This is a walkthrough of my attempt at building a realtime assistant (J.A.R.V.I.S.) that can respond on demand to voice, text, and your laptops screen. It was highly technical, including a lockless ring buffer, FSM (finite state machine), and concurrency via asynchronous channels. All in rust! I will attempt to show you from first principles how this complexity arises. Unlike related projects, you don’t have to click a button to get a response.| Hugo ʕ•ᴥ•ʔ Bear
1 Preliminaries Symbol Meaning $d$ width of the residual stream $L$ number of Transformer blocks $V$ vocabulary size, so logits live in $\mathbb R^{V}$ $h^{(\ell)}$ residual-stream vector entering block $\ell$ $r^{(\ell)}$ the update written by block $\ell$ $W_U\!\in\!\mathbb R^{V\times d},\;b\in\mathbb R^{V}$ un-embedding matrix and bias Additive residual stream. With (pre-/peri-norm) residual connections, $$ h^{(\ell+1)} \;=\; h^{(\ell)} + r^{(\ell)},\qquad \ell=0,\dots,L-1. $$ Hence the ...| Hugo ʕ•ᴥ•ʔ Bear
I recently saw a tweet[1] about how people should go about starting startups/businesses, and it caused me to formalize my intuitions around two distinct types of tech businesses I am familiar with. I’ll call them $O(n)$ startups and $O(n^{2})$ startups. Throughout this essay, let n represent the time elapsed since launch. An $O(n)$ startup grows its key metric (revenue, users, etc.) roughly linearly with time—double the time, double the metric. An $O(n^{2})$ startup accelerates, with grow...| Hugo ʕ•ᴥ•ʔ Bear
dad told me i can make ~3 MAJOR mistakes in my teens and still recover. i have 1yr left and 0/3 mistakes. watch closely. https://t.co/wjzzWedk1n — rohan (@oceanmoist) May 1, 2025 I’m Rohan. I’m 18, study CS at Purdue, and I’ve been building things since I was 9. I’m raising a pre seed round because I need as much leverage as I can get to build the future. I’ve shipped across infra, ML, product, trading, and more.| Blogs on rohan ganapavarapu
NOTE: This is purely a theoretical exercise. Gambling can be highly addictive and have far-reaching social and economic consequences. If you choose to gamble, do so responsibly. how to gamble like an addict I’ve spent some time recently thinking about a deceptively tricky problem: What’s the optimal betting strategy if your goal isn’t necessarily to make money, but to maximize the fun and length of your gambling session? Most betting strategies you find discussed–Martingale, Kelly cri...| Hugo ʕ•ᴥ•ʔ Bear
This essay is about the nature and character of algorithms that would lead to artificial super intelligence (ASI) primarily through recursive self improvement. It is a window into my intuitions at the current time. current bottlenecks OpenAI just recently released a long form podcast talking about primarily pre-training gpt4.5 and the various lessons that it taught them. I never thought I would see such transparency from OpenAI and it serves as valuable insight into the current state of ML re...| Hugo ʕ•ᴥ•ʔ Bear
This is a reference, not a guide. In a modern LLM, the “weights” consist of several distinct collections of matrices and tensors that serve different functions during inference: Token Embeddings - Large matrix mapping token IDs to vector representations - Used at the very start of inference to convert input tokens to vectors - Typically shape: [vocab_size, hidden_dim] Attention Mechanism Weights - Query/Key/Value Projection Matrices: In standard attention: 3 separate matrices [hidden_dim,...| Blogs on rohan ganapavarapu
tl;dr: There’s more to the story than “distribution is the last moat”. It’s clear that the price/difficultly of creating software is going down while intelligence is increasing. I don’t yet think people have internalized the cascading effects of this on the way software companies monetize, and how that is affected by the speed of increasing intelligence. I think there are generally three types of software companies: B2B SaaS companies that sell to other software companies (tech-focu...| Hugo ʕ•ᴥ•ʔ Bear
I have a midterm tomorrow (actually)[0] and instead of studying I decided to write this piece on exactly how I spend as little time as possible on school while achieving decent results (top ~5-10% on tests). This protocol is heavily tailored to me and my background specifically. I do think some people can get a few ideas from this though. tl;dr: Step 0: Optimize health and mental state (sleep, nutrition, meditation).| Hugo ʕ•ᴥ•ʔ Bear
These are my thoughts on zero sum games. From poker to quant trading. Zero-sum games are fundamentally informational contests. The theoretical optimal strategy is often suboptimal in practice because real advantage comes from information asymmetry and behavioral exploitation rather than mathematical perfection. In markets, prices have no inherent meaning beyond what participants collectively assign through their actions. Value is constructed through narrative, liquidity, and behavioral dynami...| Blogs on rohan ganapavarapu
My vision for how AI will create the future. AI won’t replace programmers, but rather make it easier for programmers to replace everyone else. — Naval (@naval) January 3, 2025 core belief Every company will become a tech company People said this during the web era, and they were partially right. Every company has a website, every business manages some online presence. But the web was merely the protocol over which business was done.| Hugo ʕ•ᴥ•ʔ Bear
Disclaimer: These ideas are untested and only come from my intuition. I don’t have the resources to explore them any further. intro CoT and test time compute have been proven to be the future direction of language models for better or for worse. o1 and DeepSeek-R1 demonstrate a step function in model intelligence. Coconut also provides a way for this reasoning to occur in latent space. I have been thinking about the geometric structure of the latent space where this reasoning can occur.| Blogs on rohan ganapavarapu
This post is a critique of the direction of current AI models, particularly the path of scaling of current AI models. AI as collective intelligence Humanities intelligence and progress is really collective, everyone builds off the work of others. Take Tesla and Edison, each vehemently championed two sides of the same coin, AC and DC. Unlike a beehive where individual bees act purely on instinct for the colony’s benefit, humans paradoxically achieve collective advancement through fierce indi...| Hugo ʕ•ᴥ•ʔ Bear
github: https://github.com/Ocean-Moist/rohlang3 I’ve long been obsessed with minimalistic languages that still manage to have expressive power. My usual approach: start with a bare-bones combinator calculus (like SK), then keep adding “one more extension” until it looks suspiciously like a full-blown typed language. rohlang3 is exactly that experiment. It’s a small language built in Rust that tries to be point-free, homoiconic, and (somewhat) dependently typed, all on top of an SK-lik...| Hugo ʕ•ᴥ•ʔ Bear
I don’t mean this in the traditional sense of “perfect is the enemy of good”. I rarely held that idea to much credence anyways. What I mean is that perfection is an unimaginable idea that we have tricked ourselves into accepting as an ideal for various human endeavours. what does it mean for something to be perfect? Generally perfection is ascribed to things that have been analytically decomposed and interrogated. Once split into discrete pieces we apply metrics, metrics that have an up...| Blogs on rohan ganapavarapu
There’s something interesting and beautiful about evolution that is often overlooked: it works through emergence. No central planning, no intelligent design, just countless micro-interactions creating macro-level optimization that appears almost deliberately crafted. the diversity problem Think about what human society needs to function. We need visionaries and executors, leaders and followers, risk-takers and conservatives. We need people who can focus intensely on details and others who c...| Hugo ʕ•ᴥ•ʔ Bear
premise The parallels and invariants between artificial and natural intelligence propose unique practical and philosophical insights into both fields. assumptions Intelligence is prediction–all intelligent systems build models to anticipate future states Self-modeling is computationally necessary–systems must represent themselves to reason about their actions Implementation is secondary–the principles are the same whether in neurons or silicon You can find a defense of these assumptions...| Blogs on rohan ganapavarapu
premise I want to explore where mainstream programming languages are headed, using Java and Go as my primary subjects. java is underrated The reason I say Java is underrated is because it is oddly both easy to learn and its type system is strong enough to make many incorrect states irrepresentable.[1] It has product types and sum types, and makes managing state explicit and easy enough.[2] Java is very “simple” precisely due to its verbosity.| Hugo ʕ•ᴥ•ʔ Bear
purpose Physicists long balked at colloquial usages of “entropy” (especially by philosophers). I believe some usages of entropy as resistance to systematic understanding has precise mathematical grounding in information theory. I am tired of arguing with my physics major friends over this. definitions Shannon entropy H(X) for discrete random variable X: H(X) = -∑ p(x) log₂(p(x)) measures average bits needed to encode X’s outcomes. For any compression scheme C mapping sequences of X ...| Blogs on rohan ganapavarapu
premise After I saw that Github Copilot was able to do my DSA coursework with little guidance (though, I never did use it to cheat) I for sure thought that LLMs would ruin kids ability to learn how to code and that it would handicap their learning of not only coding but the meta-cognitive skills around learning coding (ie, “learning how to learn”). In the past few weeks my view has switched entirely.| Blogs on rohan ganapavarapu
clarification What I mean is that, conceptually, “MOATs” are not a useful abstraction. There is a saying that “all models are wrong, but some are useful” and I am saying that, as a model, MOATs are not useful. I think that asking a founder “what their MOAT is” will not offer useful information about their project. I am also specifically referencing tech startups. what is a MOAT From what I gather, a MOAT is defined as “a company’s (unique) ability to maintain a competitive adv...| Hugo ʕ•ᴥ•ʔ Bear
the analogical physics of creativity Consider paint mixing: two colors swirling together create temporary, unrepeatable patterns. These patterns exist only in transition—between separation and uniformity, in a parallel to mental states. Like those paint swirls, certain cognitive configurations exist only in transition, in states of productive instability. This state mirrors the thermodynamic concept of entropy: a measure of a system’s possible microscopic arrangements. High entropy means ...| Blogs on rohan ganapavarapu
backstory (skip if only interested in technical explanation) Yesterday I had a CS midterm from 8-10pm. I was somewhat stressed as I had basically gambled by not properly studying for it. However I finished early, and I was feeling pretty good. Even though it was a Monday night I felt it was cause for celebration. So all of us CS kids gathered (and one engineering guy with a chem midterm at 9am the next day…) and broke bread, comforting those who lost their battle with Java/concurrency and d...| Blogs on rohan ganapavarapu
This is an example workflow of how to integrate LLMs into your software dev. First script (prompt.sh) does one thing: dumps your entire codebase context into a text file. It first appends the file tree while ignoring what you tell it to ignore (node_modules, files not needed contextually, etc), and recursively finds and auto includes what you want to include (e.g. *.ts, *.tsx), and adds any extra files you need (API docs, project plans).| Hugo ʕ•ᴥ•ʔ Bear
Recently I have been working on what I guess is called an LLM agent, but it’s more just building more ergonomic tools for myself to use LLMs in my life. The code base keeps growing, and I use it to help write itself. What I have noticed is that the patterns I have applied for making code more intelligible to LLMs also apply to making code more intelligible for humans. I think the reason for this is that LLMs can only successfully hold so many details in their context window, and we can only...| Hugo ʕ•ᴥ•ʔ Bear
disclaimer I am not affiliated with YC or have any personal insight into the inner workings of YC. It is hard to track YC’s success (with precision) as those metrics are largely not public, and this article is mainly a speculative guess based on limited history. the thesis Generally, the YC thesis makes a lot of sense. Remove the barrier to entry and networking required to get VC funding and fund smart (curious) people.| Blogs on rohan ganapavarapu
The college startup pipe dream I write this article to pose a question: Will the next trillion dollar startup be started by a college-aged kid in his dorm? Can the pattern of Microsoft, Meta, Amazon, Apple, and Google be repeated? (Bezos being a sort-of exception to this, as he was ancient when he started Amazon at 30). An analysis of historical successes What’s interesting is that all these incumbents started as consumer startups, and most hyperscalable startups that have started in the in...| Blogs on rohan ganapavarapu
inspiration I frequently lost pencils and never took notes for most of my academic career. Eventually I got to the point in my self-studying and in my learning were notes was necessary. I find that writing things down can 2-3x the amount of stuff I can hold in my head at once, while making it easier to reference, and makes me remember it for longer and quicker. pencils I prefer pencils to pens as I can erase pencil but not pen.| Blogs on rohan ganapavarapu
about me I am 18, born in 2006. This is generally a good thing as I am in the prime of life currently. I am not one of those people who think they were “born in the wrong decade”, I think I was born at the perfect time to take advantage of superlinearly growing technological advancements. the internet today I generally greatly dislike social media, although I am an avid user of it.| Blogs on rohan ganapavarapu
In 9th grade my English teacher told me I was writing an assignment with broad over-generalizing maxims and beliefs. I tend to write absolutely, especially when being un-technical/un-scientific. I think this is because this makes my writing more concise and more clearly communicative–unencumbered by the recognition of my limitations or the limitations of my ideas. This is because I take for that granted everything is limited. No idea or set of ideas can be truly be always “true”.| Blogs on rohan ganapavarapu
I recently rewrote the interests section of my blog to be more concise. The primary interest I wrote down was “making things that alter thinking at scale.” When I distilled what I believed to be one of my long-term goals I landed on that. Recently I thought about how this is both similar and different to YC’s goal of “Make something people want”. I find that successful startups do both. People want it and it changes the way people think.| Hugo ʕ•ᴥ•ʔ Bear
Most people fundamentally misunderstand risk. Risk is a broad, somewhat useful, metric that people use to predict future outcomes. Risk is only correct relationally to the system that its defined by. It’s a way to quantify a lack of knowledge in relation to predicted future outcomes. Risk is inherent to all systems designed to predict the future as all systems of human understanding are limited. So how to succeed in high risk circumstances?| Blogs on rohan ganapavarapu
preface I don’t care if you believe or don’t believe in God. I care if you dogmatize science. the problem Unfortunately it is becoming common to portray people who are religious as dumb and claim that God is dead due to science. Our mathematically understanding of reality (physics) is not reality. We are not uncovering the hidden code of the world. It’s all symbolic, operations over symbols that aid in survival.| Blogs on rohan ganapavarapu
There is no shortage of deeply intelligent but characteristically uncreative and cynical people. This is wildly unfortunate. This piece is about one piece of advice I would give this archetype of people. the sum of the total is not the total of the sum Think about cooking. The product has much more “value” than the constituent ingredients alone. A painting is much more than just paint and a canvas. A song is much more than some 808s and a nice melody.| Hugo ʕ•ᴥ•ʔ Bear
The Problem Crypto becoming a commodity is the single stupidest thing to gain widespread popularity. It’s the paragon of post-neoclassical/postmodern economics. If Adam Smith saw NFTs he would shoot himself. What I mean is that value is deconstructed to peoples own assignments to objects, abandoning the notion of any essential value in things. This is closely related to the abandonment of traditional metrics of valuations in the tech sector. The original idea was very good–a radical and e...| Hugo ʕ•ᴥ•ʔ Bear
The Addictive Idea of “Grinding” There is this idea, often sold to young men, that “you just need to choose to be successful,” that success is only a function of how much suffering you can put yourself through. The reason this is addictive is that it lowers the perceived bar to peoples dream life. The problem is this only works if, on a fundamental level, you truly believe it. You believe that your success is unbounded.| Hugo ʕ•ᴥ•ʔ Bear