140: AI for Science, From the Beginning to Now | Dialogue with Zhang Linfeng and Sun Weijie from DeepWisdom

1. Key Themes

Theme 1: From First Principles to AI-Accelerated Scientific Computing

DeepSeek was founded on the breakthrough of using machine learning to dramatically accelerate quantum mechanical calculations. Zhang Linfeng describes the founding moment: "When we actually did this with a notebook, we could beat supercomputers - achieving over 6 orders of magnitude acceleration. In the morning I could discuss with my advisor, change calculations in the afternoon, and continue discussions the same day." [00:31:00]

The company's core insight was applying deep learning to solve the "ab initio" problem - calculating molecular and atomic interactions from first principles using Schrödinger equations. Sun Huijie explains: "All matter in the world is composed of atoms and molecules. If we can clearly understand the interactions at the atomic and molecular scale, we can theoretically solve all materials, pharmaceuticals, and chemical reactions." [00:49:50]

Theme 2: Building the "Microscopic Dassault" - Industrial R&D Platform

DeepSeek positioned itself as creating microscopic-scale industrial R&D tools, analogous to how Dassault Systèmes revolutionized aerospace with CAD/CAE software. Sun Huijie articulates this vision: "In 2020, we proposed a 5-year plan to build a complete microscopic industrial R&D platform... Past industrial software utilized solid mechanics, fluid mechanics, electromagnetics well, but quantum mechanics - humanity's ultimate science dream - hadn't been well utilized." [00:54:10]

The strategic decision to remain a platform company rather than pursue specific drug pipelines reflects this vision. Sun explains: "If we serve specific stage needs with the same technology platform, our platform could serve 1,000 or 10,000 customers, enabling rapid iteration and self-growth. This better aligns with our original intention - to be a technology company originating from China that leads the world." [01:30:46]

Theme 3: The Four Waves of AI for Science Evolution

The company identified and capitalized on four major technological shifts:

1. Machine Learning for Physics (2017-2020): Initial breakthrough with DeepMD for molecular dynamics
2. Pre-training Models (2021+): Development of Uni-Mol (molecular), Uni-Fold (protein), and atomic foundation models
3. Large Language Models (2023+): Integration with BioNavi/SciMaster for scientific knowledge access
4. AI Scientists/Agents (2024+): Current frontier toward autonomous scientific discovery

Sun Huijie notes: "AI Agents will change not just the three basic scientific production tools (computation, databases, experiments), but the fourth production element - humans themselves. Past waves improved tools; this wave makes every scientist have an AI assistant at the professor level." [01:11:00]

2. Contrarian Perspectives

Perspective 1: Choosing China Over Easier Silicon Valley Path

Despite having opportunities to relocate abroad with significantly more resources, the founders chose to build in China. Sun reveals: "We've had opportunities to become an overseas company... with promises of far more resources than we have now. But looking back, I feel our determination to build this company with its mission was something worth remembering." [01:22:30]

Zhang adds context about their advisor's influence: "Weinan E told me three things I'd never thought about: You should graduate now, return to China, and continue this work through entrepreneurship." [00:35:08] This was contrarian given Zhang's position at Princeton and typical career trajectories.

Perspective 2: Platform Over Pipeline Despite Resource Constraints

When facing 100x less funding than competitors like Schrödinger and DeepMind, DeepSeek chose the harder platform path rather than pursuing specific drug pipelines that would offer clearer near-term returns. Sun explains the trade-off: "Pipeline companies can achieve hundred-billion valuations, and platform companies can too. But platform companies serving everyone's information access needs represent a different kind of value. The core question was which resonates more strongly with our inner calling." [01:31:00]

Perspective 3: Scientific Paradigm Shift Is Inevitable and Imminent

Zhang offers a provocative prediction about Nobel Prizes: "Can we ask from another angle - looking forward 10 or 20 years, will there be any Nobel Prize that exists separate from AI? I don't think any future Nobel Prize will be disconnected from AI - they'll all be based on AI-accelerated science, just applied in different fields." [01:33:50]

Sun adds: "In 10-15 years, I believe cancer becoming a manageable chronic disease like cardiovascular disease is hopeful... I'm more optimistic than Elon Musk or Sam Altman about these predictions." [02:03:00]

3. Companies Identified

Schrödinger (Reference/Competitor)

Description: Established computational drug design company using classical molecular dynamics Quote: "Schrödinger does what Linfeng wrote as the third equation - using previous generation empirical molecular dynamics to calculate drug-protein interactions for drug design. Their software and services have become very mature overseas." [00:44:50] - Sun Huijie

Dassault Systèmes (Analog/Inspiration)

Description: Industrial software giant that digitized aerospace/automotive design Quote: "Dassault put all airplane and car blueprints into computers... using solid mechanics, fluid mechanics, electromagnetics, and optics to simulate whether planes and cars can run safely. This became critical discovery in 2020 - a huge industrial opportunity with deep scientific principles at the intersection." [00:51:20] - Sun Huijie

AlphaFold/DeepMind (Competitor/Inspiration)

Description: Google's protein structure prediction breakthrough Quote: "When AlphaFold came out, we felt this was definitely Nobel Prize-level achievement, just a matter of time. We didn't expect it so fast... but AI for Science had already seen the opportunity - if you just pushed forward without everything else." [01:36:40] - Sun Huijie

Meta (Competitor)

Description: Developing competing foundation models for molecules and atoms Quote: "Our atomic large model's best competitors besides DeepMind should be Meta/Facebook. Our gene large model Uni-RNA competes with DeepMind's EvoForge. Molecular large models also compete primarily with DeepMind and Meta - all are our competitors, and currently we're still in the first tier." [01:15:20] - Sun Huijie

4. Operating Insights

Insight 1: Forced Innovation Through Resource Constraints

Rather than viewing limited resources as purely negative, DeepSeek used constraints to force methodological innovation. Sun explains: "Facing competitors with 100x or even 1000x our resources, how do we unite limited forces? We need more clever R&D approaches to maximize our success rate and produce desired results." [01:25:40]

This manifested in creative approaches like competing for prize money: "Besides financing, we participated in competitions because they had prize money - this counted as the first bucket of gold. The timing was particularly coincidental - the day after Linfeng's wedding, we had a major project defense. We fortunately won that disruptive innovation competition - 120 million yuan over 3 years." [00:47:30]

Insight 2: Building Cross-Disciplinary Talent From Scratch

Unable to hire ready-made interdisciplinary talent (physics + chemistry + computer science + business), DeepSeek created internal training systems. Zhang describes: "We found no single major or education stage produced ready people. Those studying chemistry don't know algorithms, those knowing algorithms can't do software engineering. So we had to train people ourselves from sophomore/junior year." [00:58:05]

This extended to creating their own educational infrastructure: "AI4Science versions of Colab, Kaggle, Coursera etc. - all needed to be built because they were unfamiliar to people in chemistry, materials, physics, computational science, astronomy, geology fields at that time." [01:01:40]

Insight 3: Dynamic Founder Role Allocation

Rather than rigid CEO/CTO divisions, the founders maintain fluid responsibilities based on capabilities and needs. Sun characterizes their difference: "Linfeng is externally cold but internally hot - hard to approach but full of passion inside. I'm externally hot but internally cold - easy to approach but my inner principles and judgment bar is quite high." [01:53:00]

Insight 4: Embracing the "Best Last One" vs "First One" Philosophy

Zhang articulates a counterintuitive success metric: "Nobel Prizes reward being the first. But business should aim to be the best last one - when AI does everything on this platform. If you can drive a paradigm shift, a good platform might be the best in all three dimensions: first to accelerate discovery, last one everyone uses (network effects), and best product-market fit." [02:00:20]

5. Overlooked Insights

Insight 1: The "Ab Initio Paradox" - Education Undermines Innovation Capability

Zhang reveals a profound educational problem that few discuss: "When I learned general relativity in sophomore year and got good grades, I later realized math students only learned Riemannian geometry in junior year. Something was wrong - we only accepted a series of rules, could make derivations from these rules, but missed the actual discovery process that Einstein went through." [00:02:40]

This connects to his later observation: "My biggest regret is I already knew how to derive it, so I lacked the opportunity to derive it from scratch myself - I lacked the possibility of making this kind of innovation... I happened to solve one problem as an undergraduate originally because I didn't know, but when I learned three months later, I no longer had this opportunity." [02:00:45]

Significance: This suggests current scientific education may be systematically undermining innovation capability by teaching finished theories rather than discovery processes. As AI takes over more routine scientific work, this educational flaw becomes critical - humans may lose the muscle memory of genuine discovery before AI achieves it.

Insight 2: The Quiet Geopolitical Science War

Zhang subtly reveals an underreported trend: "Science knows no borders, everyone says, but this assumption is becoming increasingly questioned. Since AI is driving science, and underlying AI models increasingly have this problem... Even at Princeton, I discovered projects we thought we'd do were having funding stopped." [01:24:00]

Sun adds: "Harvard and Princeton have many very famous, internationally influential professors whose funding has been completely cut." [01:25:05]

Significance: This suggests an unreported but systematic breakdown of international scientific collaboration, particularly affecting ethnic Chinese scientists, that predates and exceeds public discussion around export controls. The founders' decision to build in China may reflect not just patriotism but recognition of this geopolitical reality reshaping the scientific community. This context makes their achievement of maintaining "global leadership" from China even more remarkable and strategically important.

---

Timestamp Format Note: All timestamps in [HH:MM:SS] format reference the original Chinese transcript timestamps.