I think people are underestimating just how much industrial knowledge Germany is sitting on, and how easily it could lose its competitive advantage over the next decade. Germany has spent generations building some of the most sophisticated industrial companies in the world. Many are family-owned Mittelstand businesses, operating in obscure niches, making products that most people have never heard of, but which entire global supply chains depend on. Their competitive advantage was built over decades. It lives in the accumulated experience of their engineers, the intricacies of their manufacturing processes, and an intimate understanding of problems that competitors may not even know exist. And yet, these companies face an increasingly uncomfortable future. Energy is expensive, regulation is burdensome, skilled workers are difficult to find, and competition from China is intensifying. Meanwhile, technological change threatens to erode advantages that took generations to build. I find it troubling how casually we discuss Germany's industrial decline, as though we are talking about abstract economic statistics rather than businesses that families have spent their lives building. But I also believe AI offers these companies an extraordinary opportunity. Consider the sheer quantity of knowledge accumulated inside a fifty-year-old manufacturing business. Engineering drawings, maintenance records, production data, and decades of experience troubleshooting problems on the factory floor. Much of this knowledge remains fragmented, difficult to access, and dependent on a handful of experienced engineers. Imagine making that accumulated expertise accessible across the organization, allowing engineers to understand complex dependencies and make better decisions in a fraction of the time. This is precisely the kind of problem we're working on at @FastCodeAI. Germany's greatest industrial strength is its accumulated engineering knowledge. Air-gapped secure customized AI offers an opportunity to make that knowledge considerably more valuable. And I believe preserving this advantage should be a national priority.

Sound on. 🔊 Beep. Hum. Pour. Flip. Click. Clack. Strum. A day at @FastCodeAI, in 35 seconds.

A self-driving car can make the right decision for the wrong reason. For an LLM, that can be annoying. For a self-driving car, where the decision can literally be a matter of life and death, being consistently causally wrong is really fucking dangerous. So we built a benchmark to test whether autonomous driving models actually understand why they should act, not just whether they produce the right action. That work has now been accepted at NeurIPS. Here is a brief explainer:

We were genuinely surprised by the amount of pushback, and in some cases outright hostility, to our announcement of Research as a Service. A few people seemed to read it as some new scheme for companies to buy papers or rent academic credibility. That is not what we are doing, and the underlying model is much less exotic than the reaction suggests. I spent years in academic research, first at Max Planck and later at NYU. Research has never happened exclusively inside universities. Companies commission research, universities take on industry-funded projects, pharmaceutical companies work with CROs, and specialist labs are brought in when an internal team does not have the expertise or bandwidth for a particular problem. This has been normal for decades. We are doing the same thing for machine learning. There are companies with difficult ML problems, unusual datasets, and potentially novel technical questions that do not justify hiring an entire research group. Building such a group can take a year. A focused research project might take four months. So they bring us the problem and the data, and our researchers work on it. We design the experiments, build the methods, run the evaluations, iterate when things fail, and try to produce research that is good enough to survive peer review. If that work leads to a paper, authorship follows contribution, as it should. The conference reviewers still decide whether the paper deserves to be published. Paying FastCode funds the months of research required to get to that point; it does not purchase the outcome. I suspect some of the reaction came from the phrase “Research as a Service.” Fair enough. New packaging can make an old practice sound stranger than it is. We published a longer explanation from the @FastCodeAI account because I think the model deserves to be understood before it is dismissed.

5'7 is the ideal founder height. There are a suspicious number of extremely successful founder-CEOs hovering around this height. Mark Zuckerberg is 5'7. Jeff Bezos is about 5'7. Jensen Huang is reportedly around 5'7. Jack Ma is even shorter. There is an interesting reason this might happen, and it has less to do with founders themselves than with the unusual way in which someone becomes a founder-CEO. Nobody hires a founder to be CEO. Zuckerberg never had to persuade a board at Facebook that he possessed sufficient executive presence. Bezos did not have to look commanding enough to be selected to run Amazon. Jensen Huang did not spend twenty years moving through Nvidia's management hierarchy until a succession committee finally decided that he looked and sounded sufficiently like a chief executive. They started the companies, and at some point the companies became enormous while they were still running them. That is a very unusual selection mechanism. A professional CEO, by comparison, has usually spent decades being selected by other people. Think of Tim Cook at Apple, Satya Nadella at Microsoft, Sundar Pichai at Google, Dara Khosrowshahi at Uber or Bob Iger at Disney. Before reaching the top, they have usually survived countless promotions, interviews, presentations, executive meetings and succession decisions, all involving people who already have a fairly well-developed idea of what a senior executive should look and sound like. Competence obviously matters in this process, probably more than anything else. But competence is never evaluated in isolation. Height, voice, appearance, confidence, social fluency, posture and the countless small signals that together produce what is rather vaguely called "executive presence" all enter into the judgment. If someone has to pass through twenty years of other people deciding whether they seem ready for the next level, those signals have twenty years to matter. Founders largely escape this filter. Their selection process is different and, in some ways, harsher: something has to be built that actually works. The market does not particularly care whether someone looks like a CEO if they own the company.

No, AI will NOT kill us all.

It is 2001. I arrive at my JEE Mains centre in a Home Minister's car. "Lal batti" flashing. An armed guard riding along, if I remember right. My classmate Kuber's father was the home minister of Madhya Pradesh. I have stayed the night at their Bhopal bungalow, and his mother cooked dinner for me herself. That is the kind of school Daly College is. My friends' fathers run states. I am the kid whose aunt paid the school fees. The car waits outside through the whole exam. But the paper does not care who drops you off. Inside the hall, I already know it has gone badly. I can attempt only a few questions. I had bet two years on that exam - quit hockey, quit the quizzes and debates, studied while all my friends played. I placed the bet myself, and I lost it. The backup plan was a year in Kota. My father did not have the money for Kota. So: RV College, Bangalore. Cauvery Block, 1st Floor. Nothing about my situation said "researcher." No brand name, no coaching, no one watching. I pulled IEEE Transactions off the library shelf anyway, because the desire was mine and I stopped waiting for someone to authorize it. A year in, I gave the JEE one more shot from my hostel room. Second attempt: 3363. Petrochemical at IIT-BHU. An IIT seat, at last - and I turned it down. I already had one paid-for year of computer science, and just three years left before I could start earning. Starting over meant four more years of fees, in a branch I did not want. Fourteen years later, IIT Bombay calls. Not with a rank. With an offer to teach. In 2016 I walk into CS763, Computer Vision, as the professor. Every student in that room cleared the exam I could not. It never once comes up. The exam's authority, it turns out, expires. I teach there four years. Graduate 2 PhD students. One of them, Rishabh, is a professor today at Max Planck - the institute where I did my own PhD. The credit is his. At most I held a door open. But the symmetry stuns me: my student teaches where I trained, and I taught where I could not study. So if an exam has just told you who you are - if you are stuck, and you know you are capable - hear me: it is not the end. The exam decided where I started. It had no authority over where I ended. Neither does yours. Do not wait for permission. Work like the desire is already authorized, because it is. It will work out - not on their schedule. On yours. So tell me in the comments: the exam that said no to you, and the person you became anyway. Someone sitting tonight where I sat needs to read your comment more than my post. 𝗚𝗼 𝘁𝗮𝗸𝗲 𝘄𝗵𝗮𝘁 𝗶𝘀 𝘆𝗼𝘂𝗿𝘀.