I build generalist policies and share my learnings here | ML @bracketbot | Researcher @UCLA's BAIR Lab | Software Engineering @UWaterloo
Joined September 2023
- Tweets 889
- Following 1,397
- Followers 2,707
- Likes 2,648
66 Photos and videos
Pinned Tweet
Jun 5
I just spent months handwriting a 200 page guide on the entirety of ML foundations and math from scratch.
The guide features:
- Neural Nets (Backprop, Adam, SGD, Batch Norm)
- ML Algorithms (SVM, Grad Boosting, K-means, PCA)
- Hardware (Tensor Cores, Systolic Arrays, CUDA)
- Transformers (Multi-Head Attn, KV Cache, LoRA)
- Vision (ViT, Convolutions, MAE, IoU, NMS, VLM)
- Agents (OpenClaw, ReAct, Memory, Orchestration)
Everything I wish I had years ago, for free.
143
337
2,828
276,626
Arjun Virk retweeted
Jun 5
I just spent months handwriting a 200 page guide on the entirety of ML foundations and math from scratch.
The guide features:
- Neural Nets (Backprop, Adam, SGD, Batch Norm)
- ML Algorithms (SVM, Grad Boosting, K-means, PCA)
- Hardware (Tensor Cores, Systolic Arrays, CUDA)
- Transformers (Multi-Head Attn, KV Cache, LoRA)
- Vision (ViT, Convolutions, MAE, IoU, NMS, VLM)
- Agents (OpenClaw, ReAct, Memory, Orchestration)
Everything I wish I had years ago, for free.
143
337
2,828
276,626
Jun 5
I just spent months handwriting a 200 page guide on the entirety of ML foundations and math from scratch.
The guide features:
- Neural Nets (Backprop, Adam, SGD, Batch Norm)
- ML Algorithms (SVM, Grad Boosting, K-means, PCA)
- Hardware (Tensor Cores, Systolic Arrays, CUDA)
- Transformers (Multi-Head Attn, KV Cache, LoRA)
- Vision (ViT, Convolutions, MAE, IoU, NMS, VLM)
- Agents (OpenClaw, ReAct, Memory, Orchestration)
Everything I wish I had years ago, for free.
143
337
2,828
276,626
May 19
Life Update: I've moved to SF to build the future of robotics learning @bracketbot with @sincethestudy.
My research focuses on unlocking continual learning for robotics policies.
More soon.
ALT San Francisco
18
4
138
8,710
Arjun Virk retweeted
May 14
Every ANN method either gets slower as your corpus grows or starts retrieving the wrong things.
We built one that does neither.
50K vectors → 10 µs / query
1M vectors → 10 µs / query
5M vectors → 10 µs / query
Introducing MeshRAG: retrieve anything with the same speed, same accuracy, no matter how much data you have.
Here is how we did it 👇
12
18
69
34,025
Arjun Virk retweeted
May 11
People talk, listen, watch, think, and collaborate at the same time, in real time. We've designed an AI that works with people the same way.
We share our approach, early results, and a quick look at our model in action.
thinkingmachines.ai/blog/int…
464
1,958
15,786
7,748,866
Arjun Virk retweeted
Apr 7
Releasing the Unfolding Robotics blog!
Time to unfold robotics: we trained a robot to fold clothes using 8 bimanual setups, 100 hours of demonstrations, and 5k GPU hours.
Flashy robot demos are everywhere. But you rarely see the real story: the data, the failures, the engineering.
We’re sharing everything: code, data, and details in the blog → huggingface.co/spaces/lerobo…
30
138
793
280,321
Apr 2
Interesting how OAI is suddenly acquiring all these companies.
Similar parallel to Meta’s strategy.
Apr 2
TBPN has been acquired by OpenAI!
The show is staying the same and we’ll continue to go live at 11am pacific every weekday.
This is a full circle moment for me as I’ve worked with @sama for well over a decade. He funded my first company in 2013. Then helped us fix a serious logjam during a critical funding round a few years later. When I took my second company through YC, he was president at the time, and then when I joined Founders Fund, the first deal I saw in motion was the post-ChatGPT round in late 2022. And as we started growing TBPN last year, he was the very first lab lead to join the show.
Thank you to everyone that has been a part of TBPN until now. The last year has been the most fun and rewarding part of my career and we’re excited to have more resources than ever going forward.
5
900
Arjun Virk retweeted
Apr 2
Sam Altman predicted in 2024 that a one-person billion-dollar company "would have been unimaginable without A.I., and now it will happen."
He just emailed the NYT saying he won a bet with tech CEO friends over when it would arrive, and that he "would like to meet the guy."
The guy: Matthew Gallagher, 41. Spent $20K and two months building a GLP-1 weight-loss telehealth company out of his living room in LA.
The stack: ChatGPT, Claude, and Grok writing code. Midjourney for images. Runway for video ads. ElevenLabs handling customer calls. Custom AI agents stitching it all together.
$401M revenue in year one. On track for $1.8B this year.
216
413
5,484
1,401,014
Liftoff.
The Artemis II mission launched from @NASAKennedy at 6:35pm ET (2235 UTC), propelling four astronauts on a journey around the Moon.
Artemis II will pave the way for future Moon landings, as well as the next giant leap — astronauts on Mars.
3,808
55,159
177,851
14,275,621
Apr 1
If this came to Waterloo, classes would overfill.
Bring this roster to the loo…
1
21
1,792
Mar 30
Very bullish on Icarus
Mar 30
We made it possible to make humans extremely intelligent using non-invasive hardware.
Far beyond biological limits.
My friend from high school, Rayne, and I moved in together for college and accidentally developed an ML architecture that learns at a pace uncommon to traditional ML.
Out of curiosity, we connected it to a high-volume trading pipeline, and it began averaging 1.8% daily returns on real market data with simulated capital.
Our architecture creates extremely compressed memory representations, something closer to episodic memory.
It learns patterns unusually quickly, retains them, and uses them to synthesize new insights.
That's when Rayne and I had a four-hour conversation about the future of AI and realized that if everyone is building AI to replace humans, we want to build AI to enhance humans.
Since middle school, I've been rigorously studying ML to make superintelligence possible. But superintelligence could make humans obsolete.
The only way to safely harness superintelligence is to democratize it by giving everyone access to systems that amplify human thinking rather than replace it.
We made that possible at Icarus Cybernetics.
Not a tool or an assistant, but a new way to think.
I want to augment the intelligence of our entire civilization so we can do things we simply couldn't before.
I think we've made that possible with some of the other research we've done, which won't fit in a single tweet.
We're in SF now, raising our seed and looking to meet a small number of people thinking seriously about this.
If you're interested in chatting, DM me.
---
Old photo of my roommate/co-founder after seeing the 1.8% daily returns.
1
14
1,660
Arjun Virk retweeted
Mar 29
let’s break this list down so it’s actually useful
• JEPA / H-JEPA: avoids predicting every single pixel (too expensive) and rather predicts in latent space. H-JEPA adds hierarchy - short term details vs long term planning ie. how humans actually learn
• I-JEPA: built for very efficient vision models. Masks image patches and predicts the semantics and in doing so bypasses heavy compute of traditional autoencoders
• MC-JEPA & V-JEPA: both of these are built for videos. MC-JEPA separates content (what an object is) vs motion (how it moves). V-JEPA masks video features with no text labels making it perfect of action tracking at scale
• Audio-JEPA: filters out background noise by treating sounds like visuals
• Point-JEPA & 3D-JEPA: used primarily in AVs. Uses LiDAR point clouds & volumetric grids
• ACT-JEPA: filters out real world noise to learn manipulation tasks efficiently via imitation learning
• V-JEPA 2: predicts future physical states of the world caused by an action before it happens
• LeJEPA: replaces techniques like masking with an Energy-Based Model (EBM) which mathematically prevents "feature collapse" & ensures the model scales reliably as data increases
• Causal-JEPA: for learning true cause-and-effect physics by applying object level masking
• V-JEPA 2.1: great for spatial grounding since it combines a dense predictive loss across image & video
• LeWorldModel: built directly on LeJEPA's math but super compact - 15M params
• ThinkJEPA: uses dense physical prediction with VLM reasoning. Best used when long-term strategy is needed
Mar 29
14 most important and influential types of JEPA
▪️ JEPA / H-JEPA
▪️ I-JEPA
▪️ MC-JEPA
▪️ V-JEPA
▪️ Audio-JEPA
▪️ Point-JEPA
▪️ 3D-JEPA
▪️ ACT-JEPA
▪️ V-JEPA 2
▪️ LeJEPA
▪️ Causal-JEPA
▪️ V-JEPA 2.1
▪️ LeWorldModel
▪️ ThinkJEPA
Save the list and check this out to explore these JEPA milestones as a map of AI progress: turingpost.com/p/jepamap
10
95
661
82,342
Mar 28
wow this is better than tinkercad
Mar 28
SOMEONE VIBECODED AN APP WHERE YOU CAN BUILD AND TEST ELECTRONIC CIRCUITS DIRECTLY IN YOUR BROWSER.
Community note
The video shows an unmaintained circuit simulator from 2022, not a newly "vibecoded" app. The developer confirms it's no longer updated and should be used at your own risk.
x.com/austin_malerba…
x.com/austin_malerba…
1
1
8
1,216
Mar 27
Best example of taking something ordinary and building something out of the box. Who knew you could use your mac in this way?!
By far the most impressive demo @socraticainfo.
50
4,240
Arjun Virk retweeted
Mar 24
Introducing TurboQuant: Our new compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x speedup, all with zero accuracy loss, redefining AI efficiency. Read the blog to learn how it achieves these results: goo.gle/4bsq2qI
1,007
5,712
38,826
19,379,867
Arjun Virk retweeted
Mar 24
🚨 BREAKING:
Amazon acquires Fauna Robotics! 🤯
Another week, another Amazon acquisition in robotics space.
This time it's a consumer humanoid company.
@amazon just acquired @faunarobotics, a startup making friendly humanoid robots.
Sprout is a $50,000 bipedal robot that's about waist-high on an adult and weighs 50 lbs (22,5 kg). It's designed to be "approachable and human-friendly, meaning it looks and acts less intimidating than typical industrial robots.
Fauna was founded in 2024 by former Meta and Google engineers in New York.
Early customers already include Disney and Boston Dynamics.
Rob Cochran, Fauna's CEO: "We are thrilled about what joining the Amazon team means for our future. We will proudly operate as Fauna Robotics, an Amazon company."
They already have millions of customers who trust them in their homes (Alexa, Ring, deliveries). Adding a helpful home robot might be a good next step. 🏠
Read more: cnbc.com/2026/03/24/amazon-h…
~~
♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com
7
25
152
13,770
Arjun Virk retweeted
Mar 24
Google DeepMind 🤝 Agile Robots
Our new research partnership will integrate the Gemini foundation models with their hardware to help build the next generation of more helpful and useful robots.
Find out more → goo.gle/4lKu7de
59
217
1,741
331,221