Don't Worry About the Vase Podcast DeepSeek v3.1 Is Not Having a Moment
Aug 22, 2025
Dive into the hurdles faced by a Chinese AI company with their new DeepSeek v3.1 model, grappling with tech woes and geopolitical issues. Discover the model’s innovations and how it stacks up against competitors like Claude Opus 4 and GPT-5. The analysis covers impressive advancements in coding and reasoning, pricing updates, and user reactions. Explore the future potential of DeepSeek and its performance in AI benchmarks, revealing the evolving landscape of artificial intelligence!
AI Snips
Chapters
Transcript
Episode notes
Geopolitics Slowed Model Progress
- DeepSeek delayed model progress after being pushed to use Huawei Ascend chips instead of NVIDIA, which caused training failures and slowdowns.
- Company used NVIDIA for training eventually and Huawei for inference, costing them time and competitive ground.
V3.1 Is A Unified, Marketed Upgrade
- DeepSeek released V3.1 as a unified model claiming agent skills, 128k context, and hybrid inference improvements.
- Marketing emphasized "Think" and "Non-Think" modes and API updates rather than a clear leap in capability.
Benchmarks Look Strong But Conflicting
- Public benchmark visuals show V3.1 outscoring prior DeepSeek models and claiming strong SWE and multilingual gains.
- Yet other observers report only minor improvements and similar ceilings to earlier versions.