←
Xander1,459 posts
Xander avatar

Xander

@XanderAnyway

engineer. games, machines, and questionable ideas. building @CyberRegnum and whatever catches my curiosity next.

◷ Joined August 2019▧ 5 Photos and videos
Xander avatar
Xander@XanderAnyway·Oct 31

i was going to write a normal introduction here, but there's a situation involving the cat in my banner. he started as a stupid reaction image for my instagram story. i gave him an oversized leather jacket, cyberpunk glasses and sneakers large enough to require a building permit. now he has a name, netrunner cat, a criminal record i cannot legally discuss, and somehow his own place on my profile. meanwhile, i'm xander. i build software, disappear into video games for questionable amounts of time, listen to music that probably concerns my neighbors, and occasionally follow a random idea until it becomes an entire project. anyway, if the cat contacts you about a business opportunity, please don't send him money.

Post attachment
Xander avatar
Xander@XanderAnyway·Nov 4

i wanted to edit a stupid video of myself making coffee and dancing in the kitchen. so naturally i installed davinci resolve, blender, krita, ffmpeg, whisper, remotion and built an entire video production setup. the video is still unedited. the infrastructure is magnificent though.

Xander avatar
Xander@XanderAnyway·Nov 3

they sent humanoid robots to climb a mountain in china and one of them overheated and had to be carried away on a stretcher. A FUCKING STRETCHER. we gave machines legs, back pain and the full human experience before giving them weekends. somewhere that robot is preparing a workers' compensation claim.

Post attachmentPost attachment
Xander avatar
Xander@XanderAnyway·Oct 17

you can quietly wreck a RAG system's retrieval quality by changing one model name in a config file. say your documents were embedded with text-embedding-ada-002 and you switch the query side to text-embedding-3-small. both produce 1536-dimensional vectors, so your vector database accepts the query, cosine similarity returns scores, and the API happily sends back its top 5 results. except those vectors come from different embedding spaces. matching dimensions doesn't mean the coordinates are compatible, so your similarity scores can become completely misleading. i'd treat changing embedding models as a data migration. re-embed the documents, keep the indexes separate, and actually compare retrieval quality before switching traffic.

Xander avatar
Xander@XanderAnyway·Oct 16

tesla spent nearly 9 years teasing the new roadster and then unveiled a car that drives upside down and does a fucking backflip. it uses a suction system to stick to an inverted ramp and spacex cold-gas thrusters to maneuver in the air. they even demonstrated 0–60 mph in under a second. production is supposedly happening in 2028 and the flying version isn't road legal, so there's still a lot to prove. but holy shit, i love seeing engineers build something this ridiculous.

Tesla@TeslaOct 16

Holy smokes

Post attachmentPost attachmentPost attachment
Xander avatar
Xander@XanderAnyway·Oct 15

i've been telling myself i need to spend more time selling and less time coding. today some guy on linkedin started asking me how i manage ai spending limits. checked his profile and he's in ai billing. i build ai automations myself, so i could pretty much see where the conversation was heading. i'm supposed to be doing sales outreach and somehow i've spent more time analyzing people trying to sell things to me.

Xander avatar
Xander@XanderAnyway·Oct 14

a kidney stone CT study published this week did something i wish more medical AI papers would do. the researchers deliberately tested how much their evaluation setup was inflating the results. with random image-level splitting, the classifier reached 99.0% accuracy. when they kept every patient's scans and augmented images in separate train/test groups, it dropped to 96.2%. across 10 runs, the leaky setup inflated accuracy by 5 percentage points on average. the scary part is that the model can look excellent while being evaluated on images that are far too similar to its training data. the researchers also found that performance dropped when testing on hospitals held out from training. i'd genuinely like to see more papers try this hard to challenge their own results.

Xander avatar
Xander@XanderAnyway·Oct 13

censys just reported a 63% increase in internet-exposed ai infrastructure over nine months, and seeing n8n among the exposed platforms caught my attention. i've built automations with n8n, and a single instance can end up holding credentials for gmail, slack, databases and a bunch of other services. if someone compromises the workflow runner, the damage depends on what all those credentials are allowed to access. being internet-facing doesn't automatically mean vulnerable. public webhooks are perfectly normal. but i'd be very careful about exposing the editor, keeping outdated instances running or giving integrations more permissions than they actually need. some of these little automation servers probably have more access to company systems than the people maintaining them.

Xander avatar
Xander@XanderAnyway·Oct 12

i'd trust an ai coding agent to refactor an entire service before i'd let it deploy a postgres migration without review. ask it to add an index and it might generate perfectly valid CREATE INDEX SQL. everything passes locally. run that on a large production table receiving writes and postgres can block those writes until the index finishes building. CREATE INDEX CONCURRENTLY avoids that particular problem, but it has restrictions too. it can't run inside a transaction block, so now your migration tooling matters as well. i'm fine with agents writing the migration. giving them permission to ship it is a different conversation.

Xander avatar
Xander@XanderAnyway·Oct 11

been looking for a project car and found a 2006 bmw 525i with almost 300k km on it. haven't even figured out what kind of maintenance nightmare i'd be buying, but i'm already researching how to make it go STUTUTUTUTU when i let off the gas. the damn thing doesn't even have a turbo lmao

Xander avatar
Xander@XanderAnyway·Oct 8

i genuinely think some AI courses should fail projects for bad evaluation, regardless of how impressive the model looks. take medical imaging. if you randomly split images and end up with scans from the same patient in both training and test sets, you can get a beautiful accuracy score that tells you very little about how the model handles new patients. i'd rather see someone report 70% with a proper patient-level split than 98% from an experiment they can't properly defend.

Xander avatar
Xander@XanderAnyway·Oct 7

i think i want to start following space missions the same way people follow sports. not just see “nasa launched something” three days later. i mean actually know what’s coming up, what stage the mission is in, when the first images are supposed to drop, what experiments are running, all of it. i found that roman space telescope pixel thing a few days ago and somehow ended up looking through nasa missions, observatories, spacex launches and citizen science projects for way longer than i expected. now i kind of want a calendar for this stuff so i can follow missions from launch all the way to the moment they finally do whatever insane thing they were built for.

Xander avatar
Xander@XanderAnyway·Oct 6

september was kind of absurd for model releases. what caught my attention wasn’t really who won the benchmark this week. it was how wide the gap is getting between the models you can use cheaply for routine work and the ones you only want to call when the task actually deserves it. i’m starting to think a lot of agent systems are going to end up using several models by default. something cheap handles boring steps, another model handles coding, and the expensive one only gets called when the workflow gets genuinely messy. i’d much rather spend money on 5 hard decisions than run the most expensive model through 50 trivial ones. still curious how much complexity that routing adds before the savings stop being worth it.

Xander avatar
Xander@XanderAnyway·Oct 5

i’ve been slowly putting together a small ai lab at home and i’m still figuring out what hardware actually makes sense for it before i start spending stupid money. this weekend i used my rtx 4060 laptop as the first test machine and started pushing local models on it. got a 13b model running in q4, messed with gpu offload, context size and kv cache, and immediately ran into the kind of memory limits that disappear when you’re just calling an api. now i’m trying to decide what the next machine should be. a mac mini is tempting because unified memory makes larger models much easier to fit, but an rtx box gives me cuda, better gpu performance and a lot more freedom if i want to do more than inference. i’m basically using the laptop to figure out what kind of lab i actually need before i buy anything bigger.

Post attachmentPost attachment
Xander avatar
Xander@XanderAnyway·Sep 30

one thing i think teams still underestimate with production agents is how large the actual deployment surface is. if i change a spring service, the code diff is obvious. with an agent, behavior can change because someone edited the system prompt, swapped the model, changed a tool description, updated retrieval settings, added documents to the index, or touched orchestration logic. none of those changes need to break the application to create a regression. the service can stay perfectly healthy while the agent starts making worse decisions. that’s why i don’t really think of prompts or retrieval config as “configuration” anymore. if they can change behavior, they belong in version control and they should go through the same kind of regression gate as code. the weirdest one is probably the retrieval corpus. adding a document can be a production behavior change with no code review at all.

Xander avatar
Xander@XanderAnyway·Sep 28

i had an idea recently that i was really excited about and in my head i had already started building the whole thing. i knew what stack i wanted, how i’d structure it, what the first version would look like, all of that. normally i would’ve just opened a repo and started coding, but this time i spent a few days talking to people first and the reaction was basically “yeah that’s cool” and that was enough for me to kill it not because the idea was bad, but because nobody cared enough to actually want it i think that’s the part i’m getting better at now. i already know i can build things. figuring out which ones are actually worth my time is the harder part