AI summary by TwinMind
Did DeepSeek burst the AI bubble or will the Stargate Project make it worse?
Sabine Hossenfelder
7m 32s
Stargate Project Announcement
- January 21st: Trump announced $500B AI investment project
- Named after 1994 movie Stargate (symbolizing wormhole to ancient pyramid)
- Private investor umbrella, not governmental project
- Initial investment: $100B, target $500B by 2029
- Investors: OpenAI, SoftBank, Oracle, Abu Dhabi-based MGX fund
- Elon Musk criticism: claimed they don't actually have $100B
DeepSeek's Disruptive AI Models
- Chinese firm DeepSeek released two open-source models:
- R1: reasoning model similar to GPT-4o
- Janus Pro: image generation model
- Neither best in category but among top performers
- Cost efficiency claims:
- $6M development cost vs OpenAI's $100M+
- 2,000 GPUs vs OpenAI's 100,000 GPUs
- Built on Meta's open-source Llama
- Likely trained on GPT output
- DeepSeek attributed success to multiple small efficiency improvements
- Debate over number accuracy:
- $6M figure for V3 training, not more powerful model
- Expert consensus: efficiency gains broadly plausible despite exact number disputes
Market Impact and Reactions
- NVIDIA stock plunged 17.7%, wiping out $500B+ market value
- Gary Marcus compared OpenAI to WeWork (bankrupt 2023 office leasing company)
- Marcus previously gave OpenAI "Theranos vibes"
AI Bubble Analysis
- Current AI companies hugely overvalued but bubble not ready to burst
- NVIDIA decline due to devalued American AI companies (their chip customers)
- Lower training costs will open market to more players
- Stock market recovery expected - ultimately good for technology/economy
Bubble Burst Prediction
- Will burst when large language models proven unprofitable
- Experts (John Lecun, Gary Marcus) doubt LLMs achieve general intelligence
- Market will react once LLMs can't be rescued with further updates
- Money will evaporate rapidly at that point
Strategic Implications
- Stargate project comparison: like spending $500B on first semiconductors instead of waiting for market development
- Similar to dot-com bubble but with "hallucinating chatbots" instead of free t-shirts
- DeepSeek demonstrates value of building on American R&D investments
- European "wait and see" approach may prove advantageous
- Risk aversion strategy: let Americans do heavy lifting, then build on knowledge
Brilliant.org Sponsorship
- Courses on neural networks, large language models
- Interactive visualizations, follow-up questions
- Topics: coding, Python, computer memory, quantum mechanics
- Host's quantum mechanics course available
- 30-day free trial, 20% off annual premium subscription
