Arvind (@nagaraj_arvind) header
Arvind (@nagaraj_arvind) profile picture

Borderline boomer

Bengaluru
Joined November 2016
Photo by Arvind (@nagaraj_arvind)Photo by Arvind (@nagaraj_arvind)Photo by Arvind (@nagaraj_arvind)Photo by Arvind (@nagaraj_arvind)Photo by Arvind (@nagaraj_arvind)Photo by Arvind (@nagaraj_arvind)Photo by Arvind (@nagaraj_arvind)Photo by Arvind (@nagaraj_arvind)Photo by Arvind (@nagaraj_arvind)Photo by Arvind (@nagaraj_arvind)
Here's a lecture I gave on RLHF and the new DPO paper which proposes a wonderful alternative to OpenAI's PPO algorithm. If you are interested in improving the quality of your LLM completions using RLHF, please watch this video. youtu.be/Ju-pFJNfOfY
5
45
3
254
46,881
I’m trying to move on 🤣🤣🤣🤣
155
1,649
143
15,158
752,302
Done with my first ride on the new #PinkLine in Namma metro network. Kalena Agrahara 6:00 Hulimavu 6:01 Bilekahalli–IIMB 6:03 NS Palya – J.P. Nagar 4th Phase 6:05 Jayadeva Hospital 6:06 S.G. Palya (Suddagunte Palya) 6:08 It took close to 9 min for 7.5km journey, roughly one-third of the time it takes to travel by road on a typical working day.
44
109
9
1,713
81,002
Wow that escalated quickly. Was gone for a few weeks and now pretty much everyone launched their own decision model copycats: - Sep 23: Together AI, Tev1-4B - Sep 24: Alibaba Cloud, decision-model - Sep 29: Liquid AI, d1 - Sep 30: Databricks, ai_decide - Oct 1: Amazon AWS, Strands Decider 2B - Oct 1: Cloudflare, Clef and Clef-flash (Clef-omni on Oct 9) - Oct 1: Perplexity, Decider (v1.3 on Nov 12) - Oct 6: OpenAI, Decisions API (GPT-6 Luna) - Oct 7: H2O.ai, H2O-Lightning-4B - Oct 7: Snowflake, Snowflake Decision - Oct 9: Microsoft, Microsoft-Decision-1 - Oct 15: IBM, Granite Switch 4.1 with Decision adapter - Oct 21: Baseten, Mammoth-1 and Elephant-1 - Nov 6: Salesforce, Koa Decisions - Nov 11: Google DeepMind, Gemini Decision 1 According to a Vercel report, 26% of requests are now routed through decision models 📈
Took a while, but here it is, a mega write-up on language models for text classification and, of course, Jev. It's basically a visual guide to RNNs, CNNs, transformers, and calibration, with hands-on experiments on accuracy and efficiency.
57
82
6
776
42,742
Arvind retweeted
I can't believe nobody has heard of this Bengaluru-based nanotech lab. In a year or two I'm sure they'll be very well-known. They're building chips that are 100x more energy-efficient than modern GPUs. This unlocks applications like SOTA edge inference on phones and laptops. They're planning to commercialise this tech as MBrain AI, a company co-founded by Prof. Goswami with IISc colleagues and semiconductor industry veterans. Currently they're incubated at IISc's CeNSE (Centre for Nano Science and Engineering), as the MoNC-AI research group.
A lab in Bengaluru is building brain-inspired chips from molecules, designed to be nearly 100x more energy-efficient than today's GPUs. Led by Prof. Sreetosh Goswami this team of ~30 chemists, physicists and engineers is based out of the Centre for Nano Science and Engineering (CeNSE), IISc. Here's the problem they're solving. The computers we use today (GPUs included) were never designed with AI in mind. They keep data in one place and do the calculations in another, so every step means sending data back and forth between the two. Engineers call this the von Neumann bottleneck. Everyday software barely notices it, but AI does, and that's because a neural network multiplies millions of numbers at once. Even the most advanced AI hardware spends a remarkable share of its energy simply moving data around. Now imagine a chip where memory and computing happen in the very same place. Not just side by side on one chip, as in modern phones and laptops, but with the calculation taking place inside the memory itself. That is how the human brain works, and it's also what this team is building. The MoNC-AI research group (Molecular Neuromorphic Computing for Artificial Intelligence) designs tiny molecular devices called memristors. A silicon transistor stores a simple 0 or 1. Each of these molecular devices can hold more than 16,000 distinct levels. That richness enables them to carry out the central operation of AI in a single step, instead of thousands. Because these chips need only a fraction of the energy, they could bring powerful AI to phones and laptops, without relying on the vast power and cooling of data centres. The technology born in this lab is now being taken to market by MBrain AI, a company co-founded by Prof. Goswami with IISc colleagues and semiconductor industry veterans. Alongside the chips, MBrain AI is building the software that lets everyday applications run on them, and is working towards its first fully packaged chips for customers over the next couple of years.
27
157
7
976
15,596
There is no substitute for the person who Knows What To Do.
1,318
4,863
482
41,419
2,998,657
I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo…. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
1,820
6,415
2,194
41,184
6,669,022
Thank you @JensenHuang for 33 years of bold leadership. You built a small startup into America’s most valuable company, with relentless improvement. Your grit and perseverance should always be remembered. Your personal generosity with time and advice to founders like me has secured a bright future for America and AI.
286
1,266
72
22,108
611,221
I updated my quantization compiler to use e-means as a replacement for k-means in codebook generation and got Qwen3.5-4B from 8.41 GB to 1.23 GB while retaining 91.9% of original performance which I think is approximately SOTA, at least comparable to bonsai 2 but I think we could go lower if someone wanted to donate some GPU time since this is just a hobby for me. at this size, it's small enough that memory bandwidth isn't the bottleneck to decode speed so I'm interested in seeing how fast we can get larger models on consumer hardware. link below.
Introducing e-means, an algorithm beating k-means in 10x fewer epochs. With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]🧵
25
62
12
879
28,694
We’re making a personalized cancer vaccine for Yash, a 34 year-old pancreatic cancer patient, in the open with his consent. We want external experts to review our design and make their own contributions before we manufacture it at the RNA Foundry. RNAFoundry.com/yash
115
463
85
2,562
258,733
Introducing e-means, an algorithm beating k-means in 10x fewer epochs. With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]🧵
150
1,224
257
10,223
894,016
Day 22 of building GTA style Bengaluru game, 1. Now supports voice in 5 languages. 2. Source code audit by Fable 5.2. 3. Temple & Cultural rule sets, like no weapons or footwear inside sacred premises, NPCs of different faiths tagged appropriately 4. Cellphone system etc
Day 20 of building GTA style Bengaluru game before GTA 6 in @UnrealEngine Flashback mission - 2018 - Majestic! Weather systems, rain, mission systems, cinematics, cut scenes, crowd control, crowd interaction improvements, new map area: Ulsoor lake, Majestic, Racecourse, Anand Rao circle and surroundings, voice in cinematics and crowd, chirping, improved background sounds, working bus stations and lot more! Will proly iron out the NPC anim glitches next.
95
98
17
1,899
125,494
Thrilled to announce that our @GoogleDeepMind team has cracked a major challenge in software engineering and cybersecurity: formally verifying a real-world implementation of the #Signal cryptographic protocol at the machine-code level. This proof was done using AlphaProof Code, our new AI framework for formal specification and verification. * Paper: deepmind.google/documents/36… * Lean code: github.com/google-deepmind/c… * DeepMind blog post: deepmind.google/blog/introdu… A huge shoutout to my amazing teammates: @tchajed, @FerrousAndrew, Jonathan Katz, @pushmeet, @antonkovsharov, Karn Seth, Sam Schlesinger, @RandomlyWalking. I started working on formal software verification 20+ years ago as a grad student. I was persuaded by Dijkstra and other computer science pioneers’ vision of programming backed up by mathematical theorems. At that time, we had to choose between automated code analysis algorithms that scaled but could only prove shallow properties, and deep manual verification that took decades of expert human work. Thanks to progress in AI for mathematical reasoning, we don’t have to make that choice any more. Over four weeks in October 2026, AlphaProof Code agents generated more than 6 million lines of machine-checked Lean proof. We proved deep computational and symbolic security properties of the entire protocol core of Signal, all the way down to the ARM machine-code level. This is just the first step, of course, because our goal is nothing short of verifying the world’s software infrastructure. Most cyberattacks, including the AI-powered ones recently in the news, happen because of preventable code bugs. I expect that agents for formal code analysis will eliminate these bugs one category at a time. As our paper explains in more detail, this doesn’t at all mean humans are out of the picture! Proofs are now largely automated, but human experts will need to decide what is proved with AI assistance. AlphaProof Code offers an early glimpse of what such human-AI collaboration could look like, but many other proposals of this nature will emerge. We have released all our Lean specs and proofs and plan to work with the open-source community on improving them. If you are building critical software and would like to apply these ideas, come talk to us!
50
154
29
1,075
60,969
Arvind retweeted
Been running Opus 5.5 agents over the notebooks of Gauss, Riemann, Hilbert and Grothendieck. around 12,000 pages so far. Heres the best finding so far: Riemann's 1860 shock wave paper is famous for getting the energy part wrong however his unpublished sheets at Göttingen have it right the Rankine Hugoniot conditions a decade or two before Rankine and Hugoniot plus results usually dated 19061937, all in his own notation. A second sheet in another folder has the same physics. There are probably thousands of potential undiscovered stuff in the research papers of the smartest ever to live imo all of them should be made public and reviewed by ai years of work can probably be transcribed by agents in days
77
354
65
3,624
268,259
Raghunath ji, thank you for bringing this to our attention. It is at the heart of what we are attempting at @SarvamAI . We agree that you shouldn't have to count characters to be empowered by AI. We have credited your account with a lifetime access to our models. 🙏 Anyone else using Sarvam for accessibility, please write to us at [email protected]. Our team will address you directly while we work towards launching a formal program asap.
Hello @SarvamAI, I am 70, an oral cancer survivor since 2015. Cancer took away my God-given gift of speech. Today, my wife is my voice before doctors, banks and the outside world. I have spent years communicating through written notes, often unread or unanswered. At cancer hospitals, I see countless patients struggling to express their suffering. Our pain is ours, but our voice has become someone else’s. India has an estimated 1.99 million people with speech disabilities. For us, Text-to-Speech AI is not a convenience. It is our lost voice restored. A doctor may not read my 100-word note, but could listen to it in under a minute. Yet your per-character pricing makes regular use expensive. Could @SarvamAI introduce an affordable, unlimited-use Divyangjan monthly subscription like other AI models like #ChatGPT #GeminiAI? We are not asking for charity. We are asking for the dignity of speaking for ourselves. Waiting for your reply so that I can subscribe this tool right away. @narendramodi @AshwiniVaishnaw
159
549
92
2,927
91,451
Arvind retweeted
I have been watching this every morning for 11 months now.
1,006
17,811
1,296
58,379
6,467,581
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
2,365
7,256
5,752
55,936
10,065,120
スマホの傾きに合わせて、絵が回るようにしてみた。 3D化しているわけではなく、1枚の絵から360°回転の「動画」を作り、傾きに合わせてフレームを前後させて、立体っぽく見せています。 #MiniMaxH3 + ORB360 LoRA
91
1,027
148
6,850
369,078
Arvind retweeted
Nitroduck designs new colors, flavors, and disease resistance into fruits and veggies. We integrate CRISPR into breeding through pollen, making any crop gene-editable. Breeding a new specialty crop variety takes 8 - 16 years: breeders hope for random gene combinations and mutations, then spend years picking out the good ones. Meanwhile, consumer tastes, the climate trends, and pathogens change way faster. Now, accessible genome sequencing and AI can tell breeders what to aim for instead of what to hope for, and CRISPR can turn random mutations into chosen ones. The catch is that CRISPR only works on about 15% of crops. It has to go into a single stem cell that’s then regrown into a whole plant, a process that has to be worked out for each variety, which is cost-prohibitive. We skip the stem cell and put CRISPR into a pollen grain. The grain pollinates a flower, and out comes an edited seed - seamlessly tapping into how breeders already work. Matúš, Miro, and I are plant synthetic biologists who made duckweed, the smallest and fastest-growing plant on Earth, genetically engineerable. For doing so we won the Grand Prize at iGEM, the “Olympics of biotech.” During our studies, we engineered 4x more plants than the average plant scientist does in a whole career. Now we’re building Nitroduck, The Crop Design Company, which will make pesticides obsolete by 2042.
72
81
57
737
95,961
Arvind retweeted
fable 5.5 is about to mog everyone asked it to make an animation that makes me feel like it’s agi and in about an hour it cooked this is made with pure code, no external assets used. anthropic really went all with this one watch till the end cuz it gets so good
106
50
9
990
94,470
This is the kind of work we love seeing on HN! Readers are also helping shape it by requesting features, getting working updates back, and sharing their own adaptations. What an amazing project, @karpathy! news.ycombinator.com/item?id…
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5: hackertube.org/ - Excellent recs without any logins (state is kept in local browser session) - Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads - Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version - HackerTube Longs ✊ A bit more on the process: I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing. Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
16
55
4
1,270
60,625