AINEWS
Liang Wenfeng's leaked four hour investor Q&A — hours after DeepSeek closed a $7.4 billion raise — reveals what gets cut from DeepSeek's AGI mainline, how the lab actually runs, and why its founder says compute, not will, is the ceiling.
DeepSeek, one of the most-watched Chinese AI labs and the team behind some of the most-cited open-weight models of the past year, runs on an operating system that looks like neither a Silicon Valley startup nor a corporate R&D division. Liang Wenfeng, the company's founder, sketched the contours of that system across a roughly four-hour investor Q&A whose recording leaked onto Chinese tech channels and was first synthesized in analytic form by 36kr.
The recording, cross-checked against parallel transcript republishes on elsewhere.news and hellochinatech.com, gives the first sustained public look at how a top-tier Chinese AI lab prioritizes its way toward AGI. Three things stand out from the Q&A: a strictly filtered technical mainline, a deliberately chaotic research process, and a hard ceiling that is not ambition but compute.
What gets cut. Liang treats 3D generation, video generation, and "world models" as "not on the mainline" of intelligence, according to the leaked recording. Multi-modality, in DeepSeek's telling, is a "component" layered on top of the mainline, not the mainline itself. That is a deliberate subtraction at a moment when most frontier labs are racing to ship image, video, and 3D generators and when the loudest product launches of the year have all been multimodal. The same hierarchy governs DeepSeek's AGI roadmap, which the 36kr analysis decomposes into six concrete sub-steps: language model → chain-of-thought (step-by-step reasoning traces) → Agent → continual learning (a model that keeps updating from new experience rather than freezing after pre-training) → self-iteration → embodied. Agents, despite being the loudest product race of 2026, are a stepping stone. The bottleneck Liang flags after Agents is continual learning.
How the org actually works. Liang describes an organization with no fixed architecture, and says that roughly half of every researcher's time goes to unstructured, self-directed work. He calls low-bar exploration "摸奖", a phrase that translates roughly as lottery-ticket research. A small bet might land a big capability or land nothing at all. The contrast with a Western frontier lab is sharp: no quarterly OKRs, no shipped-product targets, no public benchmark ladder to climb each month. Several of these descriptions — including the absence of org structure and the roughly 50% unstructured-time claim — are Liang's own characterization of the lab, not independently verified. They describe the operating model DeepSeek's founder wants to project as much as the one it runs.
The ceiling is compute, not will. Asked about the upper limit of model size, Liang gave a sharper answer than the 36kr headline suggests. "We have not touched the upper bound," he said, according to the leaked transcript. "I scale according to my resources. The reason I stop is not that this model is enough." Compute, in Liang's view, is the binding constraint, not conviction or capital. That admission re-prices the next 18 months of the AGI race: every "we'll get there first" forecast now has to be cross-checked against who actually has the silicon, not just who has the talent or the model. The race, on this telling, is a race of teraflops.
The unglamorous work. DeepSeek's research staff still does heavy data labeling and post-training work. Liang concedes that about half of core researchers are currently tied up in labeling, gated by the cost of high-quality data. The lab also treats hallucination as a "product problem" solvable through better post-training, not a frontier research problem. The AGI ladder is built out of unglamorous steps as much as it is built out of the ones that look good on a roadmap. The most interesting frontier model of 2026 is, by DeepSeek's own account, partly carried by the same kind of human data work that powered the first generation of supervised systems.
A lab running on lottery-ticket exploration and giving half its researchers unstructured time will be tested by shipping cadence, not by benchmark climbs. The compute ceiling Liang names is the next falsifier: every DeepSeek release in the rest of 2026 will be, in effect, a small bet against the silicon the lab can actually secure.