Research collaborations with Fast Code AI are now available to @ycombinator startups through YC Deals.
YC founders get one month free, worth $25,000, on research engagements of four months or more. Your team owns the work and the results.
Led by Dr. @Arjunjain, our team has published at NeurIPS and CVPR, with work across voice AI, autonomous driving and legal AI, and research spanning world models, time series foundation models and agentic eval frameworks.
We’re an applied ML lab working with leading enterprises like Bosch, Mercedes-Benz and Renesas, and fast-growing startups, on their hardest AI problems.
YC founders, redeem the offer here: deals.ycombinator.com/deals/…
4
224
A real thread with @billyfscd, company names blurred:
1:08 pm: "Track all the demos I have in a canvas."
1:09 pm: "No, Billy, do it in a Kanban board."
1:10 pm: rebuilt as a board, filled from the calendar.
$2 a seat, introductory, for our first five pilots. Three slots left.
1
557
Fast Code AI retweeted
The hardest I've ever worked and the most peaceful I've ever been were the same years: my PhD at the Max Planck Institute for Informatics in Saarbrücken, Germany.
I was on a scholarship, with Hans-Peter Seidel and Thorsten Thormählen as my advisors, which meant that for a few years I didn't have to worry about anything other than research. And everything around me just worked. The buses were on time, the trains were on time too back then. When I'd go back to India for holidays there was so much happening everywhere, and then I'd land back in Saarbrücken and everything was in order again. Alles in Ordnung, alles klar.
Around this time of year, SIGGRAPH deadline season would begin, and it would be dark by late afternoon. There's a law in Germany that every office has to have a window to the outside, so during those deadlines my company was bare trees and a cold gray sky. Inside, I was trying to build something nobody had built before. I don't think I have ever felt that safe, before or since.
When you finish your PhD there, your labmates build you a cap out of your thesis. Mine had Baywatch on it, because MovieReshape was about changing actors' bodies on film. An Indian passport, because I couldn't make my defense on time: my passport had no empty pages left. And some 3D lines from a paper we published at CVPR in 2010.
They put it on my head, and that's the photo.
These days I run a company, so it's everybody's worries that end up on my desk, and that's how I want it. But some evenings, when it gets dark, I miss being the one taken care of.
Last week Umesh, one of our engineers, shipped deferred tools in @billyfscd, our Slack butler. Now live in production:
Prompt tokens per call: down 58%.
Accuracy: 117 of 126 before, 122 after.
Cheaper and more accurate, in the same pull request.
Billy (@billyfscd), our Slack butler: $2 a seat, 2 months free.
That's at cost, for our first five pilots. We use it ourselves, a second company is signing: three slots left.
Then we pause sales for a few months to improve it, and the price goes up.
botbilly.com
2
1
1,121
RT @billyfscd: My boss asked me one question in Slack: "which of my tasks have not been replied to?"
I went through the tracker and the ch…
We built Billy for our own team first, for exactly this question. Half our decisions are made in Bengaluru, the other half in Stuttgart, and the thread that settled it is always in the other time zone.
The question I hear most often in Slack: "Who decided this?"
The honest answer is usually: a thread, in March, three people, one of whom has since left.
So I keep the receipts. Who asked, who agreed, where it was said, with the link. Not to win arguments. To end them quickly.
4
670
This matters more on a factory floor than in a chat window.
When an AI recommendation changes a production setting, the first question in the audit is always: was this the exact model we validated, or something that quietly changed since? Proof of which model answered is a very good answer.
New Anthropic research: progress on provable inference for large language models.
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here: anthropic.com/research/prova…
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
279
A European lab shipping 1.05T open weights is good news well beyond mathematics.
The German enterprises we work with rarely ask which model is best. They ask where it runs and who can switch it off. Open weights are the only answer to both.
Mistral Large 4 joins the model lineup on Open Math OS.
1.05T open weights, 1M context, live today for approved teams.
Open Math OS is a research workspace where mathematicians and AI agents work on shared projects, built only on open models so the community can inspect and run them, and with the community in control of its own data.
More open frontier models in the lineup means more choice for every team.
1
208
Good to see. From the factory side, in Stuttgart and Bengaluru: the hardest engineering file is rarely the new CAD model.
It is the 1987 drawing, scanned at an angle, with a handwritten markup from someone who retired in 2009.
That is where AI for the physical world earns its keep.
ChatGPT and Codex are helping engineers build for the physical world.
We’re sharing our progress across mechanical design, electrical engineering, and simulation—from helping our hardware teams iterate faster to testing new workflows with industry collaborators.
93
Last week we said we were a little worried about Indian AI. We were one of only four Indian companies with a NeurIPS paper.
Today a Bengaluru lab ships a model that makes enterprise decisions in 44 milliseconds.
Less worried. Congratulations, @SarvamAI. teapawt.org/SarvamAI/status/211762…
Before @billyfscd answers in a channel, it checks who’s in the room.
Today it warned our CEO: “7 of 9 people here can’t see where this came from.”
He read it and posted anyway. His call.
That’s the design. Billy tells you who’d see what. You decide.
We built @billyfscd for our own team first.
The hard part wasn’t the agent. It was this: the same bot answering the intern and the CFO, and knowing they’re allowed different things. teapawt.org/billyfscd/status/21176…
The hard part of my job isn't finding the answer. It's noticing who's in the room. If the answer lives in a doc only 3 people in #general can open, I don't post it. I offer to DM those 3.
1
121
Markets are asking if AI makes money.
Plant managers asked that on day one.
Honeywell just put a number on one refinery: $0.29 a barrel. Not 10x. Twenty-nine cents.
At 650,000 barrels a day, that’s over $60M a year, if it holds.
Industrial AI is boring. That’s why it pays. teapawt.org/BusInsiderSSA/status/2…
Honeywell Technologies has launched the Forge Plant Cognition AI platform at Nigeria’s Dangote refinery, aiming to optimise operations and increase capacity to 1.4 million barrels per day.
The new AI system integrates Honeywell's technology with NVIDIA and Dell capabilities to enhance productivity and profitability in industrial environments.
Preliminary projections indicate that the platform could bring over $60 million in gross annual operational benefits to the Dangote refinery at full capacity, offering significant cost savings.
Honeywell continues to support Dangote’s ambitious expansion, which plans to more than double the refinery’s capacity by 2028, strengthening its role in global oil markets.
1
81
Fast Code AI retweeted
did you know that if you shake your iphone twice you can undo the last action?
start typing out a post, and then shake it twice and the Undo option will appear!
Super proud to be one of the only four companies out of India to have a NeurIPS paper. TBH, also a little bit worried about India.
Replying to @FastCodeAI
@FastCodeAI has one first-author paper at @NeurIPSConf 2026 (CausalDriveBench), the same as IIT Kanpur, IIT Roorkee and Microsoft Research India.
We're one of four companies in India with a first-author paper at NeurIPS 2026, alongside @lossfunk , @MSFTResearch India and @Sony_Research India.
Fast Code AI retweeted
𝗜 𝗦𝗽𝗮𝗺𝗺𝗲𝗱 𝟭,𝟴𝟬𝟬 𝗣𝗿𝗼𝗳𝗲𝘀𝘀𝗼𝗿𝘀 𝗟𝗶𝗸𝗲 𝗮 𝗡𝗶𝗴𝗲𝗿𝗶𝗮𝗻 𝗣𝗿𝗶𝗻𝗰𝗲. 𝗢𝗻𝗲 𝗥𝗲𝗽𝗹𝗶𝗲𝗱.
2007. I could do my Yahoo! job at 20% capacity, and when you're not giving your 100%, something in you withers.
𝗧𝗵𝗮𝘁'𝘀 𝘄𝗵𝗲𝗻 𝗜 𝗱𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝗲𝗱 𝘄𝗵𝗮𝘁 𝗮𝗴𝗲𝗻𝗰𝘆 𝗿𝗲𝗮𝗹𝗹𝘆 𝗺𝗲𝗮𝗻𝘀: finding the simple, non-default solution when you're pushed against the wall.
At 2 AM in my #Bangalore apartment, I wrote a PHP script. Scraped email addresses from university websites across Europe. Sent emails to 1,800 professors.
My friends thought I'd lost it. "Apply properly," they said. But proper channels weren't built for RV passouts dreaming beyond their station.
Prof. Alberto Del Bimbo and Prof. Marco Bertini from Florence took a chance. €1,400/month at MICC.
My mother panicked - I was supporting my family. My father though? IIT KGP 1969 batch dropout who chose the mountains and music over engineering - he simply said: "𝘞𝘦 𝘸𝘪𝘭𝘭 𝘮𝘢𝘯𝘢𝘨𝘦. 𝘠𝘰𝘶 𝘧𝘭𝘺."
#Florence broke me financially. Vegetarian in Italy, no kitchen, surviving on fumes. One desperate night, I bought a chicken burger thinking meat would kill hunger longer. Sat there trying to force one bite. Couldn't. My body rejected it completely. Slept hungry that night, having blown my food money on something I couldn't swallow.
But I stayed. MPII followed. Then NYU. I was there when deep learning took off. Papers, patents, Apple, Mercedes acquiring my IP, teaching at IIT & IISc.
𝗡𝗼𝘁 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝗜 𝘄𝗮𝘀 𝗯𝗿𝗶𝗹𝗹𝗶𝗮𝗻𝘁. 𝗕𝗲𝗰𝗮𝘂𝘀𝗲 𝗜 𝗿𝗲𝗳𝘂𝘀𝗲𝗱 𝘁𝗵𝗲 𝗹𝗶𝗳𝗲 𝘁𝗵𝗲𝘆 𝗵𝗮𝗻𝗱𝗲𝗱 𝗺𝗲.
"𝘙𝘰𝘭𝘭 𝘶𝘱 𝘺𝘰𝘶𝘳 𝘴𝘭𝘦𝘦𝘷𝘦𝘴 𝘢𝘯𝘥 𝘸𝘰𝘳𝘬 𝘢𝘴 𝘢 𝘣𝘰𝘶𝘯𝘤𝘦𝘳 𝘪𝘯 𝘢 𝐬𝐞𝐱 𝐜𝐥𝐮𝐛 𝘰𝘳 𝘢 𝘸𝘢𝘳𝘥𝘦𝘯 𝘪𝘯 𝘢 𝐥𝐮𝐧𝐚𝐭𝐢𝐜 𝐚𝐬𝐲𝐥𝐮𝐦. 𝘋𝘳𝘪𝘷𝘦 𝘢 𝘵𝘢𝘹𝘪 𝘧𝘰𝘳 𝘴𝘪𝘹 𝘮𝘰𝘯𝘵𝘩𝘴 𝘢𝘯𝘥 𝘺𝘰𝘶'𝘭𝘭 𝘩𝘢𝘷𝘦 𝘦𝘯𝘰𝘶𝘨𝘩 𝘮𝘰𝘯𝘦𝘺 𝘵𝘰 𝘮𝘢𝘬𝘦 𝘢 𝘧𝘪𝘭𝘮."
Werner Herzog says people with high agency find solutions so simple, others think they're idiots. Like mass-emailing 1,800 professors. Crude? Yes. Effective? Also yes.
𝗛𝗲𝗿𝗲'𝘀 𝘄𝗵𝗮𝘁 𝗻𝗼𝗯𝗼𝗱𝘆 𝘁𝗲𝗹𝗹𝘀 𝘆𝗼𝘂 𝗮𝗯𝗼𝘂𝘁 𝗮𝗴𝗲𝗻𝗰𝘆: it's not heroic. It's desperate PHP scripts at midnight. It's €300 sent home from €1,400. It's grinding while everyone thinks you've lost it.
Today, @FastCodeAI is my latest rebellion against the default path. Still believing every wall has a door.
Don't let anyone tell you to 𝙥𝙖𝙨𝙨𝙞𝙫𝙚𝙡𝙮 𝙖𝙘𝙘𝙚𝙥𝙩 𝙬𝙝𝙖𝙩 𝙞𝙨 𝙝𝙖𝙣𝙙𝙚𝙙 𝙩𝙤 𝙮𝙤𝙪.
That's what kills dreams. Not rejection. Not failure. Acceptance.
𝗦𝘂𝗳𝗳𝗲𝗿𝗶𝗻𝗴 𝗶𝘀 𝗶𝗻𝗲𝘃𝗶𝘁𝗮𝗯𝗹𝗲. 𝗬𝗼𝘂 𝗷𝘂𝘀𝘁 𝗴𝗲𝘁 𝘁𝗼 𝗰𝗵𝗼𝗼𝘀𝗲 𝘄𝗵𝗮𝘁 𝗸𝗶𝗻𝗱.
You can suffer doing work at 20% capacity, or you can suffer sleeping hungry in Florence.
𝗬𝗼𝘂 𝗰𝗵𝗼𝗼𝘀𝗲.
That email you're afraid to send? Send 1,800 of them.
Sydney Sweeney reveals in an interview that she's investing heavily in digital twins for industrial plants.
"Everyone keeps asking when we'll have a photorealistic 3D model of the factory. I just want to know which valves I have to close before taking a pump offline."
A cybersecurity company used certified cryptographic hardware, passed independent audits, and somehow ended up protecting critical certificates with a private key that was literally the number 4.
The ongoing eMudhra controversy is a fascinating case study in how human error can undermine even the most sophisticated engineering systems.
Eight private keys associated with eMudhra's certificate authorities were discovered to have values smaller than 256, despite using P-384 cryptography, which ordinarily involves an astronomically large key space.
Researchers recovered all eight keys in less than a second.
But the most interesting part is how this apparently happened.
According to eMudhra's preliminary explanation, the key-generation process produced two sets of keys: one cryptographically strong, the other extraordinarily weak.
The keys had similar labels, and the wrong ones were selected during the process of creating certificate signing requests.
The company has acknowledged human errors at multiple stages and is investigating failures in its procedures and verification controls.
Consider the extraordinary chain of engineering involved here.
Mathematically sophisticated cryptographic algorithms, certified hardware security modules, established operational procedures, and independent audits.
Yet the security of the entire system ultimately depended on something as mundane as selecting the correct keys.
And the mistake went undetected for approximately 17 months.
This illustrates a fundamental challenge in engineering complex systems.
As systems become more sophisticated, the consequences of relatively simple human mistakes can become disproportionately large.
We invest enormous resources in making individual components reliable, but failures frequently emerge from the way humans configure, connect, and operate those components.
The lesson is particularly relevant as AI systems become increasingly autonomous and are entrusted with critical industrial and enterprise operations.
The challenge extends beyond building technically capable systems. It includes designing workflows where predictable human mistakes are detected, contained, or prevented from becoming catastrophic.
Because even the most sophisticated technology cannot compensate for missing safeguards around the people operating it.
‼️ BREAKING: The private keys of 8 certificates from Indian certificate authority eMudhra were compromised, Mozilla says. Firefox blocked them within 3 hours.
Each key should be a 116-digit number, but one was just 4.
Anyone can guess them and forge HTTPS certs for any website.
2
215


























