Deep(Seek) Trouble

Jan Verhulst
January 27, 2025

Hello,

Welcome to a “breaking news” or even “emergency" edition of the Blacklynx Brief.
For the first time ever, we are publishing an extra edition.

Why?

Because a seismic shift has occurred in the field of artificial intelligence.

I spent the entire weekend reading and researching and decided this couldn’t wait until Friday. Our readers have to know what’s going on.

Buckle up. Here’s what you need to know.

Last week US President Donald Trump held a press conference in which he announced a 500 Billion dollar investment in Project Stargate. He was flanked by OpenAI’s Sam Altman, Oracle’s Larry Ellison and Softbanks Masayoshi Son.

During this press conference - a Chinese company released their large language model “DeepSeek R1”.

You can use this model right here.

In the world of artificial intelligence, we've gotten used to a simple truth: building advanced AI models costs an astronomical amount of money. Companies like OpenAI and Anthropic typically spend over $100 million just on the computing power needed to train their top models.

DeepSeek claims to have done this for $5 million by using a radically different approach to building AI models.

The surprising part: DeepSeek is matching and even outperforming the models that took so much money to build.

DeepSeek R1 is on par with ChatGPT’s o1 Pro. One is free , the other one costs you $200 per month.

In response to this OpenAI quickly hurried out their ‘Operator’ tool that makes you create ‘agents’ that do tasks autonomously. This is revolutionary in their own right and I’ll talk about that on Friday but this announcement is getting drowned in the uproar over DeepSeek.

Silicon Valley is in full panic mode right now.

The Traditional Approach

To understand why this is such a big deal, let's first look at how AI models are typically built. Think of it like constructing a skyscraper: traditional companies use the equivalent of bringing in every piece of heavy machinery available, regardless of whether they need it. They use thousands of specialized AI chips (GPUs) that cost $40,000 each, running at full power all the time.

These models are like universal experts - imagine having one person trying to be a doctor, lawyer, and engineer simultaneously, always keeping all that knowledge active in their brain. It's incredibly resource-intensive.

DeepSeek's Smart Revolution

DeepSeek took a step back and asked a simple question: "What if we could be smarter about this?" Their approach can be broken down into three clever innovations:

1. Precision Where It Matters

Instead of using maximum precision for every calculation (like writing every number with 32 decimal places), they reduced it to what's actually needed (8 decimal places). This simple change cut memory usage by 75% while maintaining accuracy.

It's like realizing you don't need a calculator accurate to a millionth of a penny to do your grocery shopping.

2. Reading Smarter, Not Harder

Traditional AI models read text like a first-grader: one... word... at... a... time. DeepSeek developed a "multi-token" system that reads in whole phrases at once. This makes their system twice as fast while maintaining 90% of the accuracy. When you're processing billions of words, this efficiency gain is enormous.

3. The Expert System

Here's where it gets really interesting. Instead of having one massive AI system that's always fully active, DeepSeek built what they call an "expert system." Imagine having a team of specialists who only step in when their expertise is needed, rather than having everyone in the room all the time.

While traditional models keep all 1.8 trillion of their parameters (think of these as pieces of knowledge) active constantly, DeepSeek's system has 671 billion total parameters but only uses about 37 billion at any given time. It's like having a huge company at your beckon but only calling in the departments you actually need for each project.

If you use OpenAi’s o1 the entire company is being deployed at great cost.

The Results Are Staggering

The impact of these innovations is remarkable:

Training costs dropped from $100 million to $5 million
GPU requirements decreased from 100,000 to 2,000
API costs (the cost to use the AI) are 95% cheaper
The system can run on gaming GPUs instead of specialized hardware.

Of course: this is China and China has a history of corporate espionage and privacy breaches (Tiktok anyone?). There is a geopolitical aspect to this that we’ll ignore for the time being.

It took me 10 minutes to pull down a 14 billion parameter model and run it locally on a virtual machine on my PC (and after analyzing the network traffic, it doesn’t appear to be contacting Beijing)

Perhaps most importantly, DeepSeek has made all of this open source. Anyone can check their work, use their code, and build upon their innovations.

Why Is Everyone Freaking Out?

Previously, only tech giants with massive resources could play in this space. Now, smaller companies and even well-resourced individuals could potentially train their own advanced AI models.

The business models of OpenAI and NVIDIA are about to be thoroughly disrupted.

This feels like one of those watershed moments in technology.

The implications are clear: AI is about to become more accessible, more competitive, and less expensive. The barriers to entry are dropping dramatically, and this could lead to an explosion of innovation in the AI space.

This is why I think AGI and ASI have now become inevitable.

Instead of a few tech giants holding this power in their hands, DeepSeek has given it back to the people.

I was going to say something about communism and capitalism here but let’s keep the politics out of this for the time being ;)

What's particularly impressive is that DeepSeek achieved all this with fewer than 200 people, while larger companies spend more just on the salaries of a few C-level executives.

It's a reminder that sometimes the best innovations come not from having the most resources but from fundamentally rethinking the problem.

And another thing: thank you European Union for the AI Act and regulating everything to shreds (ah shoot now i made it political).

I’ll leave you with a few memes ;-)