As some might know, I don't open source all the projects I work on. I do this out of a perceived [[Open Source Dystopia|moral obligation]], because I consider some of the paradigms and technologies to be possibly dangerous. I reserve the results and project reports from these exclusively for job applications to trusted institutions or for my own amusement. # Conviction My latest such project is expected to cost me around ~$10,000, and is something I have been working towards for about a year. In fact, I made my twitter account and all the open source projects I published with the explicit intent of landing a job to fund this project. I have a long METR time horizon! > So, why didn't I tell anyone earlier? Well the first problem as I mentioned is that I'm anti-open source, so I'd rather be forgotten by history than contribute to something I see as morally disastrous. I'm pretty happy as is, there's no reason for me to be that selfish. The second problem is that no one would care, I was a no-name undergrad from a no-name school a year ago, so if I had written an essay on a technology I'd just be viewed as one of the many AI schizophrenics on the internet. Today, I have a job and a reputation as a talented researcher in the AI community. If I say that I'm willing to spend nearly my entire bank account on a single project, and that I've been working towards it for a year, I'm not a reckless idiot in the throes of AI psychosis, but possibly worth listening to. Which is a pretty big step up. And this is the subject of this essay: how to differentiate yourself from the schizos. # The bitterest lesson Something that a lot of especially autistic people struggle with is empathy, not because autism is sociopathy adjacent but because if there's anything different about your brain or about your conditions, the generalizations you will make can significantly differ from other people. This is [true for LLMs](https://arxiv.org/abs/2507.12224), and it's true for people too. This makes it a lot harder to understand other people and communicate well. However, it is essential to societal functioning, team work, and yet, pretty much everyone sucks at it. One thing I notice a lot on ML twitter is people having incredible intuition and value but don't put in the effort to communicate it to others. > **Unfortunately, this is all that matters. An idea that can't be communicated doesn't scale and an idea that doesn't scale - *well you know the rest*.** ## You are not owed an ear A lot of autistic people grow bitter and frustrated over this, lamenting over how grifters and scammers raise millions of dollars for their projects, while they wallow in the trenches. However, this is a dumb way of thinking. There are infinitely many people like you that all think they know the "secret to AGI", and from the perspective of an investor, they have no idea which ones are good and which ones are bad. If there were investors competent enough to judge accurately, they wouldn't need you, they'd just do it themselves. You need to speak in their language, in their heuristics, which means establishing yourself in the AI community, having connections that can vouch for you, presenting your ideas pedagogically and well written. ## Idea guys Another good example of this is idea guys that spit out takes all the time, and complain when someone else snipes them. The problem with being an idea guy is that you have lots of ideas, and only a small fraction of them are actually any good. > If you invested the same amount of effort into your average take, you'd probably be in the red. This is why you need to take on significant personal sacrifice for an idea you feel strongly about, it signals that you are willing to put significant investment in this idea in particular, and that you're not just an armchair philosopher. If you feel passionate about an idea, put passion into it. ## Agency To give an idea the effort and resources it deserves you need agency, you need the ability to make plans over the long term, across multiple years sometimes. The path I made to get to where I am now was laid out over 2 years ago and contained many backup plans and failed investments. A good rule, for both research and agency in general, is that your plans should fail 50% of the time, if all of your plans/ideas work, you aren't trying hard enough, if all of them fail, you're shooting too high. For a sense of scale 1. **fail**: apply to regular SWE jobs for the funds for ML research 1. Even with immense software experience, a 2.7 GPA and cooked job market was insurmountable 2. **partial fail**: independent research project in junior year summer 1. failed to get it into a conference 3. **partial fail**: raise GPA (2.7 -> 3.2) 1. didn't raise it high enough to apply for a PhD 2. I also took very hard classes, but I was interested in them, so it was worth it 4. **fail**: apply to machine learning research jobs 1. got my first interviews, but didn't pass 2. I applied to a lot, nothing really worked 5. **success** apply to masters 1. got into northeastern (left after 1 semester for job) 6. **success** network on twitter in the ML community 1. immensely successful, ty kalomaze 7. **partial success** more independent research projects (this blog) 1. this was immensely successful 2. many failed projects, but even null result blogs are successes 3. I would even plan out the order of projects I did to maximize engagement 4. an idea guy has no shortage of ideas, so select the ones that are optimal 8. **success** asked for jobs on twitter 1. this worked, this was my first job 9. **success** DMed CEO of Pangram for a job 1. this also worked, this was my second job It was the summer between northeastern and graduating that I conjured this specific project, though I had become convinced of the core paradigm a year prior and is why I so aggressively tried to scale my career, so I could flesh out the idea and produce technologies from it. # How to write I've been told by a lot of people that my writing and blog posts are very well written. I have no idea how this is the case, because I don't read. I know I make a lot of grammar mistakes, but that doesn't usually take away from the content. My best guess as to why this is the case is because I spend an immense amount of time thinking to myself. Every idea gets revisited from 30 different abstractions, and so when I write, I just choose the one abstraction that is most pedagogical. Figuring out what abstractions are pedagogical takes time, and will require you to be well socialized. I suggest taking a year or two off research to learn to form deep friendships with normal people, this worked well for me. Though, it's possibly too late now. # No substitute for ability This is also good advice for research, doing anything out of distribution is extremely hard, as for something to work, absolutely every component must be correct, it's not enough to be right from one angle, you must be right from *every angle*. So it pays significantly to revisit the same idea from many different abstractions, not just one of them. Most researchers will go their entire lives without having a single good idea, let alone gain the taste and experience needed to discover the many powerful abstractions required to revisit the same idea from many lenses. A lot of this is just down to demographics and how you were raised. If you weren't raised with a STEM background, good luck. Additionally, creativity is necessary, but it's very difficult to learn it, usually it's just given to you by your environment. ## Creativity I'm left-handed, tall, first born, ESL, autistic, ADHD, and used to be very socially isolated. Individually, these things don't amount to much, but I think they've had a big influence on me as a person. Humans are non-linear dynamical systems, and even small alterations in our starting conditions can produce wholly different worldviews. People that spend long periods isolated or to themselves yet still grounded in reality will often have completely novel views of the world. I owe my skill at research to this and my conscience, and especially my roommates during my undergrad years, which keeps me grounded and out of nerdsnipes. I care very deeply about other people, and put in immense cognitive effort to empathize with them. This gives me access to far more life experience than I could get in one life alone. It also again makes it easier not to fall for obvious nonsense (like that human intelligence is unique and can't be replicated on a machine). If you are sufficiently autistic and compassionate, you will eventually figure out how your friends work on a deep psychological level, which will prove useful in ML. # Alternative paths If you can't forge your path in research, get into politics. The world needs informed people to help build policy and take risks. Your friends and their friends and so on are counting on people like YOU to make sure this all goes well. No one else can save us. I myself am getting into it, [calling senators](https://x.com/_ueaj/status/2026398533629132849), reps, etc. soon, when I am settled in NYC (I just moved) I will also start engaging with local politicians and stuff. You can just walk into their office, and making an in-person plea will hold more weight. Just ask Claude how to do this for your area.