Mauve 👁💜 @mauve@mastodon.mauve.moe

Starting my first day working on #mapeo with Awana Digital.

Gonna be responsible for making the p2p and data side of things reliable and to help develop and deploy features for folks in the field. 🥰

Here are the slides for my presentation. Gonna share the recording later this week.

https://blog.mauve.moe/slides/neural-spatial-curves/

Overall it went okay and I got to show off my ignorance and how I've been slowly learning how to do this stuff. 🤪

Move over "Millenial pause", we've got the "AI assistant processing delay".

IF you're in #Ottawa and into #ai / #LLM stuff, check out tonight's #AITinkerers event: https://lu.ma/3clyvip5

I'm going to be talking about autoencoders and some intuition I learned trying to make a simple model.

Are all those resin dodads going to be our geological era's equivalent to amber?

Fediblock lists are fun and all but I wanna dig around in postgres and see who's in my follow graph poisoning the noosphere. Secure Scuttlebutt was great because I could see exactly how assholes would get access to my whereabouts. 🤷

I'm thankful for the snow cleaning folks when I have to brave the roads but it's hard to stay positive when they're hauling snow at 3 am 🥲

My little spatial curve model is up to over 3k dense neurons with six hidden layers. It takes me 10 mins to run 300 epocs and I think I might triple that now that I've started messing with different learning rates.

Cyberpunk garf: Bitfield

"I hate cache misses"

I've been into egregious usages of enoji lately. I think it's part of my natural urge to cringemax. This is one of my faves lately: 👉😎👈

Womp womp, still can't get tensorflow-rocm to work on my machine 🤷 IT's unfortunate that it's hard to get GPU acceleration working with AMD for tensorflow 😿

Playing Dragon Age Veilguard and the "fantasy name per sentance" density is pretty high 🤪 It's like every other sentance they have to namedrop somebody.

I think I'm warming up to using local LLMs for helping me code. I don't trust them for large scale code generation but it's started saving me some time in looking up syntax and docs via DuckDuckGo. I still need to look up docs for most of the work I do but not so much for common use cases you'd find on stack overflow.

The second last layer was set to relu which messed everything up.