I thought I’d talk to you about DeepSeek for a while. I mentioned them in the last SundAI. I have been eyeballing them for some time, and they’ve really done it. This scrappy little underdog has managed to send shivers down OpenAI’s spine with six million bucks and a prayer. And they done it.
Oh yes, six million.
That’s just pocket change for Sam, Mark, Elon, Jeff, Tim et. al. They spend that much on kombucha and Teslas in a week.
But DeepSeek?
Man, they flipped that spare change into something that makes OpenAI’s billion-dollar GPT-4 look big and fat and sluggish.
Truly impressive!
Chapeau!
Yeah.
But let’s not pop the champagne yet.
Hear me out.
More schtuff after the commercial brake:
- Comment, or share the article; that will really help spread the word 🙌
- Connect with me on Linkedin 🙏
- Subscribe to TechTonic Shifts to get your daily dose of tech 📰
An David in a Silicon Valley of Goliaths
OpenAI burns billions building models that devour data like a black hole:
The company is projected to have spent up to $7 billion on AI training (and inference, that is operational cost), and $1.5 billion for staffing. This brings their total projected expenses to approximately $8.5 billion for the year.
It is estimated they generate a revenue between $3.5 billion and $4.5 billion. That means that OpenAI has a yearly loss of around 4 to $5 billion. These figures are just staggering. And until now, we thought that you could only be at the top of the game if you were able to cough up these big numbers to develop, train, and maintain the AI.
Well, for Sam, the costs are not so important. They are valued at 157 Billion and if they ever go public, he’ll walk away as one of the richest tech bros on the planet.
And along comes DeepSeek.
Ladidadida.
DeepSeek comes from Hangzhou, and that is in China. And they have been playing with AI for some time now, and they just showed the world that you don’t need Silicon Valley’s excess to make AI magic.
I am talking $6 million versus OpenAI’s $billions.
One weekend of operational costs for the big boys equals DeepSeek’s entire development budget.
This is kinda like comparing a community theater to Broadway. And in this case the local troupe might actually be better than the Broadway theater.
Oh yes. It’s that wild.
Who doesn’t love a good startup-story!
The garage band that makes Silicon Valley sweat
This isn’t a polished success story. Nope-sure-eee! This is about a team that probably worked out of a broom closet, where they were slamming energy drinks and eating Snickers bars. Well, at least the Chinese version of Snickers anyway.
And they didn’t have hundreds of engineers. They didn’t have data centers the size of the pyramids of Giza. What they had was smarts and stamina, a few clever training tricks, and a few “efficiency” tricks, that will-just-raise-a-few-eyebrows.
But hey, a win is a win.
And if you do the numbers it gets downright grotesque. For every dollar that DeepSeek spent, the boys at OpenAI spent $167.
Let that sink in.
And the brilliance of it all is that DeepSeek’s AI smashes through benchmarks. DeepSeek-V3 has created what is called a Mixture of Experts architecture. That means the model acts like a team with specialist models. It is not one huge model, like some of the other players have. And it’s making OpenAI look like they wearing cement shoes.
And instead of making everyone work at the same time on a problem, the experts are called in. This saves a lot of time and energy, and they are still getting fabulous results. This means that from a parameters point of view, only a small subset for any task is used instead of the billions in total.
But efficiency doesn’t come without compromise. When something is made cheaper or faster, you can bet your ass that corners are cut. Maybe it’s less safe, or less ethical, or it is missing something important that gets overlooked in the rush to save time and money.
Something will bite you in the butt cheeks, you can be sure of that..
But I’ll get there…
But first..
The Red strings.
A.k.a. DeepSeek’s ties to the Chinese government
I bet you knew this was coming.
That’s what you get when your readers are such smart asses: you cannot make mistakes, and sure as hell you need to let them think for themselves.
But I’m digressing…
The rule of thumb is that you can’t talk about a Chinese AI company without talking about the puppet master in the background. The Chinese Communist Party. DeepSeek might be the underdog in Silicon Valley, but back home, they are playing a very different game.
Every algorithm, every model, every move is reported back to the government. Oh yes, your shiny new AI might just be feeding the surveillance state. Innovation, meet Orwell.
And there’s another catch with DeepSeek’s AI. Everything it generates, every line of software, and every little snippet of what-have-you, is just theirs.
Not just metaphorically.
Legally.
Just check the T’s and C’s
Because hidden within DeepSeek’s terms and conditions there are clauses that can leave users exposed to serious risks. From taking full responsibility for all Inputs and Outputs to granting the company vague rights to data usage.
Take Section 4.1. Users are solely liable for any data they submit or generate with DeepSeek. That means if you accidentally upload copyrighted or sensitive material, or if your Output causes legal issues, it’s your head on the chopping block. On the input side, I can relate to that one myself, but not on the output.
And it doesn’t stop there.
Section 4.2 lets DeepSeek use your data “to maintain or improve its services”. So anything you upload is game. And you cannot opt-out of this one. And if you couple this with Section 5.1, that states that all intellectual property rights are for DeepSeek, you can bet your peachy cheeks that no matter what you do with the tool, everything that goes on, is theirs. And if you happen to misstep (intentionally or not), you could face legal action.
That means if you use their AI to build your dream software, and that software turns into the next big thing. Well, congratulations, but don’t pop the cork yet. Because DeepSeek owns the software which is the foundation, and by extension, they can stake a claim to the whole damn castle you’ve built on it.
You’re left with two options: hand over a royalty check or, in the worst-case scenario, say goodbye to the company you built. It’s innovation, but at a cost. And in this case, the price might just be your company (or autonomy)
Basically these terms, they prioritize DeepSeek’s control AND offload all accountability to its users.
Censorship embedded in their code
Ask DeepSeek V3 about the Chinese Communist Party or President Xi Jinping, or the events that took place at Tiananmen Square, and you will get a sanitized, government-approved response, or even worse, a complete shutdown of the conversation.
The model spins happy propaganda of harmony and stability while underneath they bury any mention of the darker episodes of China’s past. The AI must align with “socialist values”.
And if you search for other topics, like Russia’s invasion of Ukraine, or North Korea’s human rights violations, or criticism of Vladimir Putin, the model suddenly feels free to voice anything. But when it comes to China, the guardrails hit hard.
But the thing is that generative AI is unpredictable. And that means that even the most authoritarian governments are having a hard time to maintain total control. But DeepSeek’s V3 just shows how determined the CCP is. But, users being users, they have already discovered numerous ways to bypass the model’s censorship. For example, inserting periods between letters allows the model to discuss the Tiananmen Square protests in detail. Something it would normally suppress.
So the question is… how do you suppress freedom of speech in a system that is designed to generate diverse outputs? The party’s answer, so far, is to limit training data to government-approved sources.
But even here, you can see that cracks are forming. There is some evidence that suggests that DeepSeek V3 was pre-trained on ChatGPT-generated data.
There was a funny rant going on the internet about this, claiming that DeepSeek thought it was ChatGPT. Have a look:
If that ain’t proof of the pudding, I don’t know what is.
But censorship in AI is not “invented” by the Chinese. The first evidence of it is found at ChatGPT. The AI had wrongly accused a Harvard professor of sexual abuse. So they have now intervened and whenever you try to find the hallucination, you get kicked out of the system:
For more, read: I’ve seen the dark side of AI, and you need to know about it | LinkedIn
Systemic government control
What makes DeepSeek different from Western AI companies is not limited to its censorship alone. It is the systematic way that censorship is embedded into the development process.
Before any AI model can be released in China, it must pass state verification. State verification means it has to speak a language that aligns with “socialist values”. And this level of control goes beyond the bias, that we in the West worry about. This is manipulation. Transforming technology into a tool of the state.
Western AI models, for all their flaws and biases, don’t operate under such explicit directives. The difference is deeply concerning for me. DeepSeek V3 is the embodiment of state control dressed up to look like technological progress.
Signing off, don’t trust the shine. There’s always rust beneath.
Marco
Well, that’s a wrap for today. Tomorrow, I’ll have a fresh episode of TechTonic Shifts for you. If you enjoy my writing and want to support my work, feel free to buy me a coffee ♨️
Think a friend would enjoy this too? Share the newsletter and let them join the conversation. Google appreciates your likes by making my articles available to more readers.
Leave a Reply