Vansh Bhavsar (@vanshdot) header
Vansh Bhavsar (@vanshdot) profile picture

Researcher

India
Joined February 2023
Procedural generation is the most underrated compression scheme in computing, and a 96 KB HTML file just made the case. Abyssal Garden is an underwater world you swim through in the browser: sunbeams, manta rays, a purple grotto. There are no images, 3D models or sound files in it, and no game engine either. It talks to WebGPU directly. On an RTX 5070 Ti it runs at ~180 fps at 2560×1440. A budget Galaxy A16 phone still gets ~20 fps on Low. ~1.9k stars , MIT licensed. github.com/Noniv/abyssal-gar… #WebGPU #ProceduralGeneration #GraphicsProgramming
1
2
157
Until now, every near-quantum-limited amplifier working at millikelvin temperatures was superconducting. That just changed. OIST built a maser around a pink diamond: microwave-pumped nitrogen spins boost faint qubit signals more than 1,000x, adding only ~2.55 quanta of noise above the quantum limit. It also works inside strong magnetic fields, which spin-qubit chips need and which trouble superconducting amplifiers. #QuantumComputing #Physics #Diamond
1
100
Human reviewers pull 8.4% of their criticisms from a paper's appendix. Frontier LLM reviewers: 62%. A new study had GPT-6 and Claude 5.5 models review ICLR 2027 submissions, then mapped each weakness to the section it came from. The AI critiques keep going past page 30 of the appendix. They skew toward detail audits: more seeds, significance tests, confidence intervals. Rarely the core idea. Same pattern via API, web chat or coding agent, which points at the models, not the interface. arxiv.org/abs/2611.17811 #PeerReview #LLMs #AcademicTwitter
7
32
13
327
27,288
If you build agents on frontier APIs, the number to watch is shifting from price per token to cost per finished task. Alibaba's new Qwen4-Max claims ~60% lower average cost per task than Qwen3.8-Max on Work360, its in-house benchmark spanning 721 professional fields, while using 35% fewer tokens at High effort. Under the hood: a 2.4T-parameter MoE that activates ~71B per token, with a 1M-token context and native vision. Caveat: these are self-reported, and this flagship ships closed-weight. Smaller open Qwen4 models are promised "soon". #Qwen #LLM #AIAgents
1
31
Last SOTA: 128 GPUs for 7 days. SuperMimic: 1 GPU for 3 hours. Same job: one humanoid policy that tracks 129,514 human motions. SuperMimic (UC Berkeley) beats the SONIC baseline's return with >1000x fewer GPU-hours. The trick: learn one differentiable world model, then train the policy by backpropagating through short imagined rollouts. Failed motions (SONIC → SuperMimic): 179 → 25. With 2 GPUs for 9.8h: 0. The same policy runs on a real Unitree G1. Apache-2.0, ~1.6k stars in 6 days. github.com/HybridRobotics/su… #Robotics #ReinforcementLearning #OpenSource #GitHub
1
12
630
Vansh Bhavsar retweeted
born too late for apollo. born too early for multi-planetary travel. born just in time to witness humanity try.
23
264
14
4,023
62,029
The best AI code reviewer isn't the smartest agent. It's the one with guardrails. Alibaba open-sourced the reviewer it ran internally for 2 years. Plain code picks the files, bundles them, matches rules and pins line numbers. The LLM agent only makes the judgment calls. Same model (Claude Opus 4.6), their own 200-PR benchmark: 33.9% precision vs 7.2% in Claude Code. 385K vs 5.66M tokens. ~1.5 min vs 13 min per review. The trade-off: recall drops (20% vs 28.9%). Apache-2.0, ~51.6k stars. github.com/alibaba/open-code… #AI #OpenSource #GitHub #LLM
1
3
114
Vansh Bhavsar retweeted
Congratulations to an incredible team at SpaceX!!!
7,064
17,751
2,149
169,857
7,648,105
Vansh Bhavsar retweeted
We did it
1,816
7,817
1,245
77,146
6,872,155
Vansh Bhavsar retweeted
Mechazilla catches Starship! Congratulations to the entire SpaceX team on this incredible achievement!
6,800
27,301
9,085
140,924
17,096,830
Video models can make a scene look real. Making it move like the real world is another story. Old way: prompt the video model, hope the physics holds, retry. New way (PhysLoop): an LLM agent builds the scene, runs a physics simulation in Unreal Engine, verifies it, and only then lets a video model (Seedance 2.5) restyle the checked motion. Repairing physically wrong videos: 19% → 86% success. New videos: 73% → 85%. The catch: visual realism drops (human rating 3.72 vs 4.29). Physics first, polish second. They're releasing 100K physics-checked videos for training. arxiv.org/abs/2611.16748 #AI #VideoGeneration #AIResearch
1
45
What if an LLM didn't have to write an answer at all, just pick one? That's Google's new Gemini Decision 1. Instead of generating text token by token, it scores your candidate options in a single forward pass and returns probabilities plus a 0–1 confidence score. Yes/no, multiple choice (2–255 options) or rubric scores. Up to 1,000 questions over one multimodal input: text, images, audio, video, PDFs. ~150–250 ms on shorter inputs. $0.10 per 1M input tokens, $0 for output. The calibrated confidence is the real feature: unsure cases go to a human. #AI #LLM #Gemini
1
64
Vansh Bhavsar retweeted
No boundaries.
117
1,002
48
13,534
346,925
A Starship upper stage is circling Earth right now, and SpaceX plans to bring it home into the arms of a tower. Starship Flight 15 lifted off Friday from Starbase. Ship 42 reached orbit (only the second time Starship has) and deployed 26 Starlink V3 satellites. About 22.5 hours after liftoff comes the deorbit burn, re-entry, and an attempted catch by the tower's chopstick arms. No ship coming back from orbit has ever been caught. The booster half already missed: its tower catch was called off and it came down in the Gulf. If this works, the expensive top half of the rocket becomes reusable too. #SpaceX #Starship #Space
1
93
The weakest part of AI coding agents isn't bugs. It's that they write yesterday's web, because their training data is full of legacy code. Google Chrome's Modern Web Guidance, stewarded with Edge, Firefox and MDN, is a skill that feeds agents curated guides on modern browser APIs, fallbacks and accessibility. Search runs offline, no API keys. Its open evals: 1,104 tasks across Claude Code, Codex CLI and Antigravity CLI. Best-practice pass rate 52% → 92%. Automated a11y checks 56% → 94%. ~5k stars. github.com/GoogleChrome/mode… #OpenSource #GitHub #WebDev #AI
13
76
652
Vansh Bhavsar retweeted
Live Starship video from space
4,777
7,886
763
72,399
6,976,064
Tell GPT-5.4 a normal brain scan has a lesion, and it often agrees. Normal scans, neutral prompt: GPT-5.4 is right 87.7% of the time. Prompt assumes a lesion and its location: 51.4%. Neuro-VL on the same test: 97.3%. Neuro-VL is an 8B vision-language model trained on ~1M brain MRI/CT studies from 450k patients, evaluated on 24 public datasets from 21 countries. In radiology, a model that won't take the bait matters more than one that sounds smart. arxiv.org/abs/2611.15094 #MedicalAI #Radiology #AIResearch
2
6
192
Why does the same DNA do different things in a T cell and a liver cell? Context. Google DeepMind's AlphaGenome 2 now models exactly that: give it 1M letters of DNA plus a target cell type or state, and it predicts what that sequence does there. Trained on 100M+ single-cell profiles across 68 tissues. On TenK10K (28 immune cell types) it finds the causal variant 30% of the time vs 16% for the best prior tools, and gets effect direction right 43% vs 23%. The part I like: it adapts to new single-cell data without retraining. #AI #Genomics #ComputationalBiology #DeepMind
2
130
k-means is one of the most underrated bottlenecks in AI. Vector search and dataset curation lean on it, and most of us still run Lloyd's algorithm from 1957. e-means, from the FAISS authors, swaps full-batch passes for mini-batch updates with an exponential moving average. It beats Lloyd's 25-epoch loss in ~1.2–3 epochs. 10–45x faster on GPU, 8–15x on CPU. At scale: 10 billion DINO embeddings into 16 million clusters in 2.5 epochs. No tuning needed: batch and step size come from scaling laws. arxiv.org/abs/2611.09501 #MachineLearning #AIResearch #OpenSource
1
54
775
In 2019 we got the first image of a black hole. Then critics argued the famous ring might be an artifact of how the data was processed. The twist: a CSIRO-led team just re-imaged M87* with a diffusion model trained on CIFAR-10. Cats, cars, planes. Zero astronomy. It reads only "closure invariants", signals immune to telescope calibration errors, and starts from no assumed shape. The ring came back in EHT data from 2017, 2018 and 2021: 42.2 ± 1.5 μas wide, about 6.5 billion suns. AI as an independent referee for physics. doi.org/10.1038/s41586-026-1… #BlackHole #AI #Astrophysics
1
2
310