
Co-Founder & CTO at Stealth Startup | (research ⋃ prod) ⨉ (real-time ⋂ vision ⋂ graphics ⋂ AI/ML) | Ex @nvidia, @Google, @MSFTResearch | I read ⊂ the code.
Bay Area, CA
Joined March 2011
- Tweets255
- Following986
- Followers876
- Likes749
Pinned Tweet
3D reconstruction with 1734 images in 2 mins 15 seconds!
Sequential matcher with loop closure ended up dropping 4 images, but for the speedup over all-pairs matching (~10 mins) might be worth it.
Sameh Khamis retweeted
a tracked robot drove 9,760 metres of closed loops through two active construction sites.
revisiting the same ground makes these recordings useful for loop closure and traversability work. an RGB-D camera, a Livox lidar, two IMUs, GNSS, and wheel encoders, over 1 hour 45 minutes.
i parsed it into FiftyOne so you can scrub camera, depth, and lidar together.
huggingface.co/datasets/Voxe…
#PhysicalAI #Robotics
Sameh Khamis retweeted
Watching commercial air traffic dodging a volcanic ash plume from orbit is pretty damn cool.
Tracked a 7.8 km bike ride with an Insta360 X5 at ~40 fps!
Built ARKit-style VIO with loop closure in Rust (surprise!) - total ride ATE is 105 m RMSE on 7.8 km.
From the Insta360 images + IMU to my Android phone, no mapping, no GPS, but that'd have helped with the scale!
Sameh Khamis retweeted
Choose life. Choose a train, a statue, a pipe organ. Choose a 1967 Volkswagen Beetle, a K6 red telephone box in London.
We chose not to choose. We reconstructed them all.
Introducing 🚂 📸 Trainspotting: 12,068 objects reconstructed in 3D from 311,886 found photos.
🧵
1/8
Sameh Khamis retweeted
Chen and Snavely, "Trainspotting: Learning Object-Centric 3D Reconstruction from Found Photos"
Given a collection of photos of the same object, you can now do feed-forward geometry estimation on them. Surprisingly, it can also improve performance for normal workloads.
Oh no I went too far out of bounds!
Sameh Khamis retweeted
18 of the 20 best open models today are Mixture-of-Experts. Dense is basically over.
And yet the know-how to train a good MoE still lives mostly in the heads of a handful of teams.
So we spent 800k GPU hours running the experiments ourselves and wrote everything down. Super happy to finally release "The Ultimate MoE Training Handbook", free and open, organized around the questions training teams hit:
- how many experts, how big, shared or not?
- why is half my step spent in all-to-all, and how do I hide it?
- why is my run 25x slower with no error in the logs?
- softmax, sigmoid or sqrt(softplus) router, and why keep it in fp32?
- aux loss or DeepSeek's loss-free bias?
- upcycle a dense checkpoint or train from scratch?
- can I serve with fewer experts per token than I trained with?
⚠️ Warning: it's long (~70k words, budget a few days, sorry) but with 55 interactive figures, code you can run and numbers from our own runs.
Read it here: huggingface.co/spaces/huggin…
Sameh Khamis retweeted
Jeff Dean says recursive self-improvement running infinitely fast is “implausible,” because there’s actual physics involved
“But I think if you think about what needs to be involved in creating the base model and making it better, there’s lots of actual training and experimentation that needs to be done, and that’s not going to happen infinitely fast, right?"
"There’s just actual physics involved. You need to run some computations. You need to look at the results. The computers need to do their thing, and you need to save the data, have another model look at the results in order to decide what to do next."
"So I think some people have this illusion that you’re going to be able to run this loop infinitely fast, and I just find that kind of implausible. And it’s good to be grounded in reality. Having automated systems that can try to improve things is definitely going to improve the rate at which we can improve those systems, but it’s not going to make them go infinitely fast.”
Jeff Dean reveals Discovery Loop did 400 interviews, mostly at Cafe Venetia on University Avenue, in its first three weeks and already has about 20 automated improvement loops running
“So what have we done so far? First three weeks we did 400 interviews, mostly at Cafe Venetia on University Avenue, to find amazing people, which I talked about some of them. Finding office space, we’re in the old Zibibbo restaurant, so we have a funky space for us."
"We’ve built our own agent harness that we’re using for all of our own development and for automation. That harness has some really nice collaboration features, so you can think of it as a system that you interact with a model to have conversations and ask it to solve things, but you can also invite your teammate in to collaborate. And so then the two of you can talk to a model and an agent. And you can also have other people add comments to the thing you’re trying to do, and propose a design to the agent before you propose that."
"We’ve built really agent-first infrastructure so that it will scale to having tens, hundreds of thousands of agents doing things rather than sort of hundreds of people doing things. We think that’s really important, and we’ve got probably on the order of 20 automated improvement loops running in the way in which we imagine that we want to solve problems.”
Sameh Khamis retweeted
Video for my @rust_nyc talk, "Code Matters: Reflections on Rust in the LLM Age", is now up:
youtube.com/watch?v=5D7QPOIZ…
Sameh Khamis retweeted
did you know that it's been 21 days since an insane Linux vulnerability was publicly reported, and Ubuntu STILL HAS NOT PATCHED IT?!
debian patched it the next day. arch, alpine have also patched it since.
but not ubuntu
Sameh Khamis retweeted
Holy crap! This is not a drill - the mother lode has dropped for FREE.
The full Quixel Megascans library is going free. Fab listings will update between now & Nov 12, so check back if an asset you want isn’t free yet.
We’ve updated terms for all Quixel assets so you can train AI models on Megascans & Megaplants.
Learn more: epic.gm/quixel-megascans-goi…
Sameh Khamis retweeted
Replying to @Jvthewanderer @t_looman
What we are trying to do: Build an engine that’s more powerful than UE5, easier to develop in than Unity, and is the most efficient effective way to build with AI acceleration.
The problem with having old systems and new systems side-by-side is the tendency to mix them, and then their interoperability becomes a giant tangle of blueprints and verse, actors and entities.
We thought that side-by-side was unavoidable, but AI assisted conversion is now clearly way more effective, and will produce a way more manageable UE6 project than a hybrid, while making it way easier for us to simplify and improve the engine.
We’ll be proving it out by converting Fortnite, the game with the largest base of blueprints in existence, to the new approach.
Sameh Khamis retweeted
Obscurial. #threejs
Live: dasprinzip.com/tinker/day51/
Yes, I know - the real "Obscurus" from the movie "Fantastic Beasts and Where to Find Them" looks different. Nevertheless, I think it was the most fitting name, for my tinkering with tons of particles in WebGPU & TSL. #threejs
Get the code here: github.com/prinzipiell/tsl
#indiedev #gamedev #creativecoding #shaders #wgsl #webgpu #glsl #webgl #3d #tsl
Sameh Khamis retweeted
There's a lot going on in my GPU. I'm making a GPU Taskman to get insights to where my GPU memory is going. PRs welcome
Sameh Khamis retweeted
📢📢📢 Introducing 𝐒𝐨𝐥𝐯𝟑𝐑
Can geometric foundation models move beyond fixed-set prediction and keep solving like classical SfM?
We turn a looped Transformer into a reusable 𝐥𝐞𝐚𝐫𝐧𝐞𝐝 𝐠𝐞𝐨𝐦𝐞𝐭𝐫𝐢𝐜 𝐬𝐨𝐥𝐯𝐞𝐫
🌐 Project: ubc-vision.github.io/solv3r
🧵👇
Sameh Khamis retweeted
i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.
Sameh Khamis retweeted
For anyone questioning whether this post is overhyped, I think they are actually *underselling* the approach.
We did similar experiments and found that Claude models can write entire, functional neural net weights from scratch! csinva.io/weight-space-progr…
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Sameh Khamis retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
























