What happens when you turn a designer into an interpretability researcher? They spend hours staring at feature activations in SVG code to see if LLMs actually understand SVGs. It turns out – yes~ We found that semantic concepts transfer across text, ASCII, and SVG:

made it to Seattle, time to load from this box to this truck then to our house

Finally making the move to Seattle! Bye Dallas!

reinforcement learning is hard and complicated and maybe we just should all agree to stop doing it

ok so i got Jupyter pilled for months and then my mentor told me they don’t Jupyter anything and just raw dog scripts so… no im on the other side of the fence

jolios torng@tarngerine.bsky.social · 2y ago

i know I should be using jupyter for everything but I can’t get over the mental barrier and I’m just raw dogging the python console in terminal which is weird bc I’m always screaming about how there needs to be a Jupyter for software design a lesson somewhere here

rly hard to focus on just learning research rn,,, product quality work nerd sniping me

pretty wild to me that apple intelligence hasn’t solved the sms scam problem… these urls are so easy to classify and throw up warnings about

still in the market for a honk shoo cap . they don’t make em like they used to

i did a colonoscopy without anesthesia/meds, i don’t recommend it, bc it hurts like hell (not in the butt but in the abdomen) but I got to see the live video stream and my GI going ham on the controller like she was playing mario party

Erica Joy Astrella@ericajoy.astrel.la · 2y ago

hey friends, there's been a significant increase in colon cancer in younger people in the last 5 years. please, if you're 30-45 and can convince your doctor of it, get a colonoscopy. if not, ask them to give you at least a FIT test. take care of you! kthxbye

insulating some pipes in the crawl space

brain fried from 48 hours of transcript reading, RL derisk runs, and pipeline errors ahh

i know I should be using jupyter for everything but I can’t get over the mental barrier and I’m just raw dogging the python console in terminal which is weird bc I’m always screaming about how there needs to be a Jupyter for software design a lesson somewhere here

love to review preference data collected over 4 weeks only to realize 1/6 of the data is invalid due to a tailwind issue!!! my anti tailwind bias continues…

started playing in the local riichi mahjong leagues first season, did pretty well today! up +72.8 on 3 games!

the quality of your internal tools is a reflection of how you treat yourself and therefore your users

another night of “is this python module broken, slow, or just not working in this cluster or dev box?”

training the Next Big Model with human preference data for RLHF but what happens if their preferences are WRONG in my humble opinion

back at work in the new year and literally today this is all I’ve done

jolios torng@tarngerine.bsky.social · 2y ago

so far most of my time as an ai research engineer is either spent fighting with bespoke tooling not working and devboxes, or reading pages and pages of transcripts to make sure the data is good

my friend played valorant w pubs and they found out how old he was and started calling him “unc” i fear this is coming for me in my professional life too

so far most of my time as an ai research engineer is either spent fighting with bespoke tooling not working and devboxes, or reading pages and pages of transcripts to make sure the data is good

i have started a job where i have no idea whats going on 90% of the time it’s rare to have an opportunity to be a beginner again — i am doing a residency as a research engineer at anthropic

time for a token-optimized but still human proofreadable js framework

Simon Willison@simonwillison.net · 2y ago

I have a theory that AI-assisted programming reduces the value of abstractions like React a little, since you don't need to optimize as much for avoiding tedious state synchronization code (tedious code is less tedious now) Given that, I wonder what the best JS libraries are for building with LLMs?

in my recent round of interviews i had to unlearn the determinism of traditional engineering and embrace statistical and probabilistic problem thinking to pass the coding screens. was a big hurdle and very humbled but it explains a lot

Technology Connections@techconnectify.bsky.social · 2y ago

And that's the the crux of the problem. When my computer crashes, it stops working. When there's a render error, I can see it. Hell, if my heat stops working - my heat stopped working. LLMs appear to "work" all the time, and the failure mode is only experienced later as real-life consequences.

what r some news/media you happily subscribe to? trying to use my educational budget at work before EOY

getting more in the hang of blender!

i accidentally upgraded to ios 18 and now it’s running at 10fps

blender still driving me nuts but some progress i miss rhino

hmmmm

i think I’m actually too old to be good at valorant

kiss of life in dallas!!!