Fast Code AI

Next tuesday I fly to Denver. #CVPR. I have been going for 18 years. I still rehearse the hallway moment in my head. Not the moment with people I already know. Rishabh, my first PhD student - I will be glad to see him. An old advisor - easy. A former collaborator - fine. The hard one is the person I do not know whose talk I just sat through. Last year's CVPR. A researcher from Google had given a talk on embodied VLMs and Gemini robotics. I had been thinking about her result for the last forty minutes of the talk. I had a real question: has her team studied the reasoning traces her models output, do they have the ability to do actual counterfactual reasoning, or is it pattern matching dressed up as causal? The session ends. She walks past me in the hallway. Five feet away. Then six feet. Then ten. I look at my phone. I let her pass. The script in my head: am I worth her time? Is my question obvious? Will I be the awkward person interrupting her between sessions? There must be more important people in this hallway than me. By every legible signal I should not be running this script. I have papers. I have a company. I have my own students. The question is real, technical, and exactly the kind a researcher actually wants in a hallway. The script does not care about the signals. It came from somewhere. A boarding school where I was the low-money kid. The story was told to me in pieces - by other boys, by who got picked up first at term break, by the size of the trunks other parents brought. Then I told it to myself. Now it plays even when no one is in the room. Stuck has two kinds. Stuck in 3D - the closed door, the lost key, the offer you cannot accept - is the easier kind. You can point at the block. 𝐒𝐭𝐮𝐜𝐤 in your head is the harder kind. There is nothing in the room. The block is private. And the voice playing the story is yours. 𝐔𝐧𝐬𝐭𝐮𝐜𝐤 is one second. The second you walk up and ask the question while the script is still playing. Next week in Denver, I am going to try. Whose hallway are you avoiding this week?

"Legal AI research products were just wrappers on OpenAI. They were just not auditable, no reasoning, no hyperlinks, no hierarchy of courts." That's Laina Chan, barrister and CEO of MiAI Law, on the legal-AI category as it stood when we started working together. Pinpoint referencing - which case, which paragraph, why - was the foundational core she needed and no one had shipped. Adityan owns the pinpoint-referencing engine at our end. Anurag the database. Harsh the word plugin. Vaibhavi the retrieval pipeline. The platform now serves barristers across jurisdictions and has handled 1000s of cases where the citation chain has to hold up in court.   The interesting line in the video is Laina's: "Between us, I've solved one bit and you solved the other bit." That's the engagement shape we look for. The customer brings the domain. We bring the system. Neither half works alone. Got an AI problem no one has solved? Reach out: sanjay@fastcode.ai

#CVPR grew from 1,661 accepted papers in 2021 to 2,872 in 2025 - a 73% expansion of the conference itself. We pulled Paper Copilot's country-by-country data to see who actually rose with it. Format below: count (share of total) in 2021 → count (share of total) in 2025. Rising in count: • China: 786 (47.3%) → 1,248 (43.5%) • Singapore: 88 (5.3%) → 165 (5.7%) • South Korea: 96 (5.8%) → 154 (5.4%) • India: 17 (1.0%) → 31 (1.1%) Essentially flat in count, collapsing in share: • United States: 653 (39.3%) → 672 (23.4%) Declining in both: • Australia: 124 (7.5%) → 99 (3.4%) • Switzerland: 77 (4.6%) → 56 (1.9%) • Canada: 73 (4.4%) → 59 (2.1%) When CVPR doubles around you, holding the line means losing ground. The US is the clearest example: same number of papers, dramatically smaller slice. Singapore is the only country whose absolute count AND share both rose - everyone else either grew slower than CVPR or shrank outright. The sharper number is per capita: • Singapore: 27.5 CVPR papers per million people • South Korea: 3.0 per million • United States: 2.0 per million • India: 0.02 per million India has 230× Singapore's population and one-fifth its CVPR paper count. Per capita, Singapore produces ~1,250× more CVPR research than India. This is not a math problem. The math is in the engineers we hire from this pool every week. It is a room problem. A CVPR-accepted PhD student at Tsinghua, NTU, Stanford, or CMU sits next to senior researchers who know how the conference reads, what reviewers reject, what the bar actually is. The room produces the paper. Most Indian CS PhD programs do not have that room. Singapore has spent two decades building them. Our founder Arjun Jain flies to #Denver next week for CVPR 2026 (Jun 3–6). Paper Copilot will publish the enriched 2026 dataset within days of the conference. We will rerun this analysis the day it lands. At Fast Code AI we are building a different kind of room - a production room, not a paper room. If you run a research program at an Indian university and have a student who cannot get over the publishing line, send us a message. The gap is not the math. Sources: https://lnkd.in/gtD7z2SN. Methodology: any-author attribution; population from World Bank 2024.

My first day at boarding school. The Daly College, Indore. 1st July, 1991. I am 7. My father says he is just taking a round outside. He and my aunt never come back. I run after them. I grab the curtains. I am pulled back. The curtains come down with me. I cry myself to sleep on the upper bunk. Sixteen of us in the room. That was the first time. It has happened many more times in thirty-five years. The same one-word thing, dressed up differently. The Yahoo engineer in 2007, resigning with a one-way ticket to Florence and no plan past the landing. The PhD student at MPII in 2008, one exam away from losing my scholarship in Prof. Hein's ML course - PSD matrices, vector calculus, none of it taught at RV College. The new IIT Bombay faculty in 2016, sitting on a bench outside an empty department office while the key the admin had given me did not turn in the lock. The professor at IIT Bombay in 2017, at a whiteboard in CS 763 teaching backprop, when a student said "𝘴𝘪𝘳, 𝘺𝘰𝘶 𝘮𝘪𝘴𝘴𝘦𝘥 𝘢 𝘮𝘪𝘯𝘶𝘴." The founder in 2018, declining a permanent IIT Bombay faculty position at 9pm in my small apartment in Bangalore while my wife put her hand on mine. The CEO in 2022, muting myself on a client video call because the dashboard still said 0.96 on the test set and the new batch from the plant did not. My father dropped out of IIT Kharagpur in 1970. He worked a coconut plantation in the Andamans until my grandparents found him and brought him home to the sawmill in Jiaganj. He sent me to a boarding school away from Jiaganj so I would not have to come home to a sawmill of my own. I have been telling you these stories about institutions. About math. About neural networks. About demos. About money. About boarding schools. They are not. They are the same story, told seven different ways. Each one was a moment when the next move was either invisible to me, or visible and impossible. I have called that a lot of names for thirty-five years. Anxiety. Hypervigilance. Imposter syndrome. The wrong college. The wrong connections. The wrong timing. The wrong town. None of them was true. It was always the same one-word thing. 𝗦𝘁𝘂𝗰𝗸. Sometimes you cannot see the next move. Sometimes you see it and cannot make it. Either way, the lie is the same: the move is not yours. It comes dressed as caution, as humility, as wisdom, as "𝘵𝘩𝘦 𝘵𝘪𝘮𝘪𝘯𝘨 𝘪𝘴𝘯'𝘵 𝘳𝘪𝘨𝘩𝘵 𝘺𝘦𝘵." It is none of those. It is the most expensive thing in a capable person's career. That was mine. What is yours? What is the move that is invisible to you right now - or the move that is clear, and feels impossible? Tell me in the comments. I will read every one. Because 35 years in, I am sure of this: 𝗚𝗼 𝘁𝗮𝗸𝗲 𝘄𝗵𝗮𝘁 𝗶𝘀 𝘆𝗼𝘂𝗿𝘀. 𝗟𝗲𝘁 𝗻𝗼 𝗼𝗻𝗲 𝗵𝗮𝗻𝗱 𝘆𝗼𝘂 𝘁𝗵𝗲 𝗹𝗶𝗲.

This Friday, every screen at Fast Code AI goes dark. On purpose. Rithika G. locked in our second trip to the Cauvery shores within six months. This time to Galibore. Thirty hours. No laptops, no network, no "quick sync." The whole team boards a bus to a place with no cell signal. Job titles disappear. Cricket matches that get far too competitive. Coracle rides on the river. BBQ by the fire. Spooky stories. We build world models for a living. This weekend we are going to live in the world they're trying to model.

The fastest way to lose 32 million OpenAI tokens is an engineer's .env file. The fastest way to make sure it never happens again is the postmortem he writes the same evening. Last Monday. One of our engineers was shipping a test feature on flexa.health mobile app. He committed the OpenAI key to the repo and pushed. The internet found it before he finished his coffee. 32 million tokens, vaporized in a few hours, by some very enthusiastic stranger watching the repo. The thing that saved us was not vigilance. It was a quota. The OpenAI account was new, capped low - the kind of cap that feels patronizing when you are trying to get work done. He had been quietly resenting it for days. That cap was the only reason the bill stopped where it did. Most engineers would have stopped there. Rotate the key. Move it to the backend. Breathe out. He kept going. That same evening, he wrote up what he had done wrong and what he had done to fix it. Dropped it in Slack. Tagged the team. By the time the rest of us logged off, we had all read it. "Never. Put. Keys. In. The. App." he wrote at the bottom. "The bots are always watching." A research-grade shop is not the one whose engineers do not make mistakes. It is the one whose engineers write the postmortem before the thread goes cold.

Five Fast Code AI engineers broke our client flexa.health's iOS app in five days. And this is why no one else can break it now. None of them had written a line of the code. That was the point. April 9, 5:16 PM. Jayesh dropped a message into our new Slack channel: "Did anyone spend any time on the app since Tuesday?" The thread sat. Three hours later, Arjun added a nudge: "plz do this on priority! Takes 10 min and you actually get to do some exercise." Monday, Vishal Kashyap signed up. "Here are my findings." Eight replies threaded under it.   Tuesday, Prabhanshu, our creative director posted her feedback. Eighteen replies followed.   A back button that led to the wrong screen. A UX detail that technically worked but felt wrong. Places where the app needed more context. Moments where a real user would pause, hesitate, or get confused. Nobody was assigned ownership for "feedback." Nobody waited for a review meeting. One person noticed something. Another reproduced it. Someone else added context. The loop kept tightening. Most companies ship to their client and pray. We don't ship until five of our own have broken it. The cheapest engineer to find a bug is one who has never seen it.

Never bet against Yann LeCun. I am saying that because I did. October 2013. NYU Courant. Sixth floor. Theano was barely three years old. Out of Yoshua Bengio's lab in Montreal. Torch had been around for a decade. Lua. Battle-tested. Yann's group used it. Every postdoc on the floor used it. I picked Theano. Why? Because everyone was on Torch. I wanted to be inside the new framework. To find its rough edges. To fix them. I did. My name is on the Theano paper - the 2016 Development Team writeup. By then I was back at IIT Bombay. The work was done in the NYU Courant years. 1000+ citations as I write this. I was part of the core dev team. I also wrote half the papers I should have written during my postdoc. Every Theano build took ten minutes to compile a computation graph. Every experiment was an evening's wait. The Torch people were running three iterations a day. I was running one. Yann would walk past my desk every few weeks. He would look at my screen. Pause. "You are still using Theano? When are you moving to Torch?" I would say no. Politely. Confidently. Stubbornly. I had a thesis. I had reasons. I had a GitHub branch. Yann had taste. Twelve years later, I am clear about which one mattered more. Theano is dead. Torch became PyTorch - built at FAIR, Yann's lab - and PyTorch is now the framework most of the field runs on. The bet I made in 2013 was wrong by every measure available twelve years on. I do not make that mistake twice. When Yann talks about world models now - JEPA, energy-based, learning the structure of the world before predicting tokens - most of the field is half-listening. Same way they half-listened to him about Torch in 2013. Same way they half-listened about backprop in the 1980s. I am not half-listening. At Fast Code AI we are building world models for enterprise. Not for video generation. Not for AGI dreams. For the specific problem most enterprises actually have - what decision should I make today such that the trajectory  of my business is better three years from now? That is not a trajectory-prediction problem. It is a decision-prediction problem. The world model is the substrate; the decision is the output. We are three months into building it. It is hard. It is the right hard. I have been wrong about Yann once. Never bet against Yann. I am saying that because I did.

For some tasks, we do not promise out-of-the-box generalization. We promise calibration. October 2022. An EPC company in Abu Dhabi. The dashboard said 0.96. The new batch did not. 26 centrifugal pumps missed. False positives on isolated vessels. Line detection mAP below 0.6 on plants we had never seen. P&ID digitization is our IP. Fast Code AI built it from scratch. This was one of our first enterprise clients to put it through real production drawings at scale. The lead engineer on the video call asked the only question that mattered. "Why is the new dataset so different from the older one?" I muted myself for a second. The dashboard still said 0.96 - on data we had curated for years. Petrofac's drawings used slightly different layouts, different symbols, different leader lines. What we sold as "P&ID digitization" was a long-tail problem disguised as a generic computer vision task. I had spent three weeks telling them the system was ready. The data said it wasn't. I unmuted and told them the truth: every new customer, every new plant, every new department breaks our system in some new way. We need a two-week calibration phase per plant. We see your drawings. We tune thresholds. We add symbols specific to your conventions. We validate on your annotated sample. Performance keeps improving as we onboard more plants. But it will never be bulletproof out of the box. What we did not say - could not say - was "it works out of the box." We had thought it did. The data said otherwise. That conversation reshaped how we sell. Not just to them - to every enterprise client since. We no longer promise out-of-the-box generalization. We promise calibration. We tell every new client the first two weeks will look worse than the demo, and we tell them why. The clients who stay are the ones who trust the honesty more than they wanted the magic. If you build enterprise AI and your demo accuracy is high, that is the demo. Your first new customer's data is your real benchmark. Plan the calibration phase before you sell. Mine cost me a video call I almost avoided. Yours might cost more.

Last week, Manali ran a correlation study at Fast Code AI: every individual interviewer's vote against every final hiring decision. The question - whose "hire" votes most consistently produced an actual hire, and whose "reject" votes most consistently led to a reject? Highest signal, lowest noise. Four names came back: Narendiran, Sanjay, vaibhavi, Harsh. They are now teaching the rest of us. A masterclass on what they listen for in the room - the same way we'd treat the engineers who reliably ship the cleanest code. Find the signal carriers. Let them recalibrate everyone else. Most companies score candidates. The harder problem is scoring the people doing the scoring.