#227 - Jeremie is back! DeepSeek 3.2, TPUs, Nested Learning - Last Week in AI Recap
Podcast: Last Week in AI
Published: 2025-12-09
Duration: 1 hr 35 min
Summary
Jeremie returns to the podcast after a hiatus, diving into the latest AI developments, notably the release of DeepSeek 3.2, which boasts significant advancements in performance and affordability compared to other leading models. The episode also touches on various AI applications and hardware developments.
What Happened
In this episode, Andrey Karenkov and Jeremie celebrate Jeremie’s return after a three-month break. Jeremie expresses excitement about the recent advancements in AI, particularly the release of DeepSeek 3.2 and its competitive positioning against other models like GPT-5. They discuss how this new version of DeepSeek is not just cheaper, with a 50% reduction in cost, but also demonstrates improved performance on benchmarks, showcasing its capabilities in reasoning and efficiency through techniques like sparse attention.
The hosts delve into the implications of these advancements, noting that DeepSeek 3.2 performs neck and neck with other leading models in various tasks. They highlight the transparency of DeepSeek's developers, who provide insights into the technical enhancements that contribute to its performance, such as refined reinforcement learning objectives. Jeremie also mentions that DeepSeek R2 is on the horizon, which aims to focus on hardcore reasoning capabilities, indicating that the field is rapidly evolving with significant model releases and innovations.
Key Insights
- DeepSeek 3.2 offers a 50% cost reduction compared to competitor models.
- The model shows competitive performance, even outperforming GPT-5 in certain benchmarks.
- Technical advancements include sparse attention techniques for faster processing.
- Future developments like DeepSeek R2 are focused on enhancing reasoning capabilities.
Key Questions Answered
What are the key features of DeepSeek 3.2?
DeepSeek 3.2 is highlighted for its significant cost reduction, being 50% cheaper than its competitors like Anthropic. It also boasts enhanced performance on benchmarks, indicating that it can compete closely with models such as GPT-5. The model employs advanced techniques like sparse attention, which not only improves speed but also reduces operational costs.
How does DeepSeek 3.2 compare to GPT-5?
Jeremie notes that DeepSeek 3.2 performs neck and neck with GPT-5 in various tasks, even outpacing it in some instances. This competitive performance marks a significant milestone for DeepSeek, showcasing its ability to meet the growing demands for reasoning and efficiency in AI models.
What technical advancements contribute to DeepSeek's performance?
The episode reveals that DeepSeek's developers are transparent about their advancements, particularly in refining reinforcement learning objectives and implementing sparse attention techniques. These improvements are crucial for achieving faster processing speeds and better overall performance, which are essential in the rapidly evolving AI landscape.
What future developments are expected from DeepSeek?
Looking ahead, the hosts discuss the anticipated release of DeepSeek R2, which aims to enhance reasoning capabilities significantly. This indicates a strategic focus on pushing the boundaries of AI reasoning, reflecting the broader trend within the industry to prioritize advanced reasoning in model development.
What other AI news was discussed in this episode?
Apart from DeepSeek 3.2, the hosts touch on a variety of AI news, including new applications in business, hardware developments, and updates from major players like OpenAI and Anthropic. The episode aims to cover a wide range of topics, reflecting the dynamic nature of AI advancements happening in real-time.