Fractal Tech Log
Table of Contents
- Week 3, Day 4
- Week 3, Day 2
- Week 3, Day 1
- Week 2, Day 5
- Week 2, Day 4
- Week 2, Day 3
- Week 2, Day 2
- Week 2, Day 1
- Week 1 Reflection
- Week 1, Day 6
- Week 1, Day 5
- Week 1, Day 4
- Week 1, Day 3
- Week 1, Day 2
- Week 1, Day 1
- Week 1: Day 0
Week 3, Day 4
part of me didn't like my group project @fractaltechnyc. i forgot to outline up front how i'd like to grow from working on it, as i normally do with collaborations. when i told my group, they were super receptive! i'm now learning to integrate ai into our planning process 😃
i joked that maybe the real group assignment was to figure out how the normal development process didn't work for u and integrate other tools to reduce tedium/complement your strengths
also for the first time, i let the coding agent run commands on my server. my migration to kubernetes hit a couple bumps. letting it query my cluster and make changes (with oversight) solved the problems fast! haven't updated your stats dashboard in 7 years? np, it can migrate it
Week 3, Day 2
oops i forgot to live my life. got distracted by managing my kubernetes cluster. in a past lifetime (~13 years ago), i had dozens of projects running on a tower at home... man-in-the-middling my own traffic to see what requests my phone apps made. always a IT/dev ops boy at heart
wanting to establish boundaries with my time at @fractaltechnyc. but also wanting to see what happens when i don't. i have a reflex to create distance when an aspect of something is harming me, but i don't go back to figure out how to get the good without the harm
Week 3, Day 1
mondays are agent day here @fractaltechnyc. yesterday, we learned about separating out the brainstorm→plan→implement→review→self-improve↺ loop. kinda crazy that the coding agent can make improvements on output that it wrote only seconds ago... but i guess humans are the same
also started moving my personal server over to using kubernetes. i had a 3 node cluster ~6 years ago but it was too expensive. been using docker compose since. but now it seems the tooling is very mature and you can run a light, single node cluster with k3s. so excited!
Week 2, Day 5
i don't regret spending my fridays 'at the office'. yesterday i got excited about... writing SOPs (standard operating procedures). i made a SOP/checklist on doing projects/writing blog posts at @fractaltechnyc. having a checklist takes the cognitive load off when u work on stuff
i'm not new to checklists. i've used them for myself for years, but stopped making them at some point. though @tasshinfogleman seems to have had great success using them for personal & interpersonal projects. and honestly he wouldn't shut up about SOPs
Use Literate Checklists ☑ for recurring experiences where the checklist actually helps you complete the tasks. Data Structures for Humans [thread] 👇
Week 2, Day 4
the little guy has been trained well
i made a reinforcement learning pipeline and visualized it in the web. try it yourself or read the technical write-up
Week 2, Day 3
feel well rested today. at @fractaltechnyc i finished my write-up of the app i made last week (publish soon) and i think i'm slowly feeling more spaciousness and time. not a lot, but more. one day at a time, ulysse. wouldn't it be great if i celebrated my victories more? 🎉🥳🪩
i met with both @__drewface and @MarkEstefanos to talk about my job search in different ways and i felt seen. i was afraid the advice would be of the form "u need to work harder" but it was more like "let's figure out what's blocking u, what u want, what tradeoffs u need to make"
every 2-3 days, i update my website's code. didn't expect i'd be doing that. it's a relief to know that u can resurrect/change old projects more easily with coding agents. @tasshinfogleman recently reminded me too of the value of sharing one's code online. no effort is wasted 🩷
still feel a tension where i want to do what the main cohort is doing but also work on my projects. i guess that just means i love learning and there's nothing wrong with being curious about how everything works
Week 2, Day 2
spent all most of yesterday writing up a blog post @fractaltechnyc, most of the time trying to convey how the DQN reinforcement learning algorithm works to a technical but non-ml audience. turns out machine learning is non-trivial? who knew
i'm sorta doing my own track here. occasionally feel pulled to both do what the main cohort is doing and also do what i think is best for positioning myself for the kind of job that i want. it's mostly overlapping, but when it's not, there's an unpleasant second-gusseing/tension
one thing that's been a success is most of my lunches and dinners are home-cooked. steamed veggies + fried eggs + fruit. i got a metal lunch box and feel like a soldier when i unlatch it. also drinking lots of my favorite cold-brewed tea brought back from taiwan
still struggling to carve out "meta" time to journal and steer my experience here more. fractal tech is kind of all-consuming if u let it be. and i've been letting it
Week 2, Day 1
You can't expect anyone to remember something forever. Such is the case with Claude Code. Last week, I worked on a an AI-training project and one of the difficulties I faced was that Claude Code would work for long periods of time and forget things I asked. It was helping me to create a training program for a little AI agent that plays games. We'd put the little AI agent into environments we designed and then set up a reward system to teach it to favor certain actions when the state of the game was a certain way.
Unfortunately, if you train it too well on the environments you designed personally, then the little AI agent basically ends up memorizing them and can't take the right actions in new environments it's never seen before. But it's so tempting to do this because it is easy and straightforward. Claude Code was helping me train the little agent and it periodically would forget we wanted it to be able to handle new environments. Almost as if it said, “Look Ulysse! I got it passing all the tests! All I had to do was only show the agent the levels we designed together.” So far, I haven't found a methodical way to ensure this kind of thing doesn't happen besides periodically reminding it.
at a journaling session @fractaltechnyc yesterday, @puheenix made me realize that although i'm impressed with what i did last week , i still feel franticness... i am behind and there's no way to catch up unless i rewire my brain chemistry and start using stims or something
there was some line from @__drewface trying to distill the fear like (paraphrasing) "now that the super coding agents are here, every moment you're not claude-maxing on every part of your process, you're falling an ∞ behind"
my fear of my unemployability, of societal collapse, of my own death, makes me miss my innocence when mom and dad would shield me from life's truths. but there's “no way home”. i have to learn to love the genie instead of putting them back in the bottle
Week 1 Reflection
do u have boundaries/rules u make for yourself when programming with coding agents? this week @fractaltechnyc i realized that i'd started "pipelining" myself... juggling 2-3 concurrent coding agents tackling different problems, trying not to be the bottleneck. i did not love it
years ago, i read about spending time as maker vs manager. maker is several uninterrupted hours of flow, whereas manager is moving between many meetings/tasks. i used to code as maker, but i haven't been keeping my boundaries. now i code as manager? fs.blog/maker-vs-manager/
this optimization to Always-Be-Clauding--to never have a coding agent waiting for my input--doesn't even seem like it's good for the product/company. when do improve my skills? how can i make architecture decisions without time to think? when do the "a-ha moments" happen here?
when i was at twitter, i realized that i couldn't do my work if i was getting slack notifications. so i would just check every hour or so. nobody noticed the change. i wonder if there's something similar here? using polling instead of interrupts for your time?
like
- prepare coding agent as best as possible to work autonomously without your input (e.g. test-driven development so your expectations are clear)
- check progress every 15, 30, or 60 min
- incorporate feedback from manager/customer to make sure your output is enough
and what will i do during those 15/30/60 minutes? i guess thinking. learn skills that are widely applicable and also useful to know for my employer/project. read my coworkers' PRs. journal on how our team could improve its process. read technical articles/papers. relax some
Week 1, Day 6
finished the bulk of the ai training (reinforcement learning) visualization @fractaltechnyc. still needs some polish and a write-up but pretty good for ~1.5 days of work
i'm growing fond of coming up with the interface that i want and asking for it. today, i got the LLM to write a react component for custom materials in three.js. i likely could have written it myself, but i wouldn't have enjoyed it. this way, i get to write the parts that i enjoy
Week 1, Day 5
i planned to be spending most of yesterday @fractaltechnyc adding pizzazz to my project that visualizes training an agent with reinforcement learning, but wouldn't u know that life has its own order of teaching lessons. i spent the entire day working on the training itself
the most important thing i learned was if u want the ai to complete an entire coding task, u write an automated test that only succeeds when the ai has finished the task and THEN let it work. takes u completely out of the loop
one such test i had was counting the cyan pixels in the browser window. sometimes the underlying state of your UI seems like it should have a certain look, but in reality it doesn't. u really have to look at the source of truth (the pixels)
Week 1, Day 4
ya boy might be making training some ai this week at @fractaltechnyc. LLMs? no son, i'm talking the classics (reinforcement learning). potential first project is to train and visualize an ai that plays a game. i know it's not groundbreaking, but i will break new ground with this
really appreciating the autonomy/agency that i'm trusted with in this program. i wasn't even sure what to do with it. never been in an educational environment that kinda got me. i know i keep saying it, but i'm excited to see what i can do in 3 months
something i need to work on is maintaining my other areas of responsibility. i've been neglecting several parts of my life this week because i've been trying to really give go for it with this container. i have to be intentional about carving out time for myself
a line i read today that hit different: "If you give a positive reward for getting closer to the goal, the agent might learn to oscillate back and forth to farm 'getting closer'""
i was like yup that's me... i got know i have a reward system but i still am a prisoner of it
Week 1, Day 3
yesterday @fractaltechnyc we were reviewing material that i'd already learned and i started to worry about whether i'd spend my time here well... BUTT! a conversation with @__drewface got me excited again. we talked about how to make this experience very ulysse-shaped for me
i wouldn't say that i'm necessarily going to make my own curriculum, so much as i'll come up with tasks for myself with large nudges from instructors. that's probably just what i need right now. i'm computationally jacked. i need to start (metaphorically) saving cats from trees
i've got @tasshinfogleman shooting optimism rays at me and now i'm in @__drewface's splash zone. a few more and i might be forced to become an optimist myself. HAPPY OPTIMISTIC CREW ASSEMBLE!!
Week 1, Day 2
wow only 2 days of @fractaltechnyc boot camp so far and it's been an emotional rollercoaster
i'm fucking lost. i don't mean i don't understand what they're teaching, i mean i'm just struggling to get clarity on what kind of work i want
i loved doing #genuary2026. each day, it was the best part. i got deep into rust and compute shaders with the webgpu shading language... i learned so frikkin much
...just not sure making animations will pay the bills 😂
the instructors seem amenable to me creating my own curriculum (nothing concrete yet), which felt good. just working hard in the container for three months will be good, regardless
but what should i do? the intersection of what i want to do and what i'm good at is murky. scary
i've been using @MathAcademy in the hopes of eventually learning linear algebra properly so i can dig more into computer graphics foundations. can't hurt to know it for ML/LLM work either. i'm scared though... what if i don't actually like it? wtf am i even doing?
one win is last night i went hard into the meta with claude (normally, grieving the end of traditional programming jobs blocks me). i asked it to create a system to suggest improvements it can make to itself based on our conversations. crazy how (emotionally) hard it was for me
Week 1, Day 1
overheard/thoughts on today @fractaltechnyc boot camp 🧵
i think the biggest thing i can get from this experience is psychological... move through grieving the changes to the software industry and free up the energy necessary to fly
i'm going to use this post to write up some of the more detailed stuff i'm working on throughout the 3 months ulyssepence.com/blog/post/fractal-tech-log
what's better than vibe coding is understanding how the code works and being able to fix bugs yourself, should the ai fail
u can basically keep asking "how could i improve this" to the ai and your process for everything will just get better and your process to improve your processes will get better...
ops people probably ponder this question constantly about their company
"schoolbrained" = doing the tasks/checking the boxes and then expecting things to happen a certain way/to be rewarded (not acknowledging how much agency u have/accepting the emperor's nakedness)
juggling monkey on a pedestal = don't create fake progress/busy work for yourself or spend all your efforts on the "easy part". it's a lot harder to make a monkey juggle than to make the pedestal on which they'll stand
Augmenting my blog with Claude Code
This is my first day of the boot camp. We were asked to download and work with Claude Code to setup a blog where we could make posts to talk about how it's going/what we've learned. Today was special because I'd never shared my monolith with a coding agent nor any human before. This website is powered by a code repository I've made hundreds of git commits to over 8+ years. It's my baby.
Instead of creating a new blog, I wanted to incorporate the web programming that we'd be doing directly into this one. The challenge was, this site is basically a static site generator that I wrote in python, with all the blog content written in markdown. Where do I put the Typescript code that I'm going to write? In order to communicate what I wanted to the Claude AI, I actually had to do a lot of thinking for myself. "What kind of interface do you want, Ulysse?"
For the psychedelic graphics article series, I made a custom HTML element that would get used if I created a code block with glsl-live as the language. In the end, I decided to make a new custom code block called react that would go and compile Typescript React apps and inject them into the part of the page where the code block would have been. So something like this:
```react:shader-example const backgroundColor = '#1a1a2e'; const rotationSpeed = 1.0; ```
would go look for a React app called shader-example and then load it into this page at this location (backgroundColor and rotationSpeed override parameters in the React app):
Learning about Claude Code Skills
I asked Claude about what skills are. I originally though they were the same as the commands/agents, but it seems they have more structure and a manifest. Several of the commands/agents I used with Claude Code used external files as reference for interacting with MCP servers, but skills inherently are in their own directory and therefore encourage you to put everything the skill needs altogether in the same directory. It's the way that many package ecosystems work, browser extensions, and Unity 3D plugins: put everything in a directory with some manifest file with relative file paths to all the other assets and then it's simple to zip it up and send it elsewhere.
Week 1: Day 0
The introductory thread that I wrote:
i'm starting the @fractaltechnyc boot camp tomorrow. here's an aspiration to journal about it in public
the major feelings going in are (1) excited, (2) anxious, and (3) embarrassed
(1) i'm excited because this opportunity is potent af 🫣 i expect the instructors and participants are curious, ambitious, caring ppl... my kind of ppl
personally excited for how i can grow in this container... it's been so long since i was asked to work really hard in a cohort
(2) the anxiety is like the shadow of the excitement... what if i fail? i want to journal a lot about my experience and be really intentional about tweaking my experience to be extremely ulysse-shaped. what if i get complacent and stuck in an old pattern?
i love computer graphics. i want to cater my projects in that direction as much as possible. if i can move towards that in any capacity, i want to do it. so there's also some anxiety that i just can't cut it in the technical art world without going to graduate school or something
(3) i'm embarrassed bc i couldn't get a job. been doing my own thing for the last 5 years, but otherwise my cv is 😚👌 i've never been more technically capable. i'm computationally jacked--have u seen what i can do? "so then why couldn't u get a job, ulysse?" 🙁
i'm not angry i'm just sad about it. but i accept reality for what it has to teach me here
someone out there wants to pay me a living wage to do what i love and i just need to find them. i will find right livelihood. i will grow a lot over the next 3 months. thanks fam bisous 💋
Curious about hiring me or working together? Check out my hiring page. You can also hear more from me on YouTube or Twitter.