The Startup Ideas Podcast

Episode: DeepSeek R1 - Everything You Need To Know

Host: Greg Isenberg
Guest: Ray Fernando (Former Apple Engineer, AI streamer & founder)
Date: January 29, 2025

Episode Overview

Greg Isenberg welcomes Ray Fernando, a 12-year Apple engineering veteran and active AI streamer, for a practical and technical deep-dive on DeepSeek R1—a next-gen open source reasoning model out of China that's quickly gaining global attention. The discussion is both accessible for beginners and detailed enough for those wanting hands-on guidance, covering how to leverage DeepSeek R1 for creative and business uses, privacy considerations, various hosting options, model prompting strategies, and how to set up and run reasoning models locally and on mobile.

Key Discussion Points & Insights

1. What is DeepSeek R1 & Why the Hype?

• DeepSeek R1 is a new, open-source reasoning LLM (Large Language Model) out of China, on par with advanced models like ChatGPT-4-level in reasoning.
• Its reasoning and thinking capabilities can hit "superhuman" levels on some tasks, making it stand out.
• Open Source & Free: The model is open and freely available both for online use at deepseek.com and for local / self-hosted deployment.

“These models have now become so advanced and this specific one from DeepSeek is out of China. And what that allows you to do is... study [the model] but also, it’s on par with ChatGPT’s O1 reasoning models.”
— Ray (00:20)

2. Ways to Use DeepSeek R1

a. Direct on deepseek.com (02:32)

• Easy to start—just go to the website or download their app.
• Major Caveat: Data you input goes to servers in China, so privacy is a real concern, especially for sensitive info.

“I would be very careful as far as anything you put into this system...because it would not belong to a region you may live in or have control in.”
— Ray (02:50)

• Alternatives in the US: Use prompting services like Perplexity, Fireworks AI, or Groq (which host models stateside or in other global regions).
• Cursor (a coding assistant) was praised for hosting DeepSeek on Fireworks API, keeping data outside China.

“Cursor...told me they use the Fireworks API and that’s...not in China. So that’s great.”
— Ray (05:57)

b. Running DeepSeek R1 via APIs (07:09)

• API providers: Fireworks AI, Groq, and others allow you to host/run the model via custom setup, staying in the region of your choice.
• Benefits: Faster response, reliability, data privacy.

c. Running Locally (On Your Machine or Phone) (22:56, 37:17)

• Using Docker, Open Web UI, and Ollama, you can run R1 (and other LLMs) right on your laptop or desktop—no data ever leaves your device.
• Also possible on mobile via the Apollo app and compatible models.
• Running locally takes some setup and powerful hardware for bigger models, but small “distillation” models work anywhere—even on a plane.

3. Prompting Techniques & Chaining for Advanced Output (04:00, 14:40, 18:04)

• Prompt chaining unlocks next-level results, making the AI act like an admin or senior analyst:
  • Run transcripts through advanced prompts for analysis, summaries, blog posts, SEO optimization, verification, etc.
  • DeepSeek R1 and O1 Pro can follow multi-step, nuanced instructions with high fidelity—unlike GPT-4, which sometimes needs more human guidance.

“It’s basically like hiring an admin to go through all your stuff and make things for you.”
— Ray (04:44)

• Notable Difference: DeepSeek R1's outputs can feel “senior-writer-quality” with minimal human editing, while ChatGPT-4’s often require significant rewriting.

“What’s really...mind boggling is the fact that it almost looks...pretty human level incredible. Like a senior writer would do something like this.”
— Greg (13:54)

4. Privacy, Security, and Data Locality (05:29, 14:40, 47:51)

• Key Caution: Never enter private or sensitive data (taxes, medical records) into DeepSeek.com or any remote service outside your regulatory jurisdiction.
• Use API providers or run locally for privacy compliance (GDPR, HIPAA, etc).
• Choosing your hosting region allows adherence to different legal requirements.

5. Model Comparisons & Practical Observations

a. Distilled vs Full Models (Speed & Output):

• Full models (600B+ parameters) yield richer, more nuanced answers but are slower and hardware-intensive.
• Distilled models (smaller, faster, less resource-intensive) give simpler/shorter responses, great for quick answers or low-power devices.

b. Experimenting with Temperature (30:49)

• Temperature = Creativity:
  • Lower temp (like 0.2): More logical, precise, less “hallucination” (good for code/fact tasks).
  • Higher temp (1.0+): More creative, associative output (good for brainstorming, creative writing).
• Greg coins “Wine vs Coffee mode” for UI labels.

“Wine might get you a little more creative. If you want more rational execution style, maybe you want coffee mode.”
— Greg (31:00)

6. Step-by-Step: How To Run DeepSeek R1 Locally (22:56–37:00)

• Core Tools:
  • Docker: Installs containers to keep everything encapsulated.
  • Open Web UI: Clean, browser-based interface for chatting with local LLMs.
  • Ollama: Tool for downloading and managing local model files.
• Process:
  1. Install Docker > Pull/run container > Access local web interface.
  2. Download Ollama > Use to fetch specific models (e.g., deepseek-r1).
  3. Configure model connection within Open Web UI (API keys for APIs or local).
  4. Prompt the model with queries or run advanced chained prompts.

“To get started...open web UI, there is a getting started—it’s literally a couple steps to run. Make sure you have Docker installed...Ollama is going to show you all the different models.”
— Ray (31:48)

7. Running DeepSeek & Other LLMs on Mobile (37:05–44:00)

• Apollo app: Lets you download and run distilled LLMs directly on iOS (Apple Silicon leveraged for performance).
• Download only feasible for models small enough to fit device RAM/storage.
• Also supports custom API endpoints and Open Router, so you can use cloud-based models when needed.

8. Startup Ideas and Future Implications (45:28–47:48)

• On-device reasoning means potential for dozens of new startups:
  • Real-time negotiation assistants
  • Healthcare, translation, and context-aware apps on watches/phones
  • Local privacy-first transcription, analysis, or accessibility tools
• Future models (e.g., OpenAI Omni 4.0) will increase support for multimodal input—audio, video, tone—enabling more sophisticated applications (e.g., micro-expression detection for negotiation).

“Imagine being able to run this on your watch....You have really powerful devices just all on the sides of your wrist that can run these models.”
— Ray (44:03)

Notable Quotes & Memorable Moments

| Timestamp | Speaker | Quote | |-----------|---------|-------| | 00:20 | Ray | “These models have become so advanced...what that can do is even lead to superhuman capabilities.” | | 05:29 | Greg | “I wouldn’t put a tax return on deepseek.com...you do want to be a bit wary of what you’re putting on.” | | 13:54 | Greg | “What’s really...mind boggling is...this looks...pretty human level incredible. Like a senior writer would do something like this.” | | 30:49 | Greg | “I would rename that temperature as wine versus coffee mode.” | | 37:17 | Greg | “Is there any way to do this on mobile? Like, could you play with local models on the mobile device?” | | 44:03 | Ray | “Imagine being able to run this on your watch. Like that’ll just be...now you have really powerful devices just all on the sides of your wrist that can run these models.” | | 47:51 | Ray | “I think this is a really good primer for folks to get started on the power of prompting and especially with these reasoning models...” | | 51:02 | Ray | “Please don’t be fearful or...feel like you’re left behind. If you’re just finding out about this, you’re not that far behind.” |

Important Timestamps for Key Segments

• 00:00–02:30 — Introduction; Ray’s background and episode structure
• 02:32–05:57 — Deepseek.com overview, privacy, alternatives (Perplexity, Fireworks)
• 07:09–13:14 — Prompting, practical examples, chaining, model hosting
• 13:14–18:04 — Business value, model output comparison, cost/pricing of APIs
• 22:56–37:00 — Local setup walkthrough: Docker, Open Web UI, Ollama
• 37:05–44:03 — Mobile setup (Apollo, iOS), future of on-device models
• 45:28–47:48 — Startup ideas, new app possibilities, multimodal model potentials
• 47:51–51:02 — Recap, resources, encouragement for beginners

Actionable Takeaways

• If handling sensitive or business data, always check where the model is hosted—seek US/EU options or run locally.
• Experiment with model “temperature” for different types of creativity and logical rigor.
• Utilize prompt chaining and advanced instructions to harness the full reasoning power of DeepSeek R1.
• Install Docker, Open Web UI, and Ollama to run models on your own machine—full local setup is possible with basic command line use.
• Explore Apollo for mobile local inference.
• Follow Ray for hands-on walkthroughs and tutorials; leverage the community for startup ideas and support.
• Don’t be intimidated—the field is moving fast, but learning the basics and practicing with prompts is still the best way forward.

Further Resources

• Ray Fernando’s website
• Ray’s YouTube Channel
• Ollama
• Apollo (iOS AI App)
• Deepseek
• Greg Isenberg’s 30+ Startup Ideas Database

This summary was created for listeners and founders seeking actionable technical and strategic knowledge about leveraging DeepSeek R1 and emerging reasoning LLMs.