Apropos DeepSeek by Dan Bergh Johnsson

DeepSeek is an open LLM model promising innovation in both efficiency and transparency – but how much do we really know about what’s happening under the hood? And what does it tell us about where we are in AI development?

In this session, we examine DeepSeek from two perspectives:

Security and risk: What challenges arise when an AI model is built and distributed openly? What does this mean for reliability, integrity, and control? What risks come with DeepSeek’s Chinese origins, and how does that affect trust in the model? Technical architecture and efficiency: Has DeepSeek truly broken new ground, or is it just another example of solving inefficiency by throwing more hardware at the problem? We’ll explore the technical details of Mixture of Experts (MoE), parallelization strategies, and why LLMs still resemble steam engines – powerful but wasteful. We’ll also discuss a fundamental limitation of today’s transformer-based AI: that a model’s performance is closely tied to data availability, which in turn dictates how large and effective it can actually become. Finally, we’ll look at why the AI industry – DeepSeek included – remains so reluctant to be transparent about training methods and optimization.

Join us for a deep dive into both the risks and possibilities of open-source LLMs

Dan Bergh Johnsson
Omegapoint

Recorded at AI-fokus 2025 in Stockholm 10 June 2025
https://www.jfokus.se/ai-fokus/