CryptoPatrick

Math Major | Daily Diary D0064 Worked a bit on my website today. Visitor stats confirm it - I'm going global, baby? Nevermind the fact that I'm only averaging 7 visitors per day lol. To my visitors from China, the US, and other countries: thank you so much for blessing me with your visits. I will double my efforts and try to make the website worth your time. 😀

You're most welcome, milady.

Math Major | Daily Diary D0063 📢 Update: I haven't heard back from Prover.com despite mailing a follow-up email to my application. 😕 I'm clearly a little bit disappointed - I poured a lot of energy into my application - but that's just the part of the internship/job application game ig. 🤷‍♂️ 🔬 I'm going to continue to focus on my undergrad degree, look for a part-time job, and read/think/build/write as much as possible. 🧗‍♂️ Further to follow.

I once dreamed about becoming a writer. My parents said "No!". Tried my hand at writing fiction, for the first time, last year - in the form of this short story "The Time Lock". It sucks. So, I guess my parents were right :D Anyway, back to math hw.

Math Major | Daily Diary D0061 Over the last two days I have spent time with Bend2. Coming from Rust and Python, my first step was to try and get a sense of two things: 1. Is Bend something which I have a use for in the immediate or long-term future? 2. What's the bare-bones syntax and concetps that I need to master in order to place Bend in my "Hey, maybe we could use Bend for this?" mental toolbox. Now, Rust is blazing fast, but it's not a native in the parallel dimension - we need Rayon for that. So, the question becomes - do I know how to identify and leverage a language like Bend to alleviate hotpaths in my code bases. Well, turns out - I don't. Which is great. Because it gives me a great reason to learn Bend and work with it. My first mini-project has been to write a profiler with the specific focus on identifying not any hotpath but ones where it seems possible to divvy up the workload in ways suitable for Bend and GPUs. I wanted to share the Bend2 tutorial by Ivo Balbaert, because I had trouble finding a good tutorial myself. I'm just a poor student, so I don't have a GPU cluster at the ready to run my noobie Bend code on - but! I was thrilled to see that Bend2 supports Apple Silicon GPUs (M1, M2, M3, M4 and newer) through Apple's Metal backend. Very cool! My only gripe is that the geniuses over at @AnthropicAI have made coding so much fun that I'm blazing through my $20 sub in no -time, and then have to sit around and wait until the session resets. On the other hand - waiting for session resets is perfect for doing my math homework, exam prep, write essays, read papers, hand code Rust old school style, and read lots of books. Further to follow! github.com/Ivo-Balbaert/Bend…

Math Major | Daily Diary D0060 1. Find the joy in what you do 2. Put in consistent hard work 3. Embrace precision 4. Remove clutter 5. Question legacy 6. Think BIG! 7. Attack hard problems

Math Major | Daily Diary D0060 📎 It's getting harder and harder to remind myself that we're just dealing with regular computer programs; declarations, assignments, operations, for loops, and conditionals - nothing else. 🖇️ I spoted a PhD recruitment post for *Rational ML*, and thought "rational intelligence"? That's highly subjective - how would you model rationality and implement it in a computer program? Very interesting, indeed. 📌 Here's a list of words from the ad, which hint at AI researchers increasingly sourcing ideas from; neuroscience, psychology, behavioral and cognitive science - to expand the abilities of a conditional for loop hooked up to a transformer nn: Rationality Preference Choice Cognition Metacognition Trustworthiness Decision Uncertainty Causality Incentives Coherent Inference Calibration Abstention Faithfulness 🤝 At the center of all this, I guess we find the ambition to improve productive human-ai collaboration. 🔭 At the periphery of all this, I think we'll find a sobering realisation - without a rich and constantly updating world model, scaling collaboration with AI is going to hinge on the 🐱 and 🐭 game of identifying edge cases which turn out to be cornerstones - or, another way to phrase that: 🤬"Man, this AI is stoopid. It just doesn't *get* it.". Rational Intelligence - I hope my actions manifest those glorious words today, when I place my order at the coffee shop. ☕️🫣🍰😅

Math Major | Daily Diary D0059 Is anyone looking at this?👇 I'm running lots work on my machine atm, I don't have time to look into this - but it sounded interesting. gist.github.com/flamerix-ai/…

Math Major | Daily Diary D0059 🥱 Okay. I'm seriously tired. This week I have been coding, doing math, and writing a whole lot. Often going to bed at 2 am, and back up at 7 am. Need some sleep, for sure. 🖊️ Anyway, one of the things I've been working on is a draft of an epistemology paper exploring a topic I started thinking (and writing) about this summer: generative asymmetry. 🧫 Yes, I'm primarily talking about the challenges of Generative AI triggering a flood of results in mathematics - but it is my intention to try and extend the idea to (amd draw connections from) other domains that manifest the same pattern. 📬 The draft still needs a hell of a lot of work - but I figured I'd share it now, in order to break the fear of looking stupid. I hope that in a few months, the draft will be in a state where I feel comfortable asking someone to peer review it. 🔥 As soon as a I get some time over, I'll add a comment section to my website - with the hopes that someone might want to read the draft and suggest improvements. 🧗‍♂️ Anyway, gotta keep moving. There are lots of fun things to work on. cryptopatrick.com/research/o…

Math Major | Daily Diary D0059 😞 Very tragic news coming out of Spain, yesterday. 🚂 Having spent two weeks building a toy/prototype simulator around train collision avoidance, these news are a stark reminder that actual people are affected by collisions - there's a catastrophic reality not that far from simulated trains colliding. 🧸 We can all take it upon ourselves to help create a safer world for everyone.

Math Major | Daily Diary D0058 🔭 Exploring a Midjourney roundtrip where my website pings Midjourney to generate a new image, based on some changing condition, downloads the image and then updates the website automatically. 🛀 I'm not 100% convinced about the use case yet, but I'll try to get the roundtrip in place, as a Claude Code skill.

Math Major | Daily Diary D0057 I managed to get locked out of my LinkedIn account. To regain access they want me to take photos of myself and upload my pictures of my passport. For LinkedIn?! Please. Never touching that abomination again.

Math Major | Daily Diary D0057 🏁 Finally reached version 1.0 of RailCheck. 🏬 Background: two weeks ago I saw a software demo on the webpage of a company that I'm interested in. I got curious about how the different parts of the demo fit together. 🔨 To understand how the different parts of the demo worked under the hood, I decided to build my own prototype - RailCheck. 🎁 There's clearly room for improvements (lots of things are still broken), but at some point I think that it's more important to take a step and ask - has the prototype achieved what it was created to do? 🔥 If the answer is yes, then there is little point in shoveling more tokens into the furnace. 🚉 So, it's time to say goodbye to you RailCheck - I'm getting off at this stop - other projects await. We had a lot of fun together - and during those early mornings and late nights, I learned a lot. 🚆 "Doors closing! Right away."

Math Major | Daily Diary D0056 🏗️ Added a new section to my webpage - a page where I'll track my (tiny) oss (open source software) contributions. 😂 I realized that my oss contributions, in terms of Pull Requests to other people's projects, is pretty much zero. 🔭 I'd like to change that, so I'll be on the lookout for fun code bases to try and contribute to. ⛩️ I like the idea that you should not contribute for the sake of it, but rather because you use something and find something which could be improved. cryptopatrick.com/oss/

Math Major | Daily Diary D0056 🐶 Just for fun, I gave an online AI application called "Requirement Quality Checker" three requirements sourced from a science fiction story. 🐱 Interestingly, two of the requirement rewrites that the agent suggested, kept phrases it had just flagged. That raises the question - who checks the checker? 🐭 I wrote a post about it here (please keep in mind, I'm just a first year undergrad who's trying to learn new things - not a Ph.D. researcher like everyone else on twitter 🙌😀). cryptopatrick.com/research/r…

Math Major | Daily Diary D0055 🚃 Just for fun, I created a "product sheet" for RailCheck - a prototype of a formal verification pipeline for AI generated output in railway signaling. 🚂 I recently applied to a company, asking for internships and part-time roles, and they had a demo on their webpage. 🚋 I wanted to understand how their demo worked under the hood, so I built my own demo from the ground up. 🚊 Even though a demo like mine is 50,000 miles away from a real professional solution - building a prototype from the ground up like this, is a fun a valuable learning experience. I've learned quite a lot about formal verification of ai generated system constraints. 🚉 I've created a list of features that I'd like to add, and I'll try to make small additions to RailCheck over the coming months. ⚠️ Just want to stress that RailCheck is just a fun prototype and a work in progress - it's very far from anything that should be used in a professional setting.

Math Major | Daily Diary D0055 🎃 Imagine that we ask an AI to sort Halloween trick-or-treaters into two groups; kids who get candy and adults who get a trick. 📐 We give the AI what we believe is a crystal clear rule: “Little monsters get treats.” “Adult monsters get tricks.” 🤔 "What about an adult monster who's a little short?" 👥 There's ambiguity at play here, and neither the AI nor we humans may always perceive it, especially in cases that are much more subtle than this example. ☑️ Just because our AI correctly sorts ten cases, doesn't mean we've established who (we or the AI) it was that provided enough information to reach the performance we wanted. An eleventh, incorrectly sorted case, might surprise us. 🎲 Basically, we don't know if the AI worked out the missing context, or if we were just lucky. 🔬 This highlights the importance of providing the AI with the right context, by carefully choosing words that reduce ambiguity.

Math Major | Daily Diary D0055 Adding some minor fixes to my prototype formal verification project, and learning new things. The concept of "saturation" is really interesting. 🎃 For example, in the context of Halloween, imagine that we're part of an exhausted trick-or-treat squad, having checked "every" house in the neighborhood. But did we? 🏘️ Just because we've knocked on many houses, doesn't prove there's none left to knock on. 🧟 Let's assume that our young squad of professional trick-or-treaters is equipped with a list of 10 houses to visit, making this a bounded search. We also tag each house on the list with a "VISITED" tag as we "process the neighborhood for candy". 🏁 Once we've processed the street, it's easy to draw the conclusion; "That's it. There's no more candy.". We've saturated the search - having exhausted the "search space" for candidates to know on. 🍬 However, just because there's no more candy on our list of 10 houses, doesn't actually answer the question "Is there more candy in the neighborhood." 🍭 What we need is to operate under the assumption that there is an UNKNOWN number of houses with candy... at least until we've established that our search is complete and that no more candy remains to be had. 🧛 The "danger" of bounded search aka "Let's check 10 houses for candy.", is that we might unknowingly bound our search based on our perceived dimensionality of things ("10 houses? That's a lot.") - which may or may stand in stark contrast with reality ("Wait, isn't there like 100 houses in our neighborhood?"). 🧬 A Japanese systems biologist once joked at conference; "Physicists are used to working with a handful of variables. You show them a system with 10,000 variables ...and they panic." 😂