Deb (@debsahu) header
Deb (@debsahu) profile picture

🫟Splatting Physical AI 🧲 NMR/Biophysics PhD 〽️GoBlue · ⚙️ Freshnox · 💻 S3 Tech Innovations

Ann Arbor, MI
Joined May 2009
Photo by Deb (@debsahu)Photo by Deb (@debsahu)Photo by Deb (@debsahu)Photo by Deb (@debsahu)Photo by Deb (@debsahu)Photo by Deb (@debsahu)Photo by Deb (@debsahu)Photo by Deb (@debsahu)Photo by Deb (@debsahu)Photo by Deb (@debsahu)
Pinned Tweet
I’m back on Xitter! The research community reach here is far bigger than BlueSky. Now that I’m (@S3TechInno) venturing into Physical AI along with EarthByte Lab (earthbytelab.com), I’ll post progress & my contributions to various OSS can be found in GitHub.
1
1
546
FERRIC v0.18.0 introduces a unified joint 1H + 13C-detected IDP backbone solver in Rust (--regime idp-joint). Fuses HN + 2D CON across Prolines & broadened loops, pools CAS typing & N-N edges, auto-calibrates TROSY offsets & abstains on shift-predictor bias.
21
For the tech people who don’t have any friends in bio: Feng Zhang’s lab literally discovered a new gene editing mechanism every 6 months with like 20 people and 1/1000th of the funding. Using thousands of overpaid scientists for this weak of a headline is actually lame af.
Dario Amodei reveals Anthropic discovered what looks like a new gene-editing mechanism and says it may discover dozens a month within months “I’ve been talking to our biology team several times a week. We discovered what looks like a new gene editing mechanism, and I think within a few months we may be discovering dozens of them every month. We’ve seen, with the speed at which this technology moves, these little sparks will soon turn into a firestorm.” “I think that we can find a way to build this technology safely, in a way that benefits people, that addresses all the risks, and that delivers a world that is so much better than the world we have today. We haven’t done that yet. We haven’t earned the right to talk about how great those things are, but we’re starting down the path.”
25
126
5
1,918
81,887
In silico ≠ in vivo, we are not there yet!
Today, we enter a strategic research collaboration with @MSDInvents. The collaboration brings together Isomorphic Labs’ world-leading frontier AI capabilities with MSD’s deep discovery, development and manufacturing expertise. Read more: isomorphiclabs.com/articles/…
23
Built FERRIC (open-source #NMR workbench w/ MCP) in @antigravity using Gemini Neon Extra High & threw a brutal Pro/Ser-rich IDP at it. Manual solve got 16/159 residues (10%) & Opus stopped at 24/159 (15%). Neon Extra High hit 87/159 (55%) blind Tier-A (47+ cross-validated).
1
1
138
Install Ferric > Turn on MCP > Copy MCP config > Paste config into your favorite AI and ask it to assign your NMR signals! Again you can manually do every step on your browser (see demo link in repo) without installing anything.
13
Apple just released Depth Pro 2 (github.com/apple-aiml-resear…). We ran it on indoor frames from our in-house capture rig, next to MoGe-2. First impression: very clean depth.
7
6
181
8,674
Where it shines: fine detail. Chandelier arms, a railing, hanging shoes. The normals from Depth Pro 2 are crisper. One catch: it reads mirrors as a window into the reflected room.
2
7
903
Scored against LiDAR on 9 views, MoGe-2 still edges it: 0.063 vs 0.086 error, 95.3% vs 94.9% within 25%, 0.78 s vs 0.99 s per image. Small test, Depth Pro 2 L model. Sharper is not yet the same as more accurate.
1
7
717
Let’s bring this to spirula studio?
🧊  BENCHMARKED IN 3D: @Apple's Depth Pro 2 open weights on my 64GB M5 Pro Mac mini, across Core ML, PyTorch and MLX Verdict: Apple’s model running natively (& blazingly fast!) on Apple hardware using Apple’s ML frameworks 🤯 Core ML (single frame inference): - I outperformed Apple’s published numbers (on an M4 Max MacBook Pro) for all 20 model size x input resolution combinations! - For the small model, I got 149 fps when processing 360p video (and 31 fps for 720p video), making it faster than real time! - For the biggest model (H+), I got 495 ms per frame when processing 720p video (Apple got 826 ms per frame on their M4 Max machine). Other fun numbers: - It took 1.94 s to turn a 1920×1440 photo into a depth map with model H+ - Depth maps generated with my MLX port correlate extremely well (r = 0.9999+) with Apple’s depth maps (generated with PyTorch) Your Mac can see depth now. Literally!
2
262
Ok just saw the license, it’s not going to happen. Also it’s not any better or faster than MoGe-2!
1
19
Bring this to 3DGS?
We shrank 1 GB of Wikipedia to 105.39 MB, losslessly. Introducing tufazip: the best entry on the Large Text Compression Benchmark using a neural network as its sole predictor. Everything above it mixes a transformer with hand-built models. Full post 👇 tufalabs.ai/research/tufazip
25
OK I’m going to say it, @antigravity neon medium is trash 🚮 makes tons of mistakes - but neon high with a hydrogen reviewer is 🚀 && neon high feels closer to opus 5.5 medium.
39
AVM looks 🔥
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
139
I have stopped reviewing scientific papers in last 2 years, for this exact reality!
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...
50
Wow a vaccine for leukaemia! Incredible!
A first of its kind vaccine for acute myeloid leukemia, a deadly blood cancer, succeeded in a Phase 3 clinical trial, drugmaker Sellas Life Sciences and Memorial Sloan Kettering Cancer Center announced. wapo.st/4t6oIRy
138
1,534
39
10,295
393,874
Hey @antigravity , dealing with “Our servers are experiencing high traffic right now, please try again in a minute.” all morning! Any fixes coming soon?
1
40
Maybe this is the reason ? I’m down to 56% weekly usage - no actual work done. Keeps trying to connect to servers, errors and loops all over.
🚨 Gemini 4 Neon's benchmark is floating around. It's beating Google's best model, Carbon, by a fair margin and leads on 7 out of 8 parameters it was tested on. These are unverified and rumoured benchmarks, so take them with a pinch of salt.
1
29
Heard of @docusealco ?
DocuSign wants $8,100/yr from a 15-person team just to sign PDFs. So I built SignuDoc: open source, one-click deploy to your own @vercel account, $0 per seat. White-label branding, full audit trails, tamper-evident seals and your documents never leave your cloud.
46