Jan 29, 2025 · 51m · startup-ideas
DeepSeek R1 - Everything you need to know
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Ex-Apple engineer Ray Fernando joins Greg Isenberg to deliver a practical breakdown of DeepSeek R1, detailing prompt engineering techniques, private cloud and local hosting configurations, and the transformative economic impact of open-source reasoning models.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Greg holds 12.4% of the talking time here. How this is scored →
Greg as informed peer 3.6 Guest teaching 6.3 Guest disagreement 0.4 Greg pushing back 0.4
speaking balance: gold is Greg, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 2:15 Caveat on Chinese data sovereignty
Ray directly warns listeners about data sovereignty and regulation risks when transmitting sensitive company information directly to servers hosted in China.
Hardest push from Greg ▶ 17:31 Warning on compounding token costsGreg challenges the casual assumption that token costs are negligible, pointing out how ongoing business workflows rapidly accumulate expenses.
Biggest teaching moment ▶ 23:30 Local LLM deployment masterclassRay provides a detailed technical demonstration of spinning up Docker containers and pulling Ollama models to run private inference locally on a Mac.
Greg holds their own ▶ 13:02 Evaluating reasoning models for startup competitivenessGreg demonstrates his entrepreneurial expertise by articulating how reasoning models replace extensive human labor and create competitive moats for founders.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Greg as informed peer | Guest teaching | Guest disagreement | Greg pushing back | Why |
|---|---|---|---|---|---|---|
| Navigating Data Privacy Concerns and Prompting DeepSeek | 4 | 5 | 1 | 1 | Ray details the privacy concerns of sending data to Chinese servers when using DeepSeek directly. Greg contributes by emphasizing practical privacy boundaries, noting he would not upload tax documents and highlighting Perplexity's US hosting. | |
| Sponsor Spotlight: Startup Empire Membership | 1 | 7 | 0 | 0 | After Greg's brief sponsor read, Ray walks through model distillation, parameter sizes, and how to avoid Chinese routing using Open WebUI and Grok or Fireworks APIs. Greg remains entirely passive during the instructional walkthrough. | |
| Economic and Strategic Advantages of Reasoning Models | 6 | 5 | 1 | 1 | Greg frames how reasoning models provide startups with an unfair advantage, comparing DeepSeek's structured output to senior human copywriters and pressing on the cumulative cost of token consumption. Ray concurs and details specific API pricing comparisons against OpenAI o1. | |
| Prompt Refinement and Web Search Fact Verification | 2 | 6 | 0 | 0 | Ray illustrates how to refine prompts using OpenAI's playground and demonstrates automated fact-checking on Marc Andreessen's manifesto. Greg observes before asking to see the local deployment workflow. | |
| Local DeepSeek Installation with Docker, Ollama, and Open WebUI | 2 | 8 | 0 | 0 | Ray gives a comprehensive step-by-step masterclass on running DeepSeek locally via Docker containers, Ollama model pulls, and configuring system temperature parameters. Greg acts as an attentive learner following the technical tutorial. | |
| Human-Centered AI Interface Design and API Integration | 5 | 6 | 1 | 1 | Greg contributes product design insights by reframing model temperature as 'wine versus coffee mode' and critiquing the lack of playful interfaces in AI products. Ray agrees and continues educating on model quantization levels and API backend setups. | |
| Running Local Reasoning Models on Mobile with the Apollo App | 3 | 7 | 0 | 0 | Greg initiates a query about mobile inference, leading Ray to screen-mirror his phone and demo local reasoning on Apple Silicon using the Apollo app and OpenRouter. Ray shows the model running completely offline. | |
| Exploring Multimodal Intelligence, Audio Analysis, and AI Startups | 6 | 6 | 0 | 0 | Greg pitches practical startup use cases such as real-time audio analysis for emergency detection and negotiation assistance. Ray expands on the concept by explaining the distinct capabilities of native audio multimodal models like GPT-4o. |
Statements from this episode (8)
Fernando: DeepSeek R1 Is Open Source and on Par with OpenAI o1
“Like they've made it open source so that it's available for us to study but it's apparently also on par with ChatGPT's like O-one model, O-one's reasoning models.”
Fernando: DeepSeek's official apps and website route user data to China
“To start out to use these models, you have a couple of options and one is going directly to deep seek.com and this is actually currently hosted in China. So a little bit of a background here is that your computer is here. Like it was, for example, I'm in North…”
Fernando: Cursor serves DeepSeek via Fireworks API, avoiding China-hosted servers
“One of my favorite apps for coding is actually Cursor. And I asked them, hey, where do you have your DeepSeq model hosted? And they told me they use the Fireworks API. And that's, you know, actually not in China. So that's great. So it's like, okay, cool. That…”
Fernando: DeepSeek via Fireworks costs $8/M tokens versus o1's $75
“I think it's about eight dollars per million tokens where normally I think ChatGPT was like 15 input and 60 dollars for output for O-one Pro.”
Fernando: OpenAI o3 launch will likely trigger reasoning model price drops
“Open AI is currently promised that the O three model will come out and the mini model will come out, which would be on par with this model. So that prices will also probably significantly drop as well because they just get more efficient with time.”
Fernando: DeepSeek web search is unavailable through third-party APIs like Fireworks
“The fireworks and the Grok are specific API endpoints. And right now there isn't like a specific web search thing that's currently tuned into them. So if you want to do web search, you have to go through the deep seek.com route or the app”
Fernando: GPT-4o natively understands audio and tone, unlike DeepSeek R1
“So one thing I just learned very recently about GPT-IV and ChatGPT's Omni models is the fact that this model's breakthrough, a little bit different than R-I, is the fact that it can actually understand audio and tone and all these extra implications that we do…”
Fernando: OpenAI's upcoming o3 model will take the industry by storm
“And so it's going to be really exciting when they actually dropped O three. I think a lot of people are going to be taken by storm of like, what's actually really going to come out from them. It's going to be a really, really big leap.”