Foundation Model Researcher Tian Keyu Never Lacked GPUs

"ByteDance Chemistry Department Entrepreneur"

You all know what happened by now. Yesterday Tian Keyu stirred up some drama, and a whole crowd came out to debunk Tian God — not just with timelines and code, but with their own GitHub profiles thrown in for good measure. Academics really are rigorous about this stuff 👇

https://keyu-tian-sabotage.github.io

Put yourself in the shoes of the interns on the other side of Tian God. It was pure despair.

Every day they'd see their model training parameters getting messed with, but they couldn't figure out where or how. They'd try to troubleshoot but couldn't trace the root cause, because Tian God was right there in the troubleshooting meetings with them — he knew their every move.

That's literally a Sophon, isn't it? The Sophon knows everything humans are doing, but humans don't know what the Trisolarans are thinking or doing, or where the Sophon even is.

Years from now, tech historians will surely record: In 2024, Trisolaran Tian God launched a Sophon attack on ByteDance interns.

Afterward, everyone wondered how he could sabotage things while his own model progress stayed smooth — then they discovered the model he'd deployed internally for team testing was just an open-source Text-to-Image model.

To avoid detection, Tian God had set it up so every user prompt automatically got a "City" prefix tacked on. That's when everyone finally understood why all their generated images had city skyline silhouettes in the background.

This maps perfectly to The Three-Body Problem, where humans discover the Sophon and it projects "YOU ARE BUGS" across everyone's retinas! So Tian God is absolutely a Sophon reincarnate.

Of course, all this is what the accusers are saying — it's rival slander against Tian God.

The case ByteDance filed against Tian Keyu is now in its second trial. The first-instance verdict ordered him to pay 500,000 RMB, which compared to what he's raised is like a bottle of iced black tea — just pay it and be done.

But ByteDance appealed, because what they actually want is 8 million RMB from Tian God.

How do I feel about this? Let's say everything Tian God said is true — no two months of sustained attacks on training code, just "I had GPT write me a program that stopped or froze things," and the timeframe was shorter, "about ten days, roughly five times total."

Then ByteDance's losses weren't that serious at all, so what's ByteDance's angle suing for 8 million?

It's not like Tian God manually attacked five times within ten days instead of the two months of sustained attacks the accusers claim, right? Tian God swore he's telling the truth.

So I can only conclude that ByteDance must be out of money for GPUs and is trying to squeeze some cash out of Tian God to buy more 😭

Damn, ByteDance is so evil. Once my Tian God's world model can predict the world, your company is getting destroyed.

Also, this is a civil case, no criminal liability, so Tian God isn't doing prison time.

You cannot call him "Prison Tian," and you definitely cannot slap "Prison Tian" on beverage bottles. That would be wrong.

Xianyu posted on Jike asking everyone to discuss rationally: Is Tian Keyu being able to raise money a dividend of the era, or a tragedy of the era?

He thinks it's a dividend of the era. Because back in the Mobile Internet era, everyone called Justin Sun an idiot, but Justin Sun got the biggest outcome of that whole wave of entrepreneurs.

We believe that with Tian Keyu's depth of scheming, getting wrecked by ByteDance and still exploding to a $200 million valuation off $30 million from 5Y Capital, IDG, and Zhongding — that kind of ability to win while playing from behind — his future achievements won't fall short of Sun's.

Sun was just saying the other day that whoever lets him join OpenAI, he'd pay $50 million for the privilege.

Honestly just give that $50 million straight to Tian God — that's basically joining a world-class AI lab right there, plus you get to treat your younger self to dinner again.

Tian God debating the details is a narrow path. What's the point of arguing "I didn't poison" — what does denying it at this moment even accomplish?

Better to just get a personal trainer with Sun and start building that lean muscle, chugging water all day. In this solar system right now, who's more meta than these two?

Of course, technically Tian God didn't poison anything, because poisoning is a chemistry concept. As I understand it, since everyone's physically healthy, Tian God didn't poison anyone, so stop slandering him.

Oh right, and it's a remarkable coincidence — IDG was Justin Sun's investor back in the day. IDG's eye for talent is incredible, they've basically invested in another Justin Sun. I'd say the top three are decided, what do Sequoia and Hillhouse even have to compare?

But my friend Brother Cheng offered the opposite take:

"Tian Keyu broke the law; Justin Sun never did. Justin Sun is a thoroughgoing idealist, the opposite of Tian Keyu.

And Sun's values are solid, he hates evil like a personal enemy. He repeatedly told his father not to be corrupt, and here's another example: he sent all his employees who embezzled from his company to prison."

Fair point, I guess.

What I personally admire most about Tian God is that he calls himself a foundation model entrepreneur.

That's right — not large language model, not video model, and definitely not world model, but foundation model!

Because hyping yourself as world model or embodied AI is a narrow path now — too many grifters, too much noise, you can't tell who's legit.

"Foundation model" is the perfect term. You can scale up to a $50 billion Kimi, or scale down to a nice little $200 million exit.

What's a $200 million valuation? Let me give you an example: ByteDance's former Trae lead Shi Yang, teaming up with ByteDance's data security lead to start something — that's also a $200 million valuation. Two executives combined equals one Tian God.

What does this tell us? It tells us that ByteDance's strongest former intern is equivalent to 4-2. Melvin Chen was 28 before he made 4-1, and he'd been hyping that level forever. Tian God is a post-00s, so who's higher between Tian and Mel — no question.

Mel God is dead, long live Tian God!

Though a senior investor did point out to me that Tian God perfectly hits every VC old-timer's pleasure point, because VCs love investing in rebels — they just underestimated how much public fury Tian God would spark among researchers.

To which I can only say: there's rebellion and then there's rebellion. Che Guevara and Pol Pot are different categories. Of course, Tian God is a Trisolaran, beyond the five elements, outside our judgment.

As usual, we'll directly distill Tian God's interview quotes, with random selections of debunking from fellow ByteDancers and our commentary:

01

"I never lacked GPUs myself; the sabotage was mainly out of a sense of justice"

1. I started interning at ByteDance in 2021, and in 2024 a new intern joined. After that I felt the whole team was on edge. He started with some soft resource grabs — like saving seats, taking resources that belonged to others — then started directly stopping other people's programs.

My Tian God says the other intern stopped people's cards first, so don't blame him for not being polite. This is individual heroism.

2. Actually he never stopped mine (my cards), he stopped people with softer, more introverted personalities. We gave our leader feedback on this,

but it wasn't resolved. So I used my method — this method was of course wrong — I entered his program and froze or stopped it. That way his excess resources would get reclaimed, and others could use them normally.

Ah, Tian God is so full of justice, he wasn't even personally affected. How could this leader not do anything, making my Tian God face a lawsuit 😭

3. Actually I personally never lacked GPUs.

Even if ByteDance lacked GPUs, they wouldn't let our Tian God lack them. OpenAI lacks GPUs but Tian God never does!

4. I must say this: "saying I framed others" — this absolutely didn't happen, and it was what I found most incomprehensible and what made me most emotional at the time. I'm not someone who does things but won't own up to them.

I did call the police. Actually it was because the company found me and asked if I knew who made these posts, I said I didn't know, and wanted to see if police could find out who posted them. After confirming with the company, I went to report it.

Salute to the most honest, most courageous Tian God who admits his mistakes. But real talk, if you're so brave and accountable, why not just admit it when the company came asking? Why call the police to find out who posted?

5. (I swear), all this about viruses, about attacks and defenses — it's all baseless rumor.

Tian God swore an oath, how could it be malicious?

6. This wasn't poisoning, it didn't involve attacks and defenses. I had GPT write me a program that stopped or froze programs.

Also, I manually, point-by-point stopped the other person's program, absolutely no two months of sustained attacks... about ten days, roughly five times total.

Listening to Tian God, I get the feeling the other intern was just so noob they got wrecked by a GPT program, my Tian God just clicked a few times and they got blown up.

Now look at the other side's description: "According to post-incident review, Tian's sabotage operated on multiple levels. First, he directly modified PyTorch source code in the training cluster, intervening in random seeds, optimizers, and data loaders.

These components directly affect model training data ordering, parameter updates, and numerical computation, so such tampering could cause unpredictable deviations in training results.

Second, he randomly terminated some processes in multi-machine training jobs, causing normally running large-scale experiments to suddenly hang. More insidiously, the attack code used checkpoints as a vector to establish login backdoors, enabling subsequent attacks to trigger automatically and randomly terminate training processes.

The common feature of these attack methods was disguising deliberate sabotage as random software failures, numerical anomalies, or infrastructure issues."

7. I think ByteDance is great both in intern resource support and communication atmosphere.

Other side's view: "Tian consistently participated in project debug communication groups, even attending daily troubleshooting meetings for the training cluster. Team members reported newly discovered anomalies every day, discussed possible causes, shared troubleshooting approaches, and planned follow-up experiments and fixes based on this.

Tian could directly access this information, and based on colleagues' investigation directions, continuously and targetedly modified his attack code."

The communication atmosphere really was excellent.

8. I'd call my behavior at the time "naive justice," but it was indeed very immature. Like if I encountered something on the street, I might handle it in an impulsive way.

Other side's view: "On the morning of July 30, 2024, while investigating training anomalies, team members unexpectedly discovered obviously abnormal code in the DataLoader section of PyTorch source code in a machine's Docker environment. This didn't look like normal development changes, but showed clear signs of deliberate malicious tampering.

After investigation, the team reported this discovery that same day, and further checked login records for that machine. Besides the machine's owner, only Tian had logged into this machine.

On July 31, 2024. Another team member discovered large amounts of code suspected to be for sabotaging the training system in Tian's personal directory, with shocking contents."

9. I'm not the villain from the rumors.

Correct, Tian God is a good person. He did this to test ByteDance's cybersecurity capabilities, didn't expect them to be so fragile. ByteDance needs to reflect on this.

02

The self-cultivation of a foundation model entrepreneur

10. My definition of world model is predicting the future. Too many definitions today, too little implementation. We set this direction in 2023, mainly focused on implementation.

Tian God realized he was going to build world models during his 2023 ByteDance internship, a full two years ahead of Fei-Fei Li, absolutely not riding coattails. Once Tian God can predict the future, you haters are all finished!

11. I don't care whether our company gets labeled (world model) or not, neither actively seeking it nor avoiding it. Anyway we're just building a multimodal foundation model.

Right, everyone's definition of world model is outdated now, please use our full name: multimodal foundation model.

12. We're the most pioneering of all today's (world model) approaches.

On world models, we've observed the same pre-training Scaling Law as language models. This near-hundred-billion-parameter unsaturated Scaling Law is the only one.

Amazing, Tian God has achieved enlightenment at the universe factory, the Scaling Law for world models is coming.

13. The ultimate form of world model is Predict everything.

For example, if I take the state of that clump of matter at the Big Bang 10 billion years ago and feed it to World model, it can simulate forward 10 billion years.

Another intuitive example, like Doctor Strange in Marvel: when he fought Thanos, he saw many different futures, and only one succeeded.

I am enlightened — Tian God must be Doctor Strange himself, he found the only path to funding among thousands of futures: standing up for those who lacked GPUs!

14. What attracts me to this field is that without human teaching, it can spontaneously learn things.

For example, after seeing massive amounts of images, it can vaguely sense the difference between cats and dogs. Language models can self-learn too, but language is human-created first. I prefer distilling nature, not distilling humans.

Tian God has seen the endgame of language models 😭 How far can your eyes see, Tian God

15. When we want to tell fortunes, I toss three coins, and in that instant they're in the air, World model should be able to predict it. This used to be hard, but at our current stage we've already achieved it.

What more is there to say, the haters slandering Tian God today — he probably already predicted this 👍

16. If I'm lying, then my research will suffer setbacks for life.

Wishing Tian God the next Best Paper award.

Finally we'll leave you with a thought exercise: Can Tian Keyu count as a ByteDance-affiliated entrepreneur?

If we're talking about Tian God's entanglement with ByteDance, his connection to the ByteDance network, this bond is definitely deeper than most ByteDance employees.

After all, most employees have never been sued by ByteDance, while Tian God was formally sued by ByteDance — probably the first intern in history to be sued by them.

So Tian God's bond with ByteDance is extremely deep, making him unquestionably a ByteDance-affiliated entrepreneur.

Though I think he should count as a ByteDance Chemistry Department entrepreneur, suited for AI4S.

(This article's cover image was generated by ChatGPT, text is purely human-written)

⬇️

Subscribe to our Substack funeralai.substack.com