DeepSeek: What Actually Matters (for the everyday user)
About this video
Drawing heavily on Ben Thompson’s DeepSeek deep dive, Jeff Su busts 10 myths for everyday AI users. The claimed $5.6 million cost covered only the final training run, excluding infrastructure such as a reportedly 50,000 Nvidia Hopper GPUs worth around a billion dollars. DeepSeek innovated within export controls using the less powerful H800s rather than the restricted H100s. Its reasoning model R1 matches OpenAI’s o1 on performance but leads on efficiency, not overall capability, since OpenAI had already demonstrated the more powerful o3 and then released o3 mini free — arguably under DeepSeek pressure. Comparisons must be like-for-like: base model V3 compares to GPT-4o, while R1 is a reasoning model. R1’s visible chain of thought is a UI choice, not a technical breakthrough, and DeepSeek allegedly used model distillation from ChatGPT outputs — not illegal, he says, but against OpenAI’s terms of service. On safety, the native app sends and stores data in China, but users can access the models via Perplexity, Venice AI, or Cursor, or run them locally with Ollama or LM Studio. He invokes Jevons Paradox to argue efficiency may increase demand for Nvidia chips, says Amazon, Apple, and Meta may benefit, and likens the moment to Google’s 2004 infrastructure openness rather than Sputnik. His advice: enjoy free access to R1 and o3 mini, protect your data, and do not switch tools just because something is trending.
Key takeaways
- The $5.6 million figure was only the final training run; it excludes infrastructure like the reportedly 50,000 Hopper GPUs worth around a billion dollars.
- DeepSeek optimized for efficiency within export controls on H800 chips; R1 matches o1, but OpenAI’s o3 is more capable and o3 mini was released free.
- Compare like with like — V3 is the base model comparable to GPT-4o, R1 is the reasoning model — and visible chain of thought is a presentation choice, not a capability breakthrough.
- If data stored in China concerns you, use DeepSeek models through Perplexity, Venice AI, or Cursor, or run them locally with Ollama or LM Studio.
- Do not pay a switching tax for hype: switch only if DeepSeek gives a clear advantage for your specific needs and workflow.
Episode details: published February 3, 2025 · 13:23 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are Jeff Su’s own claims, shared for education — they are not typical or guaranteed.
← Previous: You’re Not Behind: Become AI-Native (without the overwhelm) · Next: I Switched 50% of My AI Work to Claude, Here’s Why →
All Jeff Su videos · Jeff Su profile · Jeff Su (official site)
