GPU‑Thieving Intern Wins NeurIPS Best Paper
Key Points
- An intern at ByteDance (TikTok’s parent) stole a large number of GPUs by sabotaging internal AI training pipelines, leading to a $1 million lawsuit and his termination in August 2024.
- The intern, named Kouan, used the stolen compute time to develop a paper on “Visual Autoregressive Modeling: Scalable Image Generation via Next‑Scale Prediction,” pushing the field beyond token‑ or pixel‑level prediction toward reasoning over larger image concepts.
- Despite the theft, Kouan submitted the paper to NeurIPS (the premier AI conference) and, after a blind review that judged the work solely on merit, the conference awarded it Best Paper in December 2024.
- The award sparked controversy, as the conference organizers knowingly recognized work done with stolen resources, while ByteDance remains outraged and continues legal action against Kouan.
Sections
Full Transcript
# GPU‑Thieving Intern Wins NeurIPS Best Paper **Source:** [https://www.youtube.com/watch?v=6A3NOedlPWI](https://www.youtube.com/watch?v=6A3NOedlPWI) **Duration:** 00:04:54 ## Summary - An intern at ByteDance (TikTok’s parent) stole a large number of GPUs by sabotaging internal AI training pipelines, leading to a $1 million lawsuit and his termination in August 2024. - The intern, named Kouan, used the stolen compute time to develop a paper on “Visual Autoregressive Modeling: Scalable Image Generation via Next‑Scale Prediction,” pushing the field beyond token‑ or pixel‑level prediction toward reasoning over larger image concepts. - Despite the theft, Kouan submitted the paper to NeurIPS (the premier AI conference) and, after a blind review that judged the work solely on merit, the conference awarded it Best Paper in December 2024. - The award sparked controversy, as the conference organizers knowingly recognized work done with stolen resources, while ByteDance remains outraged and continues legal action against Kouan. ## Sections - [00:00:00](https://www.youtube.com/watch?v=6A3NOedlPWI&t=0s) **Untitled Section** - ## Full Transcript
this is the story of the craziest
internship I have ever heard of happened
in AI it's still unfolding this person
defrauded the company they stole gpus
which is the most precious resource in
AI they've been sued for a million
dollars and they're not done yet they
just won best paper at the most
prestigious AI conference on the
planet their name is
kouan and he started at bite dance which
is the parent company of tick tock back
in the middle of 2024 so like
jish immediately things went wrong
something started to happen so his
colleagues models would
fail his their training runs would crash
naturally during during large training
runs there would be small innocuous file
edits that would pass and somehow the
pipelines would be sabotaged and no one
figured out what was going on but the
net net of it was he was able to change
model weights he was able to hack
machines and he caused enough of the AI
training and research pipeline at bite
dance to fail that he freed up a
significant number of gpus which he used
for his own academic paperwork that's
what he wanted his whole goal was to get
access to
gpus well when bite dance figures this
out in August they terminate him bye-bye
fired fired for malicious interference
bite Dan then reports his behavior to
his university and begins investigating
the extent of the damages he's caused
they are very upset about this but Tian
isn't done writing he keeps writing and
in October of 2024 he submits his
research paper visual autor regressive
modeling scalable image generation via
next scale prediction to nurs which is
the most prestigious AI conference on
the
planet talk about like wow right like
the the the willingness to basically say
yeah I stole the gpus but look at what I
did it's so incredible you have to look
at this that was what happened and if
you're wondering what scalable image
generation via next scale prediction is
he's moving past just next token
prediction or next pixel prediction and
actually looking in images at how you
can have a larger concept to translate
scale more effectively and one of the
things that is at The Cutting Edge of AI
in late 2024 is how do you reason
against larger chunks than just a to
we saw it very recently with um deep
seek V3 doing double token prediction
we've seen it with a paper from meta
that's looking at reasoning across
Concepts this is very much in that vein
but apparently it was such a good paper
that in December very recently the
judges at NPS blind awarded the best
paper at nurs to Kon the intern who
stole the
gpus and and I say blindly because they
measured the paper quality without
looking at names they didn't know this
was who it was now obviously the
conference organizers knew who it was
when they awarded it and they still
chose to award it and there's a lot of
controversy about awarding best paper to
someone who stole
compute and bite dance is certainly mad
about it because when they saw that the
paper was submitted using their stolen
GPU time they sued him in Beijing
demanding a public apology and demanding
$1.1 million in Deb Imes roughly this
was back in November that court case is
still pending so this guy now has a
court case for a million bucks against
him best paper award at nurs and a
massive controversy around what he
did and the where where I come down on
this at the end of the day is you have
someone who is brilliant enough that
they can figure out how to hack the AI
modeling pipeline of a major model
builder and AI researcher and they can
do that for their benefit and they can
get a groundbreaking Innovation out of
it you want to employ them you just want
them to have a very good manager with
tight constraints if you don't employ
them it will be worse because they will
figure out a way to contribute to this
field it is evident that they will not
be stopped from contributing to the AI
field it's about whether you employ them
or not so I would expect that someone in
the model maker space is going to decide
to bite the bullet cover the liability
for the damages sued for or settle out
of court and get this guy employed as
long as they have very very tight
constraints because they want the
innovation in the house they just don't
want the liability that comes from him
being a loose cannon so we will see what
happens it's still unfolding but the
story of Kon is already the wildest
internship story I have ever heard you
tell me if you've heard something Wilder
but this is just nuts