the decade‑old Adam, which treats each parameter element independently, Muon considers the whole matrix structure, yielding better learning per data point. Such algorithmic tricks are essential to push model intelligence when data isn’t scaling as fast.
Wrapping up, Yang draws on the philosophy of The Beginning of Infinity: problems are inevitable, but problems are solvable. AI progress mirrors a post‑Enlightenment dynamic society—each solution expands the frontier and breeds fresh challenges. He no longer sees AGI as a fixed summit; rather, it’s a direction. Just as the steam engine took decades to reshape society, AI’s deep societal impact will unfold over long timescales. For him, the reward is the climb itself—higher vistas, constant learning, and the hope that the mountain never ends.
