Practical AI · Practical AI LLC

Deep-dive into DeepSeek

January 31, 2025·51 min·5 clips

A Chinese startup trained a model rivaling OpenAI's best for just $5 million—how did they do it?

Daniel Whitenack and Chris Benson open this episode of Practical AI by noting the Changelog network restructuring before turning to what dominated AI news: DeepSeek R1. The model is a reasoning-class LLM from a Chinese startup that produces chain-of-thought output before giving final answers, making it directly comparable to OpenAI's O1 model. On benchmarks, DeepSeek achieves functional parity with O1 while costing an estimated $5-6 million to train the final run — a fraction of what OpenAI spent. The hosts immediately complicate this narrative: the published figure covers only the successful final base model training and excludes experimental runs, data generation costs, and reinforcement learning, meaning total spend was materially higher. Chris notes the DeepSeek team was not a bedroom startup; they had access to tens of thousands of GPUs and had been doing credible open-source AI research on Hugging Face for years before this release. The geopolitical dimension gets substantial attention. Chris, whose employer limits use of certain Chinese technologies, tested the model by asking about Tiananmen Square 1989 and received a non-answer, confirming the model is trained on a Chinese government-approved dataset. The hosts distinguish between using the DeepSeek app — which routes data through Chinese servers — versus downloading and running the open model locally, a distinction they see as critical for privacy and security. They note an irony: Meta and OpenAI also withhold full training details in their technical papers, so DeepSeek's selective disclosure is industry-standard, not uniquely opaque. The episode explores why the open-source release on Hugging Face matters regardless of national origin: it enables independent analysis, reproduction attempts, and research that proprietary models do not. The hosts credit Allen Institute for AI as a counterexample of genuine open science. They discuss whether the cost efficiency signals a broader trend — that constraints force innovation — and whether this foreshadows open-source models generally closing the gap with closed frontier labs. The conversation also touches on how this development affects enterprise AI procurement decisions and the accelerating geopolitics of AI capability.

As heard by us

A careful DeepSeek check that separates model from product and keeps the security question grounded.

The piece treats the DeepSeek discussion as a useful check on what is actually happening and where the real risk sits. Chris and Daniel do a clear job of separating the model from the hosted product, which matters for security, data handling, and the concerns people often…

Read the full review in PlayNext →

Why you'd press play

Want a clear read on DeepSeek without the panic or the hype?

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Practical AI