YouTube Review

Liang Wenfeng's 2023 DeepSeek AGI Interview

Techonomic China Insider's video is useful only if its source type stays visible. The YouTube description says the material comes from a Waves interview dated May 24, 2023 and was translated and adapted for educational purposes. The Elsewhere translation of the Waves article gives a clearer textual anchor: it presents Liang Wenfeng and High-Flyer at the moment the firm had named its independent large-model venture DeepSeek and was explaining why a quantitative hedge fund would pursue AGI.

The video is therefore not best treated as original footage. It is a narrated, translated packaging of an early founder interview. That makes it weaker as evidence for exact wording, tone, and production context, but strong as a public artifact of the founding doctrine later readers used to interpret DeepSeek. It belongs beside Liang Wenfeng, DeepSeek, Open-Weight AI Models, AI Compute, DeepSeek Founder Interview, and OpenAI is Not God.

Founding Doctrine

The strongest signal is that Liang frames DeepSeek as a research-first AGI project before it has the public identity that V2, V3, and R1 later gave it. He says the effort is not a finance vertical and not primarily an application company. The goal is general artificial intelligence, with language models as a likely path and later expansion toward vision. That matters historically because it shows the organization presenting itself as a lab before it became a symbol of open-weight reasoning and compute-efficient frontier competition.

The interview also explains why venture-capital pressure was a poor fit. Liang says research is expensive because it requires experiments, comparisons, compute, and skilled people, while pure replication can ride on public papers or open-source code. High-Flyer could supply compute and an engineering team, which he describes as already providing part of the required capital. The Spiralist lesson is institutional: who can pursue long-horizon AI research depends not only on model ideas, but on patient capital, compute reserves, and tolerance for failure before product-market proof.

Compute and Openness

The compute story is the second major contribution. The interview traces High-Flyer's GPU growth from small research setups to 100 GPUs in 2015, 1,000 in 2019, and roughly 10,000 A100-class cards by the 2021 Firefly 2 era. Elsewhere's translation similarly describes Firefly 1 and Firefly 2 as large High-Flyer compute investments, with Firefly 2 at roughly 10,000 NVIDIA A100 GPUs. Read cautiously, those numbers explain why DeepSeek was not a normal app startup entering the model race from zero.

The openness claim also appears early. Liang says the company hoped to make most training results public so that small applications could use large models at low cost rather than leaving the technology concentrated in a few companies. DeepSeek's later R1 release shows the path that claim eventually took: an MIT-licensed release, open distilled models, and a public argument that R1 could compete with OpenAI-o1-style reasoning. The 2023 interview does not prove the later technical result, but it makes the later release strategy less surprising.

DeepSeek's V3 technical report and R1 paper supply the retrospective technical evidence. V3 describes a 671B-parameter mixture-of-experts model with 37B active parameters per token, 14.8T pretraining tokens, and a reported 2.788 million H800 GPU-hours for full training. R1 describes reasoning-model training through large-scale reinforcement learning, cold-start data, multi-stage training, and distillation. The 2023 interview is not those papers. It is the organizational preface to them.

Innovation Culture

The final theme is management. Liang argues for hiring on ability, creativity, passion, and foundational skill rather than prior experience; he says inexperienced people may experiment more and find solutions suited to the current situation. He also describes an innovation culture with few KPIs, minimal intervention, room for mistakes, and important tasks given to people who can grow into them. The rhetoric is romantic, but it is also operational: a research lab's output depends on the kinds of people it recruits and the constraints it does or does not impose.

That theme should be read with discipline. "No KPIs" is not a magic recipe, and lack of management can hide failure, waste, or uneven accountability. But the interview helps explain DeepSeek's later public image: a Chinese lab with a quiet founder, a young technical team, a compute base outside the usual internet giants, and a release strategy that challenged assumptions about who could produce frontier-adjacent models.

Evidence and Limits

This review is published on July 1, 2026. The video was uploaded on January 28, 2025, but the source interview dates to May 24, 2023. That chronology matters. It is pre-V2, pre-V3, and pre-R1; it should be read as a founding-position document, not as proof of later capability. The channel's transcript includes obvious translation and transcription artifacts, including "Phantom fund" where the source context points to High-Flyer, so exact phrasing should be checked against the Waves/Elsewhere text rather than treated as a verbatim record.

The useful conclusion is that DeepSeek's later open-weight shock had an earlier institutional story: High-Flyer compute, curiosity-driven AGI research, low-cost access rhetoric, resistance to platform monopoly, and a culture that prized young technical talent over inherited playbooks. None of that proves DeepSeek's later claims by itself. It explains the doctrine that made those claims coherent when V3 and R1 arrived.

Return to YouTube