Building Foundation Models • Founder at @soketlabs • @iiscbangalore • HPC ♥️ AI
Bangalore
Joined July 2009
- Tweets1.4K
- Following636
- Followers1.7K
- Likes2.4K
Abhishek Upperwal retweeted
Does an instruction actually decide its output?
@prit_sk explains our @emnlpmeeting paper on the Task–Specificity Score: with just 5% of the data, TSS/TSS++ win 8 of 9 settings.
@upperwal @lingoiitgn @iitgn
arxiv.org/abs/2602.03103
#EMNLP2026
Abhishek Upperwal retweeted
We're hiring a Cybersecurity Specialist, Agentic Cybersecurity, in Bengaluru.
Help build Soket's cybersecurity frontier model with MITRE ATT&CK mapping, exploit and CTF datasets, agentic red and blue workflows, and evals.
You'll need 3 to 5 years in red teaming, pentesting or AppSec.
Apply: in.northstarz.ai/job/candida…
#Hiring #Cybersecurity
Abhishek Upperwal retweeted
Our Partnerships Lead @KochharSoumya represented us at #AadhaarSamvaad2026 and spoke on AI and the future of India’s digital identity infrastructure.
Organised by @UIDAI on the 17th Aadhaar Day under “Trusted Identity, Expanding Possibilities.”
Five initiatives were announced — including face authentication for ePDS and a multilingual UIDAI website.
AI is already in identity. The next step is sovereign foundation models that strengthen the intelligence layer around it — while keeping security, privacy, and human accountability at the centre.
AI should automate the work, not the accountability.
#AadhaarSamvaad2026 #UIDAI #DigitalIdentity #SoketAI
Why we built it: agents are moving from minutes to hours on critical workloads. At that point the harness is infrastructure, and it needs to be fast, small and boring.
Loop is built for exactly that.
Try it and tell us what breaks. Do give it a star if you like it.
GitHub: github.com/soketlabs/loop
Install is where it gets silly.
Loop: 9s, one 21 MB file.
Pi: 501 MB and 25,693 files.
Claude Code: 42s.
opencode: 77s.
Native Rust core: 14 MB resident, 8 ms turn latency.
It does the same work with far less.
Loop used 803k tokens against OpenCode’s 1.82M and Goose’s 1.33M. That works out to $0.13 vs $0.29 for the full suite, and it finished in 12 minutes vs 23.
We benchmarked 4 harnesses on identical coding tasks (website, JS, Rust), all running GLM 5.3 Flash and graded them.
Loop: 13/13 passed
Fastest median task time: 29s (OpenCode: 72s)
Lowest CPU time: 1.7s (OpenCode: 8.5s)
We just open-sourced Loop: a Rust agent harness that’s 44x lighter than opencode at idle, 5x lighter on CPU, and finishes the same tasks in half the time.
Initial benchmarks look promising.
Same model. Same tasks. 13/13 passed.
Developer preview, out now. 🧵
> curl -fsSL loop.soket.ai/install | bash
github.com/soketlabs/loop
Add cross entropy to that list
Do you really need QKNorm to stablize your training?
Does your "failed training" really failed by bad design?
The answer may lie in your kernel!
We found cudnn, FA4, flex attn ALL have severe precision issue in backward!
And we bring a better kernel here:
github.com/KohakuBlueleaf/Ko…
So much useful data especially for Indic languages can come out of such initiatives
Kudos to you and the team @kingofknowwhere
Release 🚨
America dot gov is boring. I am happy to release 275k public Indian gov documents parsed from 15k gov websites. 322k official PDFs and on 128 topics, available to plug and play into any RAG frameowrk of your choice. Both raw data and cleaned data available in csv, parquet and SQLite format. At least 50k pages are useful. :D
Will be updating this to 20 Million documents by month end (process running on JarvisLabs autonomously. Happy to share tracking link for those interested)
Go build your Indian gov chatbot. No license, no credits needed. <3
Rust is going to places. C still remains my fav
If you are a pro in Rush, would love to talk. We have been building our harness from scratch in Rust and the performance is phenomenal
Introducing CUDA Rust: Two Tracks for Writing GPU Kernels
> NVIDIA launches CUDA Rust with two tracks—cuda-oxide (SIMT, nightly) and cutile-rs (Tile, stable)—for writing safe, native GPU kernels directly in Rust.
ALT Image: Introducing CUDA Rust: Two Tracks for Writing GPU Kernels > NVIDIA launches CUDA Rust with two tracks—cuda-oxide (SIMT, nightly) and cutile-rs (Tile, stable)—for writing safe, native GPU kernels directly in Rust.
Abhishek Upperwal retweeted
Our first OCR model - live soon on HuggingFace and Bodhan
Text (Indic-Translate). Speech (Indic-Transcribe). Now vision.
IndicOCR, the new kid on the block. Built by @ai4bharat and @Bodhan_AI
Handwritten notes, degraded archives, 9+ scripts, all in a 0.8B model that holds its own against models 4-5x its size.
Live 05.09.2026.
Short video here 👇 full version in the reply.
Horrible experience with @wandb. Have been trying to upgrade for weeks now and all 9 cards are declined. Seems like they dont support India for billing. Lost access to the account and the support takes 3 days to response even with priority support. Moving back to tensorboard.
If an Indian entity is working on Wandb alternative would love to get a subscription. Also, India lets please break this US dominance on all essential softwares.
