DeepSeek rocks the AI boat, but few in publishing shed tears

A roller coaster week at the cutting edge of tech has caused a surreal blend of consternation and delight

by Rob Corbidge

Published: 17:27, 30 January 2025

Last updated: 18:08, 30 January 2025

What in the name of God's bananas has happened this week? 

A guy who is primarily interested in using Machine Learning to predict market behaviours has caused a market shift that no one, using flesh or silicon, could predict. 

OpenAI, a company founded as a non-profit which collected a great deal of money for exclusive use as a non-profit, before becoming a for-profit, has been leapfrogged by an open source project created and given away by that most piratical of capitalist endeavours, a hedge fund. 

Open AI, a company which has spent years harvesting other people's content without much regard to its creation or ownership is now crying foul that someone has stolen their IP.

Did we miss anything else in the DeepSeek saga? 

Possibly. It was known by the AI cognoscenti that DeepSeek's Chinese researchers were progressing rapidly a year ago. Curiously, news of their progress seems to have been delayed akin to a medieval monarch receiving messengers from a distant battle. Informed folk were enthusiastically flagging the processing improvements seen in DeepSeek's model V3 at the back end of last year. 

Why then, four weeks later, are we suddenly plunged into a multi-level frenzy about the capabilities of newest model only after technosage/investor Marc Andreesen posted that "DeepSeek-R1 is AI’s Sputnik moment"? 

It's all a little perplexing and it seems doubly so for Wall Street, which has been flinging money at technology most investors don't understand in the hope more money is spat out from the other end of this wondrous digital confection.

It seems that rather than investing in artificial intelligence, a little more in the way of market intelligence would have been useful.

The simple answer is that a hype train is rumbling down the tracks, and woe to those in its way. It's a shame, because it's harder to have a calm discussion about AI systems with all the noise and juddering that is the signature of a hype train's transit. 

More noise, less understanding

In the UK where I am, and likely in much of the developed world, it doesn't help that we have effortlessly transitioned from a generation of politicians who knew nothing about new technology, and let Big Tech get away with numerous horrors, to a generation of politicians who feel they must pontificate broadly and frequently on tech, despite largely matching the same vacuum of understanding as their predecessors.

The truth behind DeepSeek's success is still cloudy. 

• Did they actually use a stockpile of inferior Nvidia H800 chips and come up with a superior solution requiring less compute power? 
• Did founder Liang Wenfeng use a team of young developers unfettered by conventional thinking to outfox less adventurous rivals? 
• Have DeepSeek made a mockery of the previous US administration's "small yard, high fence" attempts to protect critical technological know-how?

Perhaps for our own publishing industry, a moment of schadenfreude can be taken from OpenAI's statement that they believed DeepSeek had "distilled" data taken from ChatGPT.

Refreshingly for those of us who have to force ourselves to understand technical terminology, "distilling" in this case resembles distilling in the more usual sense: certain data is separated out from a larger mass of data, and then used in its more concentrated form. In this case, the "larger mass of data" belongs to OpenAI and they say DeepSeek used ChatGPT's outputs to train. 

Forgive us all for not crying, but as one observer put it, "I'm so sorry I can't stop laughing. OpenAI, the company built on stealing literally the entire internet, is crying because DeepSeek may have trained on the outputs from ChatGPT."

As my colleague Rich put it a couple of weeks ago, "the OpenAI pot [seems] unable to call the rival kettle black because to do so might prejudice its defence in separate legal matters." 

Will they sue, we all ask.

As it turns out, distillation is a known method of producing lower cost LLMs, which use data from larger and hence pricier models, and reduce the large model to a smaller one, which then operates on this tighter set of data, while using the larger model's reasoning capabilities. 

If they did do that without an OpenAI licence, then DeepSeek have simply learned from the masters of data harvesting. It's a competitive world, Mr Altman, and price matters.