Sameh Khamis (@sameh_khamis) header
Sameh Khamis (@sameh_khamis) profile picture

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
Photo by Sameh Khamis (@sameh_khamis)Photo by Sameh Khamis (@sameh_khamis)Photo by Sameh Khamis (@sameh_khamis)Photo by Sameh Khamis (@sameh_khamis)Photo by Sameh Khamis (@sameh_khamis)Photo by Sameh Khamis (@sameh_khamis)Photo by Sameh Khamis (@sameh_khamis)Photo by Sameh Khamis (@sameh_khamis)Photo by Sameh Khamis (@sameh_khamis)Photo by Sameh Khamis (@sameh_khamis)
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.
8
18
1
209
11,012
Sameh Khamis retweeted
Say what you want about LLMs - and how smart you are without them. I can say without hesitation that they’ve made me smarter. Being able to interrogate an idea, ask anything, and go as deep or sideways as I want has changed how I think. Bottom line: I’m learning more, and better, than at any other point in my professional life.
92
160
29
1,528
50,875
Sameh Khamis retweeted
Apple just released Depth Pro 2 Sharp and Temporally Consistent Metric Monocular Depth HF: huggingface.co/apple/DepthPr…
56
75
18
871
89,949
Sameh Khamis retweeted
just being able to figure out the *precise* 3d location of your real world captures unlocks so damn much - from change detection to ar/vr and robotics.
22
33
3
307
18,210
Sameh Khamis retweeted
Getting streaming autoregressive view synthesis to hold together like this takes getting a long list of tricky details right. Great execution by @jensenzhoujh! Check out the results on the project page.
Meet AVM, a generative in-context Autoregressive View Model. Given one or more images or videos of a complex 3D scene, AVM streams 2D views as you steer a virtual camera. It reconstructs what the context supports and imagines the rest without building an explicit 3D representation. Find out more at autoregressive-view-model.gi….
1
1
33
2,030
Academia has a problem: Every author wants to use LLMs to write papers, but no reviewer wants to read AI (assisted) content. And every reviewer wants to use LLMs for reviews, but no author wants their paper judged by an AI. Authors & reviewers are the same people though...
211
365
123
2,866
294,624
Sameh Khamis retweeted
15
24
2
272
8,609
Totally agree. There are many risks, but AI deciding to kill us all is BS, instead humans will use AI to do bad things (and good things too), just as with other tech (only faster).
absolute fanfic. AI has real, immediate, tangible dangers that we need to focus on, for instance the impact on the job market and making sure random script kiddie can't use it for cyber offense. But this "oh once it's superintelligent it will take over and ensave us, maybe kill us all!" is pure sci-fi fanfic and distracts from the real (and difficult) issues, which linear probing won't solve. It's worth repeating the LLMs don't just "decide" to do something and then just "do it" in a void. They generate tokens based on our prompts, only as long as we call `model.forward()`, and on giant clusters with terabytes of weights, that's impossible to "quickly self-replicate". We also need to regularly remind that the OAI/HF incident was specifically a _guardrail-free_ cyber eval where the model is _instructed_ to hack, to find out how good it is at it. Some recent personal event made me realize there needs to be more push-back on this doomsday nonsense, and we need to re-focus the discussion on real dangers/issues. I won't argue much in the replies though, not interested in peer-reviewing Terminator fanfic.
14
24
4
332
37,298
Sameh Khamis retweeted
my AIE talk is up! come hear how the grok @bot team ships so quickly without compromise on quality youtu.be/O_zcc6Xl12g
463
983
297
10,597
3,014,327
Sameh Khamis retweeted
We're excited to release the first open-source VIO to beat Aria on LaMAria! Industry level SLAM for everybody will power tons of applications, can't wait to see what people build :) So how did we beat Aria's SLAM? (1/4)
when LaMAria was introduced a year ago, Meta's closed-source Aria SLAM was far ahead of academic systems. one year later, we've surpassed it in 3/5 categories: medium, long, and low light. here's what closing that gap looks like:
32
80
11
858
435,479
Sameh Khamis retweeted
Some of you asked for my 🌋 Krakatoa repo, so here it is. Real-time 3D Navier–Stokes smoke. One HTML file, only 121 KB, pure WebGL2. Made with Claude Opus 5.5. GitHub repo in the comments. Enjoy!
4
1
4
340
Thanks Kostas! I'm still trying to figure out how it works...so far it's been mostly dog videos. Strange because I am more of a cat person actually...
7
1
66
4,436
Sameh Khamis retweeted
how should multimodal models encode visual input? → pretrained encoder? → encoder-free? → before the LLM? published models bundle these choices together. we trained 200+ models from scratch to see which of these choices matters only one of them does 🧵 [1/n]
25
139
29
855
133,158
Sameh Khamis retweeted
Full 3D replay of the FAA ground stop, and everything we know so far, all in 3 minutes. I used OSINT to reconstruct cascading impact across US airspace, the reopening, plus what officials say caused the longest nationwide ground stop since 9/11.
6
12
1
70
5,818
Sameh Khamis retweeted
Watch the U.S. skies clear out in a 34 second timelapse in God's Eye View. Pulls together public ADS-B, geostationary satellite imagery, FAA advisories and ATC audio. - Blue = planes in the air - Orange = planes on the ground - Rings = aircraft piling up at each airport, with the count beside it What we know based on advisories and tonight's White House briefing: - This is the longest nationwide ground stop since 9/11. - Started at Chicago O'Hare at ~3:10 PM ET with "comms issues" - Soon after radar & radio instability spread to air traffic control centers nationwide - Wasn't a complete blackout. Sec. Duffy says old analog switches and routers got overloaded, leaving controllers with intermittent radio and radar, so the FAA stopped departures. - DHS Sec. Mullin says patches installed during an equipment upgrade are the primary suspect. Essentially "the new technology… was overwhelming the older system." - No evidence yet of a cyberattack, though DHS hasn't ruled out either yet. Notably, Sec. Duffy says the FAA's new AI routing tool (SMART) didn't cause it. - More than 3.5K planes were in the air, and every departure nationwide was stopped at 3:54 PM ET - The stop was lifted at 8:19 PM ET after about 4.5 hours, with 6.6K+ flights canceled and 8.6K delayed (per Flightradar24). This is a wild 4D reconstruction to see. Much respect to the controllers and pilots who kept everyone safe while working with degraded capabilities.
26
94
11
423
55,590
Sameh Khamis retweeted
MiniMax-H3-Image in 2 steps, realtime pDMD LoRA on MiniMax's 33B image model realtime editing on ZeroGPU LoRA + one-file ComfyUI checkpoint: huggingface.co/akhaliq/minim… app: huggingface.co/spaces/akhali…
30
56
10
482
58,724
Sameh Khamis retweeted
This is clearly the path to scaling ground level 3d globally. It's wild how good LLMs are getting at making plausible 3d reconstructions of the real world. This literally used to be an IARPA (IC world's DARPA) project a few years ago and now it's a frontier model prompt.
Gemini 4 Carbon is out today - spent a bit of time testing its long-horizon agentic coding on a personal experiment: rebuilding ~220m of downtown Los Gatos in Blender (~50 storefronts + 3D interiors), w/ ~zero manual modeling. Quick look under the hood 🧵 (1/5)
20
38
2
400
25,211
I wanted to test Fable 5.2's quality with one-shot prompts without human feedback. Everything in this game is from scratch: models, sounds, textures. ~$550 at API pricing, ~2.7M output and 445M input tokens, heavily cached. Took 8 hours! Play it here: claude.ai/artifact/BGY1SztrA…
863
1,168
851
15,037
4,483,289
Sameh Khamis retweeted
Introducing pi-H3⚡, the fastest MiniMax H3 yet. 10 s of 768p video with audio generated in 9.2 s on a single B200. How we did it: Direct pixel output in 2 or 4 steps, no video VAE decoder. Blog: hanshengchen.com/blog/pi-h3/ App: huggingface.co/spaces/Lakoni… @StanfordAILab @LumaLabsAI
38
49
21
660
54,893
And the huggingface demo for Solv3r is out too! Jeong-gi 🚀🚀
We just released Solv3R on 🤗 HuggingFace Space! 🔗 huggingface.co/spaces/jgkwak… Try our interactive demo with your own images and see how Solv3R reconstructs your scene in 3D! 🚀
2
23
1,666
Sameh Khamis retweeted
Seventeen people who helped build past AI 100 winners now list roles at stealth startups. We read their work histories, papers and social posts to guess what they're building. Guess where these people — having built some of the most promising AI startups of the last decade — are placing their next bets. cbinsights.com/research/seve…
1
1
3
1,856