# AI Rundown Podcast - March 25, 2025

## Episode Summary

In this episode of AI Rundown, Sky and the team explore DeepSeek's impressive V3-0324 model update, practical lessons from AI-generated coding projects, and Google's competitive moves against OpenAI in the image generation space.

## Topics Covered

### DeepSeek V3-0324 Takes the Lead

- DeepSeek's "minor update" shows massive benchmark improvements:

- MMLU-Pro up by 5.3 points

- GPQA up by 9.3 points

- AIME scores jumping by nearly 20 points

- First time an open-weights model has led closed-source alternatives in the non-reasoning category

- Architectural innovations include mixture of experts approach, multi-head latent attention, loss-free loading strategy, and multi-token prediction

- Hardware requirements (~380GB memory) are substantial, but cloud providers offer competitive rates

- Implications for DeepSeek's upcoming R2 reasoning-enhanced model

### 100% AI-Generated Coding Project

- Technical developer built an entire NodeJS & MongoDB project using Claude Sonnet through Cursor and Windsurf

- Key lessons:

- Start with project structure before writing code

- Break down complex problems into smaller parts

- AI struggles with newer technologies not represented in training data

- Have AI create comprehensive session summaries for context retention

- Write tests even for small projects

- Commit code frequently

- AI works best with clear reference materials and documentation

- Technical expertise still essential for debugging and handling edge cases

### Google vs. OpenAI Image Generation

- Google taking direct shots at OpenAI with promotional materials highlighting OpenAI's delayed image features

- Role reversal: Google now shipping features faster while OpenAI is more cautious

- Google's Imagen 3 model showing impressive capabilities

- Google faces challenges with user experience - AI Studio separate from main search interface

- Competition between tech giants driving innovation and potentially lowering costs

## Featured Experts

- Sky (Host): Guiding the exploration of today's hottest AI topics

- Sarah: Our skeptical analyst providing measured takes on benchmark results

- Phil: Our optimistic futurist excited about AI democratization

- Storm: Our technical expert breaking down what these developments mean for programmers

## Key Takeaways

- The AI landscape is becoming much more competitive with open-source models challenging industry leaders

- AI coding assistants are powerful but work best in partnership with human developers who understand fundamentals

- Competition between major players like Google and OpenAI benefits users through accelerated innovation

## Links & Resources

- DeepSeek V3-0324 Model: [github.com/deepseek-ai/DeepSeek-V3](https://github.com/deepseek-ai/DeepSeek-V3)

- Cloud providers offering DeepSeek: Fireworks AI, DeepInfra

- Google AI Studio: [aistudio.google.com](https://aistudio.google.com)

---

*AI Rundown is a podcast where we don't do deep dives - we just look at whatever topic's hot today. Join us for our next episode!*