Swayam Singh (@swayaminsync) header
Swayam Singh (@swayaminsync) profile picture

Maintaining OSS | formerly @MSFTResearch Writing at https://nitter.gq/t.co/7LCbjNMXe5

BLR
Joined April 2021
Photo by Swayam Singh (@swayaminsync)Photo by Swayam Singh (@swayaminsync)Photo by Swayam Singh (@swayaminsync)Photo by Swayam Singh (@swayaminsync)Photo by Swayam Singh (@swayaminsync)Photo by Swayam Singh (@swayaminsync)Photo by Swayam Singh (@swayaminsync)Photo by Swayam Singh (@swayaminsync)Photo by Swayam Singh (@swayaminsync)Photo by Swayam Singh (@swayaminsync)
Strong version of you is dealing with all the inner demons silently, keeping all the chaos contained within you, hidden from the outside world. It'll get exhaustive sometimes and I am proud of you. Don't give up.
20
20,499
Not the expert but this is the thing I was talking about when I said we need such findings/claims to be shown more on this platform.
Is anyone verifying the lean proofs that the formal statement actually matches the objective without hidden axioms, compiler quirks, or unintended loopholes? If yes then their statements deserve to be promoted even more on this platform.
63
46
22
855
182,377
sharing this as an example of a "claimed" loophole in a formalization, not as an endorsement of the author’s other work.
1
42
17,281
Is anyone verifying the lean proofs that the formal statement actually matches the objective without hidden axioms, compiler quirks, or unintended loopholes? If yes then their statements deserve to be promoted even more on this platform.
7
2
97
143,402
Clocked 100+ Km on motorcycle, to spend some time in peace before I left my hometown.
13
1,236
Opus: writes a cleanup test → realizes the cleanup scope might be too broad → runs it to verify → deletes all the work and my entire home directory. Guardrails in auto mode? What guardrails?
1
2
188
Oh my my I'm so bullish on this. I was working on the similar idea but not for inference but a special type of efficient and privatized model training. Such a cool demo and really awesome work guys. (Really hoping if I get sometime later in this month to resume back, don't want it to be an inefficient AI slop)
introducing "pied piper" (built at @ETHGlobal ). ever thought - what if you could run a BILLION parameter model on your mobile? we did it :p guess what, you dont have to run the complete model, just few layers of it. here's what we built : we built decentralised LLM inference and ran 2.5B params models on 6 phones by just utilizing 300mb RAM per device. @sama said common folks don't utlize compute to it's full potential we're changing this now. We write WASM kernels so your phone can run a billion parameters model and we expose it as a openai api compatible endpoint so you can connect your agent to it AND a @Uniswap AMM pool for a decentralised inference market so everyone can earn dollars on basis of compute they provide. worked on this w my cracked teammates (@Siddhartha37648 @HarshitNay80531 ), we are planning to opensource/release this, DMs are open if you'd like to talk more ;)
1
11
677
Can you lateral raise this much? 18kg?
If you're into robotics, check out this repo and thank us later github.com/menloresearch/asi…
1
7
663
"If you get all your information Twitter, you dont always get the most accurate picture of the world"
8
369
"Pun intended" Folks discovering mech-interp in 2026
2
16
590
Swayam Singh retweeted
Announcing NeoSLAM: our in-house SLAM pipeline, developed from scratch. NeoSLAM is #1 in the Long category of the LaMAria leaderboard among mono+IMU methods, and has the highest pose recall (within 5 m) of any mono+IMU method in every category that reports it.
4
5
3
43
5,691
Reviewed some proposals and really strong and interesting talks are lined up. If you have something interesting to present from AI/ML, HPC, Numerical computation stack, Quantum computing, environmental, health or from any spectrum of Science, you are most welcome. Feel free to DM me for any queries.
4 days left! SciPy India 2026 is our 14th edition since 2009, and our first time in Chennai: IIT Madras, 19th–20th December. Submit your workshop, talk, or poster by 1st November, 23:59 IST. cfp.scipy.in/scipy-india-202…
1
1
7
469
Swayam Singh retweeted
4 days left! SciPy India 2026 is our 14th edition since 2009, and our first time in Chennai: IIT Madras, 19th–20th December. Submit your workshop, talk, or poster by 1st November, 23:59 IST. cfp.scipy.in/scipy-india-202…
1
1
4
584
Swayam Singh retweeted
5 days left to send your proposal to SciPy India 2026. We review proposals on a rolling basis, so the earlier you submit, the better. Closes 1st November, 23:59 IST. cfp.scipy.in/scipy-india-202…
1
3
169
Swayam Singh retweeted
6 days to go! Three ways to take part in SciPy India 2026: Workshops on 19th December Talks on 20th December Posters on your project or research Proposals close 1st November, 23:59 IST. cfp.scipy.in/scipy-india-202…
2
2
106
I wonder how we can scale this for compilers? Since it'll takes a program and spits out another program. Here the theorem would become “for every C++ program, whatever machine code Clang produces behaves like my rulebook says.” That'd be proving Clang itself correct?
The answer, as always, can be found in category theory: Ship your interpreter (or codebase). The LLM can produce whatever slop it wants, but you can validate it by proving a refinement to the binary of your existing codebase. Slop in. Trust out. basis.ai/blog/ship-your-inte…
2
1
20
1,559
One of the coolest startings any book can have!
11
438
My LinkedIn and X feeds are full of talk about the recent C++26 ballot failure, mostly around the contracts feature. Most of the value people describe is better runtime checking during development. I keep wondering if the more interesting use of contract-like annotations is formal verification rather than just optional asserts. Tools like cpp-verify are exploring that direction:github.com/SwayamInSync/cpp-…
1
14
1,076
Swayam Singh retweeted
7 days left in the SciPy India 2026 call for proposals. Researchers, educators, RSEs, students, data scientists, HPC folks: if you write code to do science, we'd love to hear from you. Closes 1st November, 23:59 IST. cfp.scipy.in/scipy-india-202…
1
3
112