Eric Chen
@erico_cheny
Followers
706
Following
767
Media
155
Statuses
378
hacking on ai and 3d at stealth_startup | gradient ascending towards my life's work @MIT 2025
San Francisco, CA
Joined February 2023
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Flamengo
• 285318 Tweets
#いい推しの日
• 129342 Tweets
Gabigol
• 129291 Tweets
Eagles
• 112741 Tweets
#GHDBT9
• 89118 Tweets
Lions
• 88124 Tweets
Saquon
• 84986 Tweets
Bell
• 70566 Tweets
Bears
• 67457 Tweets
連休最終日
• 67085 Tweets
Branch
• 60507 Tweets
Cowboys
• 55934 Tweets
Packers
• 48397 Tweets
NASCAR
• 42486 Tweets
Nekfeu
• 33853 Tweets
Riquelme
• 28271 Tweets
Seahawks
• 26135 Tweets
Caleb
• 24424 Tweets
Rams
• 24218 Tweets
Gago
• 19715 Tweets
Geno
• 19605 Tweets
Sirianni
• 18684 Tweets
振替休日
• 17318 Tweets
Byron
• 16206 Tweets
Eberflus
• 14928 Tweets
Poles
• 13270 Tweets
Jordan Love
• 12128 Tweets
Stafford
• 12071 Tweets
Jags
• 11455 Tweets
Lanús
• 10554 Tweets
Hi I’m Eric. I’m 19.
I immigrated to the US 14 years ago. Lived in the NYC projects for 11 years.
This summer, I’ll be going to Silicon Valley to work at
@zoox
on driving tools for autonomous vehicles.
Today was my first day :)
4
4
57
Today I generated 3d from text.
- Trained diffusion transformer on sunset of ModelNet10
- Got tensorboard working
Kinda blows my mind this even works.
I’ll add a lr scheduler (cosine annealing with warmup) and send it on the full dataset
3
1
40
Doing a deep dive into diffusion models. I’ll start by diffusing sine and cosine waves then work my way to building a text to 3d model from scratch.
Documenting everything in a series of jupyter notebooks. Let’s go.
Inspired by
@MajmudarAdam
6
0
30
Took some weeks off to study computer graphics. Now back to doing ai.
Will religiously read and reimplement some papers.
2
1
23
Implemented EMA for VQVAE
This improved training stability a ton. For the version w/o EMA, loss spike in the beginning and would gradually increase in the middle of the run. That doesn't happen anymore.
2
1
18
Going to speed run learning C++ and build a raytracer.
I've never done any low-level programming before so this should be fun. Nice break from ai.
5
0
15
i learned about RNNs and implemented one from scratch in numpy.
forward pass, cross entropy loss, backprop thr. time, gradient clipping, gradient descent, and sampling.
repo:
more details below
1
0
13
My FIRST time training a text-conditional model.
This is a tiny diffusion transformer that generates sine / cosine waves from text.
Kinda magically this works. Will get this working on 3d soon.
0
0
12
learning about RNNs and backpropagation through time.
this makes me so grateful autograd exists.
1
0
11
Here’s what I’ve been hacking on — a tiny text-to-3d diffusion model. This was a challenge to learn diffusion and build something cool.
Some notes and some demos :)
Prompt: "chair"
2
2
11
Kicked off VQVAE run
- loss is unstable in the beginning, probs cuz it's trying to best use the codebook
- perplexity going up
- recon on test set looks pretty good
3
0
9
will open source the rnn code after i add some bells and whistle.
- train loop
- saving + loading weights
- inference script
0
1
8
Just ran into a segmentation fault in C++. Ended up being an issue w/ dangling pointers.
How in the world do u guys even debug this? No stack trace. No nothing. Feels like driving blind.
2
0
8
Added lighting to the raytracer! Now the sphere looks 3d.
Next I'll add some shadows.
1
0
7
I figured out the secret to debugging segmentation faults.
To just not write bugs 💀.
Plz just lemme access the memory i want to access.
1
0
7
My attempt to derive backpropagation from first principles. More to go.
1
0
6
Render some cylinders rn
Once again, high school math of solving the quadratic formula comes in handy. Computer graphics just makes so much sense intuitively.
1. Hollow cylinders
1
0
6
Note to self
One of the most important lesson I’ve learned working at a startup.
Being permissionless and just doing the work.
My thoughts on this …
2
1
5
Found a bug in the VQVAE where decoder was not expressive enough. I passed the wrong channel number.
Anyhow, the quality reconstructed images improved a ton after and interpolations look pretty sick.
2
0
6
Implemented groups in the raytracer.
Lighting, shading, shadows, and transformation matrices all coming together to make this shape.
0
0
6
Something about programming something and having a visual output.
Implemented stripe, gradient, ring, and checker patterns in my raytracer.
0
0
6
ok I figured out why my diffusion model was doing ass during the big training
during validation, the guidance scale is set to 1. not enough conditioning strength. I upped it to 2.5 and it works.
1
0
5
Self-teaching myself AI by reading the original papers and training neural nets from scratch.
It's hard and painstakingly difficult. But the few days when I get something to work, it's absolutely magical. This rather "trivial" result by today's standards still blows me away.
0
1
4
This the year I become a consistent reader.
Smart people have written their life learnings and experiences in a form absorbable in a couple hours. Why haven’t I gotten started absorbing?
Here’s to reading my first book of 2024.
0
0
3
Rode in the VH6 (aka the toaster looking vehicle) today
@zoox
. Excited to see where this goes 🙃.
1
1
5
@0xluffyb
For the longest time, I associated research with academia which I did not want to do.
Then I found out it's just reading, implementing stuff, and making changes to what you've built.
This is fun and not scary.
1
0
5
This marks my first month living in SF. Working for a stealth 3D AI startup. It's the BEST decision I've made.
Some observations I made along the way
1
1
5
Going to deeply understand the output of YOLO today so that I can code the loss function
Learn, build, iterate
0
1
5
I added shadows to my raytracer.
Cast a ray from a point to the light. If an object sits in between, draw a shadow.
0
0
5
today i moved to SF. people come here to build the future. i guess i’m just one of them.
ai is what i will work on. it’s unbelievably exciting. more than ever it’s time to build
i hope to make lifelong friends here. anybody want to meet :)
1
0
4
I think I know how skip connections in UNets work now. Improved Diffusion repo from openai was a big help.
0
0
4
My learnings from Imagenet Classification with Deep Convolutional Neural Networks.
1. Architecture: 5 convolutional layers followed by 3 feedforward layers and a softmax
2. CNNs are faster to train b/c of fewer params and connections
(Cont.)
1
2
4
After ~50k steps on ModelNet10, the model does decent. Real happy with the chairs and computer monitors.
0
0
4
Downloaded and processed ModelNet10 and ModelNet40.
Next: Write a train loop for my diffusion transformer
1
0
4
Kicking off text-to-3d training on modelnet40.
153838596 param DiT (most of this is CLIP lol)
100k steps
lr=1e-3
cosine annealing w/ 500 warmup steps
batch_size=32
cosine noise schedule
Giving it some pretty out of domain prompts to validate against.
1
0
4
@CharlesLiu9
@onlychans1
@_yashkarthik
@ZSteven_
@albertlai631
@song_minjune
@freemanjiangg
@mayankja1n
@hud_zah
@sidyakinian
@edwardjxli
You guys r the best ❤️
@CharlesLiu9
@onlychans1
1
0
3
7 mil point cloud diffusion transformer at 4000 steps.
this is the worst part of training. early results look ass and u just gotta wait
1
0
4
1
0
2
Made my first raytraced scene in C++!
A lot happens behind the scenes to render this simple sphere. Thank you dot products and the quadratic formula.
1
0
4
Training VQVAE2, a hierarchy of VQVAEs. The top level latent captures global info and the bottom level latent captures local details.
Couple steps into training and the recon quality is amazing.
1
0
5
There's a big difference b/w what I know and what I think I should know. Especially in AI.
Gonna religiously dedicate time to becoming good.
Focus this week. Transformers.
0
0
3
Learned the basic syntax of C++.
- variables
- functions
- arrays
- control flow
- pointers
- classes
Onto the Ray Tracer Challenge.
2
0
3
TIL about view transform matrices.
Now I can put the camera anywhere and render my lil raytraced scene.
1
0
3
4. Reviewed backpropagation with this awesome video from
@3blue1brown
High level idea is how a tiny nudge in a neuron's weights and biases affect the network's output. We calculate this with the chain rule.
1
1
3
I've been holding this off for such a long time.
This is the month I finally start a blog :)
0
0
1
I wrote my "hello world" for ray tracing and this seems magical to me.
Shoot rays, check if they hit an object, and spit out a color.
Back to learning
0
0
2
@sam_postelnik
Sam marketed the hell of out of this. Big reason y we got to 5000+ users.
Anybody looking for a product person who can code,
@sam_postelnik
is the person :)
0
0
1
Man, PCA math is kinda hard. This blog does a great job of explaining it.
0
0
3
Yep...
A debugging trick I found for this. Rewriting the entire code b/c you can't find the silent bug.
0
0
2
Sitting down for an extended period of time to deeply understand a concept is sooo invaluable.
Been too accustomed to FAST and not struggling.
A sign I need to embrace doing hard things more often.
0
0
3
After banging my head for a week, I finally got mini AlexNet (what I called my scaled-down version) to output something “intelligent” on the test set.
Moving on to creating a web interface to further test how this model performs on unseen data.
1
1
3
Sometimes there’s work that’s so exciting, u just can’t it put down.
I think I found it today.
0
0
1
Stanford CNN for Visual Recognition Course
Lec 6: Training Neural Networks
0
1
3
Listening to
@naval
podcast on “How to Get Rich” every morning on my commute.
Learning foundations (ai, math) and becoming the type of person to attract luck (twitter, making friends with interesting ppl … etc) is gonna be the core focus of my summer.
0
1
3
@KrishivThakuria
Had a conversation about this today. Once you realize u can do anything, it’s quite empowering.
Takes a bit of rewiring to see this. Time to build :)
1
0
3
I will not get up from my desk until I get this training run kicked off ...
1
0
3
@sam_postelnik
@JosephKChoi
Love this. Andreessen’s perspective on career as a portfolio of jobs/opportunities is gold.
Makes u think more long term and willing to take risks
0
0
2
Hey
@scale_AI
I applied to the July 15 GenAI hackathon and this is why you shouldn’t accept me.
I used Gen AI for evil. Created a lil ai product
@sam_postelnik
called FGenEds (short for “fuck gen eds”) to help college students study less.
1
1
3
The entire field of probability stems from 3 axioms. A shit load of math from JUST 3 axioms.
What if the key to building a shit load of intelligence into LLM agents is simply just finding the right set of primitives (tools)?
0
0
2
Created my first "hello-world" agent from scratch to answer questions about the weather. Such a simple and trivial use case but I'm fucking blown away 🤯.
Excited to continue building
Lots of help from the ReACT paper
0
0
3
Took some time off tech to make some crepes this weekend and a virtual date with my gf.
Thinking about cooking on the regular.
0
0
2
Implemented YOLOv1's loss function in Pytorch.
Recurring lesson for me: everything's fucking hard until you do it :)
Moving on to writing the model, training it, and figuring out whatever non-maximum suppression, MAP, and blah blah is
0
1
2
just realized how much better i got at writing scripts. old me would write a for-loop and call it a day. now i use multiprocessing.
0
0
2
2 pull requests done today. Feels good to finally contribute something
0
1
2
Made an agent that reads articles from the internet, does a task on it, and sends the result to me on WhatsApp. This stuff is kinda magical ✨
0
0
2
@thedanigrant
@davefontenot
I would love to go but the waitlist is full. Any chance of an open seat 🙃
1
0
1
Starting to read You Only Look Once: Unified, Real-Time Object Detection.
2
1
2
Reviewing every god damn thing I learnt in ML
Will also need to go back to studying foundations - stats, probability, linear algebra
All in on AI. Day 1
0
0
2
still lots of gaps in my knowledge. will take some time off to read & learn before starting my next side quest
for now, I'll go to sleep
Prompt: "bed"
0
0
2
1. Denoising Diffusion Probabilistic Models
Noise an image and train a network to "denoise" it.
Questions after reading this
- where did the math come from?
- where did the math come from?
- where did the math come from?
1
0
0
#include
is pretty much like an import statement.
Need to figure out what a header file is.
Will report back after I get my first Raytraced scene.
Onto implementing matrix operations.
1
0
2