Tuesday, 28 January, 2025 / Edition 44

This week, we’re going over a hot topic and recent development in the tech world that has a lot of people drumming up intense emotions. I get it: uncertainty can be triggering. Where are we going as a society with our AI journey? We won’t know until we get there (or at least get closer). But we can speculate and make educated guesses, and in comparing different perspectives and engaging in discussions, we might learn something new. So, with that in mind, let’s dig in to see how you might be able to put your mind to work on the matter.

Sarah Compton

Editor, Enspired

DeepSeek is on the Scene

Soumybrata Roy/Shutterstock.com

DeepSeek. It sounds like something that's built for and/or targeted at the geoscience world, but alas, it is not.

What is it? DeepSeek is a Chinese company with a large language model (DeepSeek R1).

Driving the news: The buzz around DeepSeek is because it was built using fewer resources (data, power, money) and less sophisticated technology, but it still performs nearly as well as the big boys. Let’s take a closer look.

DeepSeek has taken the world by storm with two big claims:

• The money required to train the model is under $6 million.

  • When you compare that to the tens of billions invested by, and in, U.S. companies to build their language models, it seems like someone might have some egg on their face.

• The model was built after the United States banned shipments of NVIDIA’s A100 and H100 AI chips to China.

  • That ban was intended to prevent this exact event—China building their own just-as-good or better than the United States' AI model—from happening.

If these two claims prove true, DeepSeek-R1 could represent a “seismic shift in the world of technology,” according to former Microsoft employee David Plummer, who likened the development to what would happen if someone were able to “throw together a Ferrari in your garage with spare Chevy parts.”

Digging a little deeper, reveals a few complications, however:

• DeepSeek is owned and funded by hedge fund High-Flyer, which founder Liang Wenfeng built for stock-trading and which was driven by a deep learning model run on graphic processing units (GPUs).

• The background of High-Flyer indicates the “kid” who built DeepSeek isn’t a kid who just said, “I think I’ll build an AI model today.” (Throwback to the Legally Blonde line: “I think I’ll go to law school today”). He had already been building deep learning and AI models for nearly a decade before this breakthrough.

• Supposedly, in 2021, Liang began stockpiling NVIDIA GPUs for an AI project and had about 10,000 NVIDIA A100 GPUS on hand before the sanctions went into place.

Another big consideration here is that DeepSeek-R1 is open source, meaning its freely available to anyone who wants to use it, and there’s no indication that’s going to change any time soon.

A message from AAPG Academy and ThinkOnward 

Register now to join AAPG Academy and ThinkOnward on 12 February at 9am CST for a free webinar covering a unique approach to geoscience data management using AI.

Expert speakers will share:

• A practical look at how AI is transforming geoscience data management by combining various types of subsurface information into searchable, location-tagged databases that you can easily navigate using specialized LLMs trained to understand geological terminology
• An overview of the technical aspects of building a reliable AI system for geoscience, including the use of OSDU-compliant data models and specific measures to prevent AI hallucinations while incorporating expert domain knowledge
• Current insights and early results from developing next-generation G&G workspaces, with an opportunity to contribute to the ongoing development through interactive discussion and feedback

How Much Does It Matter?

Erlin Diah/Shutterstock.com

The company's logo is a whale, yes, but how much of a giant is DeepSeek in the AI space, really?

Some techies and others are panicking, concerned DeepSeek will de-throne the United States as the leading force in AI development.

Plummer’s summary of how the model works eased my mind:

• It uses larger models for scaffolding, and he believes it is a distilled language model.

• Distilled language models take a larger model and use that to train the smaller one, similar to a master training an apprentice.

• DeepSeek takes that to an extreme by compressing bigger systems into something smaller and lightweight.

• The result of that is it doesn’t need massive data centers to operate, and the power requirement is greatly reduced.

How big a deal is this? Plummer’s description of a master training an apprentice caught my attention.

Can/should society really give DeepSeek the grandiosity and attention it has if DeepSeek is built on the shoulders of larger AI programs that came before it (and which did not receive the same kind of recognition)? It’s not an entirely new innovation—just a cheaper, more compact, and more built-for-purpose one.

That being said, one thing is clear: this model represents a shift in AI the same way that the desktop represented a change to big mainframe computers (and which one are you currently using…hmmmm?).

Success and impact are TBD: Seismic shifts in tech require seismic proof, and it’s still a bit too soon to know exactly how DeepSeek got here and if the claims of “fewer inputs, nearly equal outputs” is true.

Weaknesses of even the big guns include specialty knowledge, “hallucinations,” and fake data confidently delivered, and those weaknesses seem to be more pronounced in this smaller model.

ISO domain-specific SMEs: Still, it seems the blueprint is being laid out (or begun) for geoscience to build more specialized—but still incredibly powerful—models. DeepSeek-R1 didn’t pull in only computer science folks: it incorporated domain-specific SMEs to help improve their model’s accuracy, and we geos could easily serve as some of those.

A Message From MicroSeismic

In celebration of the International Day of Women and Girls in Science on February 11, AAPG and MicroSeismic are showcasing interviews from women leaders. This week's featured leader is Anna Krylova from MicroSeismic. Here is a sneak peek from her interview:

“What still amuses me every day about geoscience is how much we know about Earth and the subsurface—and at the same time, how little. We still discover new oilfields; technology advances and we get something new even from old fields where wells were drilled for ages.” 

👍 If you enjoyed this edition of Enspired, consider supporting AAPG's brand of newsletters by forwarding to a friend or colleague and signing up for our other newsletters here.

➡️ Was this email forwarded to you? Sign up for Enspired here.

✉️ To get in touch with Sarah or suggest a topic for Enspired, send an email to editorial@aapg.org.