nhlism.bsky.social
Senior ML at 🍎
Had fun at Alexa Prize 2017.
Interested in NLP and bounding boxes around things.
Love less attention in text and more attention in vision.
OMSCS 2025(?)
Recurve Archery and weird projects
🇨🇿
24 posts
75 followers
379 following
Regular Contributor
Conversation Starter
comment in response to
post
maybe more mental health oriented analysis might be useful?
comment in response to
post
I reverse engineered the dayone2 sqlite database to access my journal entries 😬
comment in response to
post
Other tasks planned are LLM generated summaries of months and years (tbh still squeamish about sending my entire journal data to openAI but also curious) and statistics on human mentions
comment in response to
post
Using scores from ModernBERT tuned on sentiment analysis. It's definitely misleading because it the score is the model confidence in the class instead of magnitude of my joy but it makes pretty plots. I wonder if we can turn this into a "rate my day" regression task to give a score between 0 and 1
comment in response to
post
They definitely captured the spirit of the book!
comment in response to
post
Leto Artreides role in Dune 4 confirmed
comment in response to
post
I agree that 80% correctness used to be unacceptable in software
comment in response to
post
One highlight was losing interpretability when moving from symbolic AI to statistical models - I vaguely remember debugging a RDF based QA system and the biggest issue was almost always out of date / missing knowledge. So yeah I almost always knew what went wrong lol
comment in response to
post
huh literally just heard of:
Spatially adaptive Computation time
Depth adaptive transformer
comment in response to
post
point is to estimate difficulty and give model a constraint to try to fit into
comment in response to
post
Click on notifications and then settings (cog on top right)? That being said I don't have enough followers to verify it works lol
comment in response to
post
If it can repeatedly attend over the image tokens, yes. Maybe vision models need their “think step by step” moment