Here’s a summary of Liang Wenfeng’s core viewpoints, based on the transcript corrections compiled by fable from @willwangtf’s recording: From the transcript, Liang Wenfeng’s core ideas seem to form a complete worldview, which can be summarized into five main threads: 1. **Scaling Belief + Catch-Up Logic** Firmly believes that the larger the scale, the better the results—“haven’t seen the upper limit of scaling for language models.” Acknowledges that there’s a magnitude-level gap in computing power between China and the U.S. (domestic models with tens of billions of activations vs. U.S. models with 800B). Training ultra-large models is “basically unsolvable” for now, but he thinks continuing to scale within China’s current capacity is still worthwhile. 2. **Low Profit is a Natural Law** In the AI era, those who aim for high market share will be defeated by those who demand less. High profit margins “don’t align with natural laws.” Foundational model companies will eventually consolidate into just a few players. That’s why DeepSeek actively avoids raising prices and only earns “reasonable profits”—though the real mechanism is restrictive pricing: pricing based on a ten-month ROI, making third-party independent deployments unprofitable. 3. **Open Source is the Historical Direction** Closed-source monopolies “will inevitably be abandoned by history.” Open source is the inevitable path, and this year, domestic models can already replace foreign ones at an open-source scale. 4. **Technical Route Judgments** - There’s no world model in the current roadmap; multimodal isn’t a priority at this stage. - After continuous learning, there will be a gradual, continuous climb in capabilities, rather than a sudden singularity. - Hallucinations can be improved through better post-training and aren’t a core issue. - If the goal is to address labor demands, embodied intelligence is unavoidable. - Aiming to break free from NVIDIA’s ecosystem, he believes Huawei cards are usable in terms of cost-effectiveness (at the expense of 4x usage and being two years behind). 5. **People and Organizations** There’s no inherent difference between top talent in China and the U.S.; the gap comes from differences in computing power and experimental opportunities. Teams should stay disciplined and avoid chasing trends. He believes in “ordinary people doing extraordinary things.” **In one sentence:** He believes scaling hasn’t peaked yet, open source + ultra-low pricing is the correct endgame, and intelligence evolution is a gradual process. DeepSeek’s strategy is to leverage cost advantages and restraint to secure a long-term position—though the document’s author critiques this narrative for internal contradictions (talent, domestic replacements, flexibility, trillion-dollar valuation).