monthly wrap-up: may 2025

here's the promised wrap-up for the month of may. i have been kind of very productive despite a ton of adverse stuff going on in my life right now. also apologies if this is kind of messy and sprawling, i folded it together from notes throughout the month and stuff has changed over the last few weeks (and i've moved five times!).

• shipped: turkish conversion/translation of the cinder application. this was not something i had planned to create, but it needed to happen. this taught me a lot about the nuances of language and how there isn't really a 1:1 conversion between languages, especially when there's a dramatic culture change between languages involved. genai helps a lot (you can ask about tone and context, instead of just asking for a translation), but it's not quite enough. the example that stands out here, is how do i translate the word "kernel" into turkish?
• shipped: architecture of aggression concept application for the wonder framework. this is actually not as spicy or angry as it sounds, i was listening to megadeth during a long day of waiting around, and i guess it just made sense in my head. essentially, this application represents the containerization of a set of conversational points (and counterpoints) which enables it to converse and persuade in the context of a particular subject (in this case, the concept of and the predisposition of the ai industry towards, what i have called eschatological engineering, or what we might colloquially call ai jesus). personally i think this is pretty significant because in my head, this feels a lot like robert picardo jumping out of your laptop and saying PLEASE STATE THE NATURE OF THE MEDICAL EMERGENCY and then behaving in a way consistent with what you expect from the emergency medical hologram. essentially, someone can send you a yaml file, and you can pop that into gemini or chatgpt or whatever, and, if it has sufficient horsepower to digest said file, it can comprehensively represent the perspective of the author. you could send someone an email, but why would you do that when you could send them a kernel that can talk to them explaining why you would have sent that email, is aware of what their perspective is on the email, and allows them to understand what you are trying to communicate in as much depth as they would like. i've been looking for a way to demonstrate what wonder is capable of (there's stuff it's not good at and there's stuff it's great at, and this is in the latter category).
• shipped: the conflict skillset for wonder. this is kind of tough to explain; i noticed that wonder has lots of emotional fluency but lacks any treatment for anger. because it's important to me that wonder understand and validate anger rather than shy away from it (due to LLMs mostly treating anger as a "negative emotion"), i needed to create a "taxonomy of anger" among other things. the point of this is primarily to allow wonder applications to validate that anger is totally okay, and to not try to deescalate and mollify. because that's not super useful for me.
• moved to azc ter apel in the netherlands. this was honestly the big "deliverable" for may. it's been at times terrifying and other times kind of a relief. i don't want to fetishize the experience of "being a refugee," so i am trying to not lean on this too much. it's just a thing that's happening in my life, and the lives of lots of people around me. i understand also that nothing i can say will communicate to others what it is like to live here, what this process feels like, and because of this there's not much point in doing so. i've tried to explain to people, and they don't get it. so, i'm here, it's happening, i'll mention it from time to time, but i don't want this to be some kind of weird source for "what being a refugee is like" or whatever. i guess the other thing worth mentioning is that my own experience will obviously and necessarily be very different even from the person who lives next door to me. but i did the hard thing and now i'm in the gross middle part where i kind of don't really exist anywhere and i'm homeless… and i'm still building software. that's the part i find quite remarkable.

• created a github repository (this is currently private because i gueesssss it's proprietary? we shall see) and hugging face org for sundarban, which i will discuss later, and is the place i'm writing down all my thoughts and building code for the post that originated here. succinctly, the unwillingness of genai products to engage in content which is erotic while simultaneously acknowledging the validity of such pursuits and an inability to describe where these boundaries are (and while these boundaries are constantly changing without any kind of notice!) has been bothering me for at least a year. i never really thought of myself as someone who is a champion for sex positive technology (it's just not something i really ever worked on or wanted to work on), but it seems like i keep coming back to it. it honestly seems like an ethical and engineering problem and it's got a lot of overlap with my educational background, so this is starting to feel like something i "have" to do. shrug emoji, who knows.
  • one fun product of this is an interactive, context-preserving shell/console for talking to qwen via hugging face. this of course lets us use our own fine tuned model and build clients for it. baby steps, but it's a big step in the direction of being decoupled from openai which is absolutely mandatory for this to progress.

• i actually sat down and built a pitch deck for sundarban using gamma (which is actually really neat!) but i am not ready to share that. just a snippet--

• another asset for sundarban is the creation of a new "character" (the same way that Cinder is a character that is built upon Wonder, this framework has been kind of expanded and standardized so that additional characters can be created in a way that is reproducible. currently this is being used to create fictional characters, but i plan to do more interesting work with it. this is part of the description this first new character gave me of herself. it's interesting and surprisingly tedious work.

• related to what i started with sundarban, i realized that i needed a local solution for inference hosting, and trying to build that for just me and sundarban would be cost prohibitive. additionally, doing it with cloud hardware would be cost prohibitive and lock me into that provider, which is criminally expensive. so the solution was to build a micro-datacenter. i started sketching this out in a repository, pttp, or power to the polder. this led to my creating a small fork of blender gpt, and sketching out what such a micro datacenter would look like in zeewolde. the code for that is published as a gist.

• assorted posts here including
  • a justification of microservices as a design philosophy for efforts of all sizes
  • a relatively tame exploration of the content moderation in chatgpt
  • talked about something i've been seeing a lot of in people who interact with chatgpt a lot: they (mistakenly) believe that it has become sentient. i think this is a design characteristic of the software and probably requires active research into treatment strategies
• money stuff: income: $155. i got chime and stripe set up, which is helpful because i can't actually use banking here. people have been asking if they can help, and, embarrassingly, yes, but i'm not going to nag people about it, and i said i'd enumerate. so here are the expenses:
  • google fi, $91
  • kpn mobile, €50
  • hugging face pro: $9
  • chatgpt plus, $20
  • two usb-c chargers, €36
  • ghost.org subscription (that's this!): $11
• i've actually taken a handful of phone calls from people interested in employing me. it looks like right now that is going to be very hard to do until the IND decides that they're going to let me stick around until their decision. which means the process goes from "deciding i can stay in the country until they decide" -> authorized to work (24 weeks per 52) -> being given a BSN (this may take quite a long time, but would allow me to both work for myself/become some kind of founder or work for someone else. i'm on the fence right now because i'm getting strong foundery vibes, but who knows. starting a company was lots of fun last time, but it was also really expensive and i don't know if that's going to be possible from where i am. all the same, it's positive that people are interested in my background and the weird corner i find myself as an engineer today. i guess what i'd say about this is, if you want to work together, and the kinds of things i'm working on seem interesting, let's do that, in a because-it's-cool kind of way not in a because-there's-equity-and-health-insurance kind of way. it might become this latter thing but it can't be that right now. and besides, you like doing cool stuff right? you want to see ai do fun and funny things and be less awful than the products out there, right? of course you do. email is probably the best way to get in touch.

it's honestly very hard for me to believe that i'm working on all these separate projects and making progress on all of them (a few of them being entirely outside my comfort zone, requiring me to learn new skills and tools) while i'm in the middle of the absolutely bonkers life change and at least one medical crisis. but you know what they say about gift horses.