DeepSeek

DeepSeek is the point at which ToTheMoon began treating Chinese models as full participants in the race rather than followers. The hosts do not ask whether DeepSeek is better than ChatGPT; they ask what changes in the economics of AI once a strong model is open and cheap — where its output can be trusted and who is now able to run a system of their own.

February 2025 brings a dense run of episodes. In episode 043 the collapse in NVIDIA's market value is called too linear a reaction: when compute gets cheaper, people use more of it, not less; the model itself is praised for its visible reasoning, yet on the tested cases its final answer still trails o1. In episode 044 first place in the App Store sits beside two awkward questions — whether the model was trained on OpenAI outputs and where the data goes — and episode 046 splits the proposed US ban in two: banning the web service and banning the open model itself, under an MIT licence.

In spring 2025 DeepSeek becomes the benchmark for price: consumer apps are moving to it because it is cheaper (episode 048), and “DeepSeek showed up and made it all free” is one of the arguments in episode 049 against a twenty-thousand-dollar OpenAI subscription. Episode 053 adds politics: travel restrictions on key employees show that a state can release a model while keeping the team under control. By year's end, episode 088 places DeepSeek's open models beside GPT-5 and Gemini at far lower prices, and in episode 129 DeepSeek V4 Pro takes part in a test of models' political characters as one of six systems whose behaviour is set by design.