Armaghan "Rumi" Naik

Student and Trainee researchers considering joining my lab - please read: 1. We believe in approximations. 2. Start a research project by writing as much of the paper as you can, hand draw what you think the figures should be, and CITE THE PAST. 3. Only then do the experiments. [That's the scientific method, remember? Hypothesis first and then methods and then experiments.] 4. Force yourself to clarify your thinking. Do it again. 5. You need to develop good taste. That means you need to study masters. Nobody develops good taste by eating McDonald's all day, so why are you guzzling arXiv papers? You develop good taste by reading a broad range of excellent writing. 5a. I suggest reading at minimum three (3) papers a day (the way we teach in our lab, Jones's Method). Print them out. Mark them the hell up. I used to read papers on the toilet and in the shower. Study. 5b. The goal is to not have to read the literature, to BE the literature. Meanwhile you better study. 6. Learn to be dispassionate about fads. You have to know what the hoi polloi are excited about, but winners see ahead of the wave and consistently drive execution with the wave crest. 7. It is almost impossible to fight the prevailing winds of whatever field you work in because humans are pack animals, and most of them have a contemptible lack of courage. You cannot spend your PhD trying to "fight the Man." 8. We will teach you the tools to be a rebel, and win, but first you need to learn how to kowtow to the field and be exemplary at THEIR rules. 9. Most people do not care about you or your long term success. Make sure that you explicitly care about your own long term health and success. 10. Give yourself wholly to your work but remember you are more than the work. 11. Be a good teammate. We wash the glassware and leave the workspace better than we found it. 12. It is not enough to do good work. You must market it exceptionally well. 13. Pass on what you learn. If you're really the best at something, you should teach it. Ok, well you're approximately the best at something already, so teach what you can.

Why would a chipmaking startup CEO have a soft matter / biological engineering university lab? Undergrads are intensely curious, eager to learn research skills, and reveal the baseline of what a reasonably intelligent person who has no domain knowledge would figure out. That's incredibly valuable on its own. Let me give you a finance analogy. In markets the degree to which one company (say) correlates to the rest of the market is called "beta" - it's what you get "for free" by investing in the S&P500, say. But you don't want just the beta. Superior performance requires the stuff that's beyond the expected returns of the market. That's called alpha. The same thing goes for knowledge. If I'm making a superintelligence and I want to understand what a reasonably intelligent person armed with the general tools of the world (Claude, Google, Python, a laptop, etc.) would figure out, it helps me make sure I'm actually building something valuable that is BEYOND that. My job is to crush it with alpha for my customers and shareholders. But also I really love teaching and it forces me to actually know what I'm talking about in terms of simple language and kung fu depth. I am so damn proud of our cmu-prism-lab launch cohort, and grateful for my colleagues especially my Co-PI Rebecca (Bex) because these undergrads presented some amazing work. I wish Ed Clarke and Herb Simon were still alive to see what we're doing with formal methods and AI; I owe them a great debt of thanks for putting me on this path. I wish Bill Brown and Beth Jones could see how we're reshaping biology because they believed it would happen. The natural sciences (including math!!) are being profoundly affected by AI and it's such an incredible time to have tools like Lean4 (Lean FRO) to leverage. We've only just begun. And our undergrads are going to take leadership positions to drive the future. Smart money is backing us because our ideas sound stupid to experts. But where else would a revolution come from? Chipmaking is a soft matter problem in disguise. Back to building ...

It's dumb obvious that the best way to think about chemicals isn't in terms of chemical structures. It's so dumb to say that because what else is chemical besides the structure, right? Every atom is contextualized by its local chemical environment. You've got bond lengths and angles. But obviously something is missing because a hell of a lot of smart people keep trying to make yet another graph neural network or something to discover a magic representation that will make chemical reasoning easy. And so they reach for the place where middling minds go, which is to say "we need more data." What if you need less data and more perspective? But of course saying something totally stupid in public like, "the framework chemists prefer to think about chemistry might not be the right way to represent chemical properties" ... well this is an invitation for ridicule. It's easy in retrospect to mock the ancient astronomers for obsessing about epicycles but epicycles really did make a lot of sense. Circles are nice, everything becomes a matter of really clever calculation, you can make very accurate predictions of eclipses and so on. It was so stupid to suggest that epicycles were the wrong way to think about the heavens. We need more stupid ideas and fewer people who ejaculate "you need a lot of data, large language models prove data scaling wins." Aphorisms aren't proof. Everything about the apparent complexity of chemical modeling is screaming that we're thinking about the problem wrong. Do you want to be Copernicus or one of the stubborn jackasses that got in the way? Because today Copernicus could make a shitload of money for investors.

@davepl1968 Ok but it turns out that the STL really fucking sucked and everyone pretended it was the pinnacle of engineering but then you had to use boost but even boost was moronic

I had this conversation at Microsoft in 1996: Me: "Why do we have our own pointer array code?" Mgr: "Because it's solid and well tested." Me: "So is vector<> in the STL!" Mgr: "Devs don't know the STL" Me: "They're devs, they should know the STL!" Mgr: "That's great, but they don't, so no." And so we continued to use and write all of our own containers and so on. Because the STL was scary.

@exQUIZitely Norton Commander

This one is for the OGs. Norton Commander or PC Tools - which was your #1?

gem5 is slow. "it's because the I$ is too small on x86!" and then you see code like this shit /** State transition matrix */ std::vector<std::vector<double> > transitionMatrix; come the fuck on

What has been archived of Prodigy online? There were graphical and interactive novels - have they been lost? There was some absolutely brilliant new media and I wish I could share it with the Claude Code core retro futurism resurgence kids

@oxcrowx Because humans enjoy the sense of agency so much they don't want to be right quickly

To me, there are only few modern languages that can be called simple: Odin, Austral, OCaml, Scheme, Lua, C3, and Go. Most other widely loved / used languages like Rust, Zig, C++ etc. are fairly complex. Now the dilemma is: Users seem to prefer the complex languages more. Why?

@McDonalds I can eat the hell out of a burger and I'm AI native. OpenHamburgler has access to all my fries and files. I will create logistics crossovers with Domino's that will embrace, not avoid, the Noid. Then we can take on Amazon by turning every McDonalds into a McFillment Center

@LeMangy @arvidkahl I used to have a great Dr. sbaitso impression. "I cannot f--k" was so much fun to get him to say

@arvidkahl wrong

ffmpeg likely is the best piece of software ever written.

@aditiitwt I'm Commodore Basic / Turbo Pascal old

You may be old But are you this old

@filpizlo @ujkan_s It was not clear to me until this comment what point you were making.

The "types aid productivity" claim is unverifiable. I think that whether a language is more productive or worse is not correlated with whether a language is dynamic or not. There are dynamically typed languages that are unusually productive, and there are statically typed languages that are unusually productive. But, for a given language, adding or removing types might make it more or less productive.

One of the most broken memes in programming languages is that types aid productivity. They don't. Types are a preference, at best

@djtodoro Thanks for the plug

iree (iree.dev) is very powerful (ML) compiler infrastructure! We need more Toy-like tutorials there, but it is very very powerful tool.

Mostly my feed is just nerdy useful stuff but occasionally I see breathtaking antisemitism or antimuslim comments. Guys, that's just not acceptable.

@tomfgoodwin You don't suck at thinking love this

If AI can do 70% of your job, it’s often the case the remaining 30% is still invaluable

I think most appeals to "fat tails" are really just the law of large numbers (most numbers are really, really big)

@FelixCLC_ Ouch

standard libraries should be numerically consistent across implementations.

@mean_field_zane I went to CMU and had incredible peers and lifelong friendships with many classmates that have saved my ass over the years. Pretty sure peers matter more even than the incredible teachers I had.

There’s a common idea that what makes an elite university different than a state school is the peers you’ll have there, and that the classes are mostly the same. In my experience the opposite is more true. People at top schools are still very academically disinterested and divested from rigour and open debate, but you will have access to more advanced classes in things like mathematics and economics (though the quality of professor is usually equivalent).

What in Chatgpt are you talking about. Have you even seen a chip design? A real one? Chips are utterly boring corner case after corner case and physics requires them to be locally dense - you can't make long wires. That's not at all how the brain works. It's tedious but obvious how to figure out what a chip is doing.

If you looked at a logic gate level of a central processing unit (CPU) execution of even a simple algorithm, you wouldn't know exactly what is going on & you would think the CPU is brain-like because the logic gates would resemble neurons.

@runaway_vol Warum?

Germany is launching an alternative to Claude Code, it’s called Klaus Programmieren and it will run on sovereign ai system with chips made in Germany. The project is currently in the planning phase but the Federal Government has already committed a record sum of 50 Million Euros