What non-techies should know about DeepSeek

An obscure Chinese company’s release of a GenAI model that is competitive with offerings from OpenAI, Meta and Google send the AI-hype industry into overdrive and roiled Wall Street. As usual the coverage in the popular press is more heat than light. Here’s what I think it means to the broader world. As always I’d love to hear your thoughts. [email protected]

Democratizing AI. DeepSeek breaks the monopoly the tech giants have had on models capable of understanding and responding in plain English – or just about any other language. ?The fact that the model is both open source and trained on so-so hardware means that duplicating it is within reach of lots of smart folks including those in India.

Tech -> Applications. Breaking the monopoly will push the conversation on GenAI from how cool the models are to how to make them useful. Experience in other markets shows that when a 4th or later entrant comes into a market they don’t replicate the general purpose offerings of the incumbents, but build for niches, special uses. (I wrote a dissertation on how cable changed TV from mass to niche, about a million years ago. It seems quaint now.) We can expect a flowering of models that are specifically designed for different purposes, including industry verticals, vision and sound, etc.

Language ->Reasoning. If I’m reading correctly one of the interesting twists of DeepSeek is its version R1 which stands for reinforcement learning. Reinforcement learning means that during the training and finetuning of the model experts can correct errors and the model will incorporate the corrections into how it approaches a similar issue in future. DeepSeek R1 goes one further and shows its work, laying out the logical steps it took to arrive at a conclusion. This gives sophisticated users the ability to correct its logic and improve the model. Today LLMs often mess up math problems because they follow patterns vs. format logic. I am hopeful that this is a route to improving math and logical problems.

Outlook. Where does this leave the big tech incumbents and all of the investments they are making? I’m no stock analyst and am not giving financial advice, but I’m not worried for them. DeepSeek and its progeny will reduce the money the big guys spend on training models, but that same capacity will be repurposed to running models (called inference.) Overall training is a cost, but inferencing is revenue from customers. Net I think they’ll be just fine.