AI summary by TwinMind
DeepSeek’s Lessons for Chinese AI
Asianometry
16m 42s
DeepSeek Background and Initial Recognition
- Author told in Tokyo 2024 that DeepSeek was "the real deal" - cracked team, only significant Chinese AI company
- December 2024: DeepSeek released V3 base model with impressive efficiency
- Few AI experts initially impressed
R1 Model Release and Market Impact
- January 22, 2025: DeepSeek released reasoning model R1
- Works like OpenAI's O1 and O3 models
- Takes extra compute time for better answers
- January 23: New York Times article focused on V3 training costs
- Chinese engineers needed only $6 million in raw computing power
- 10x less than Meta spent on latest AI technology
- January 25: CNBC published 40-minute YouTube video with nearly 5 million views
- Monday market crash: Tech/semiconductor markets lost $1 trillion market value
- NVIDIA lost 17% or $600 billion - biggest single day loss in history
DeepSeek Technical Innovations
- Used multiple efficiency tricks including several new innovations
- Data distillation from larger models (GPT-4, Claude) - fairness concerns but unavoidable
- 8-bit floating point numbers vs industry standard 32-bit
- Mixture of experts structure innovations
- Many innovations traced to limitations of NVIDIA China-specific hardware
- Assembled all tricks into one slim model (contrast with OpenAI's complex model selector)
- Most tricks known/implemented by other companies previously
- Mistral using mixture of experts since December 2023
Key Technical Breakthrough - Chain of Thought
- DeepSeek managed to endow R1 with chain of thought reasoning without human in the loop
- Simple idea that people tried before but failed
- Required "balls" to throw big model resources behind unproven approach
DeepSeek Leadership Philosophy
- Founder Liang Wenfeng focuses on team building and innovation capability
- Wrote preface to Chinese version of Jim Simons biography "The Man Who Solved the Market"
- Interested in building lasting successful teams
- DeepSeek's MO: team capable of generating innovations, not just current innovations
Chinese Software Company Structure vs DeepSeek
- Typical Chinese software companies (ByteDance, Pinduoduo):
- Huge dev teams: dozens to hundreds of people
- Only ~50% developers, rest "operation staff" for labor-heavy tasks
- Big QA teams (vs US where developers QA own code)
- Top-down management with 20-30+ person meetings
- AI researchers guided by OKRs and KPIs written quarterly
- ByteDance has leaderboard refreshing regularly
- Anxiety when models can't keep up with top Chinese models
- Publishing freeze (same as US companies)
DeepSeek's Contrarian Approach
- Small teams with little hierarchy
- Flexible direction based on passion
- Workers allowed to publish with names attached
- Time, resources, and courage for long-term insights
- Bucks Chinese convention
Hiring and Team Philosophy
- Focuses on hiring for ability over credentials
- Example: Top 2 salespeople were outsiders (German machinery exporter, securities firm backend coder)
- No experience/resources/connections initially
- Sales team achieved nothing first year, results in second year
- No KPIs or quotas - unusual patience in China/US
- V2 team has no "sea turtle" (overseas-trained Chinese)
- Hires from Peking/Tsinghua universities
- Avoids hiring famous/already successful people
- "Those successful carry burden of not being allowed to fail"
- "Better to leave innovation to newcomers"
Scaling Challenges for DeepSeek
- Gone mainstream: Chat app hit #1 in many countries (something Anthropic never achieved with Claude)
- Dylan Patel mentioned DeepSeek trying to raise funds
- Going head-to-head with OpenAI in prosumer chatbot market
- Scaling up means adding product managers, marketing, sales - diluting research culture
- Closing gaps always feels fast; gains become harder in uncharted territory
- Precedent: Anthropic caught up to OpenAI but struggled to move past them
High-Flyer Hedge Fund Precedent
- DeepSeek's predecessor hedge fund succeeded with AI system from 2015
- Worked until 2021 when returns lagged due to higher assets under management and competition
- Liang apologized to customers via WeChat end of 2021 (covered in Wall Street Journal)
- Founded DeepSeek in 2023
Talent Poaching Issues
- Other Chinese AI teams likely offering DeepSeek employees irresistible offers
- Already happening: Xiaomi founder Lei Jun recruited Lai Fei (genius behind DeepSeek V2)
- $1.3 million salary to run Xiaomi's LLM team
- China produces 4-5 million STEM graduates annually (Croatia/Panama's worth)
- DeepSeek will lift stigma of purely domestic hires
Hardware Limitations and Geopolitical Impact
- Liang: "Money never been problem, bans on advanced chips are the problem"
- DeepSeek's hosted models running inference on Huawei Ascend 910c hardware
- DeepSeek controversy will motivate US to further tighten NVIDIA chip exports
- NVIDIA's China revenue likely going to zero
- Chinese AI companies forced to turn to Huawei accelerators
- Will strengthen Huawei SMIC industrial complex and semiconductor proficiency
- Potential TPU-like advantages for Chinese AI companies over American cohorts
Broader Industry Implications
- Popular Western sentiment that US dominates software is wrong
- China has world-competitive software industry
- Chinese are world beaters in face recognition, machine vision, other AI fields
- Chinese see themselves in war situation
- Similar to Chinese EV industry mid-late 2010s with multiple teams spending irresponsibly
- American AI labs have too much confidence in massive capex, productization lead, government relations
- Author urges American companies to push harder
- DeepSeek first Chinese lab from left field but won't be last
