Made in a few hours - interactive visualization of millions of glucose measurements from 150k people:
(I'm aiming to work at the intersection of AI & large language models and interactive visualizations such as this. DM me if hiring / know a good fit!)
Please stop what you are doing and for the sake of millions of people who died from preventable causes (the
@NIH
should never have funded this research) read this:
Thank you
@Ayjchan
on behalf of our family in Wuhan that we smuggle paxlovid for, 2x90yo
Looking for a full-time role in AI / large language models (research/eng potentially focused on health).
Know anyone to chat with?
Please RT/forward my CV ()/DM/connect me. I've built large language models, have 1000+ citations (NeurIPS, ICML, AISTATS).
Finished Grade 23!!! Just got a real PhD after a virtual defense 😆
Video of defense:
Thesis:
Write-up: issues/tips for doing it on Zoom
💌 Started: post-thesis mailing list
WOW I feel like I've seen the future of deep learning + autodiff.
Are you using jax yet??
It's a breeze compared to tensorflow. AND 3x speedup over pytorch!
Used a VAE to benchmark:
Shouts out to
@SingularMattrix
& co for all your work on this!!!
So very excited about this line of work!
Defines multiple
#causalinference
. Solves it using mutual information regularization (secret sauce: variational inference tricks!)
Like the matrix 💊 can't unsee it. Now potential applications everywhere
what an honor to hear professor
@colinraffel
speak at
@nyuniversity
about the issues with language models such as ChatGPT based on his work as professor at
@unccs
and as faculty researcher
@huggingface
. mind blown.
#chatgpt
may be doing okay but i'm on team colin for his version.
Pretty wild: 38 million people's income on a single chart (2 decades' data!!)
Check it out:
Thank you
@recifs
and
@mbostock
for the tips on asynchronous JavaScript!! This new
@observablehq
framework +
@duckdb
enables really, really cool new applications.
Woah this is much nicer thanks to
@recifs
's wizardry with Observable Plot!
Updated the Census visuals:
(& officially sold on the
@duckdb
->
@getdbt
->
@observablehq
workflow; scalable to millions of peoples'/households' data & rendered gorgeously ^_^)
had a dope time presenting at
@LightningAI
- really recommend it! def the most diverse audience i have ever given a talk to.
talked about some cool work at
@onefact_org
in health care AI :)
demo - how race to the bottom of wrapper companies works:
Insane remote teaching setup for
@unitartu
+
@kordinglab
+
@onefact_org
's first lecture on . Dm me if you wanna audit / want the livestream link to Twitch! Students exclusively use ChatGPT and CoPilot and tools like
@lexdotpage
- all open source on github.
PROUD of my 2017
@CitiBikeNYC
stats!
Top speed 25.9 km/h
Avg speed 14.6 km/h
Money saved $874.5
Distance 1325.8 km
330 Trips
== best investment I made
#OpenSource
'd my data and code:
(Easy to port to your data! Commuters – share yours?)
🚴♂️💻📈
Made a map of 100+ hospitals in the NYC area: alongside quality metrics.
Next up: adding price information from our database of 4000+ hospitals.
Ping me if you want to help with this or donate to broaden this to all hospitals! (501c3, tax-deductible)
Top Innovation 2017 to come out of
@DSI_Columbia
is
"High-Dimensional Ping Pong"
Insanely challenging and fun
#pingpong
(Here we Play on top of an optimization whiteboard with a $10 net, to practice Passive Knowledge Absorption)
Embarrassed to admit tricks in still work,
but now I'm addicted to my GITHUB FEED 😬
(the last thing I can't permaban for work)
Pro: readmes very high in info per bits
Con: compulsively refresh github, commits/stars/forks flow by slowly = less dopamine
so stoked for
@duckdb
con ---> QUACK THIS WAY <--- in a month!
fun to beta test
@Apple
's and
@awnihannun
's hard work on mlx-swift integration () and
@maxxen_
's hard work on vector search extensions for online analytics databases!
DuckDB is my new best friend for analytics. It's so fast, makes it easy to manipulate dataframes, ingests anything, and exports it to whatever form I want.
Parquet, Arrow dataframes, and DuckDB are the future for analytics.
New work! Supervised topic models work well for predicting molecular substructure from mass spectra - especially with limited data
Open access:
Open source:
Super stoked for the follow-ups!!!!
Wow this is exactly what I want! An AI app to harvest my most vulnerable moments and monetize them to remunerate shareholders and uphold fiduciary responsibility and triple ARR. Thanks so much
@ariannahuff
and
@sama
! I couldn’t be sold a brand behavioral identity as efficiently
Today Sam Altman and I published a piece in TIME sharing our vision for how AI-driven personalized behavior change can transform healthcare and announcing the launch of Thrive AI Health, a new company funded by the OpenAI Startup Fund and Thrive Global, which will be devoted to
Woah this is much nicer thanks to
@recifs
's wizardry with Observable Plot!
Updated the Census visuals:
(& officially sold on the
@duckdb
->
@getdbt
->
@observablehq
workflow; scalable to millions of peoples'/households' data & rendered gorgeously ^_^)
Fun trying out
@observablehq
's new visualization framework for the web using
@uscensusbureau
data.
Demo + code with American Community Survey data:
Framework:
Thank you
@mbostock
and
@recifs
and & co for open sourcing this!
Psyched to have helped a bit on this! ++ peep the funnest machine learning emoji 😂
Deep language models can predict patient readmission from clinical notes.
Also don't miss concurrent work from
@Emily_Alsentzer
:)
Which companies / teams are working on this?
I'm actively looking for a full-time role to help (have helped build large language models for broad use cases in health).
DMs open or email me: altosaar
@alumni
.princeton.edu :) it's an exciting time!!
Some stunning progress in medical
#AI
"The convergence of unprecedented, massive computing power and self-supervised learning of a ginormous body of human-derived data has laid the foundation for medical applications that were not previously within reach."
Musicians becoming machine learners->useful tools for creativity!
Thibeault trained neural nets to predict guitar effects and the results are impressive:
Bonus: crazy new dataset of Moog sounds
Data/code/blog:
So stoked to give a talk at
@duckdb
's annual conference called "DUCKCON" in a few weeks in Seattle! Hit me up if you'll be in the area:
Only plan is to go see the Bullitt center for our sauna financing and drive to
@esalen
Weekend: visualizing all 850k real estates in NYC :)
What should I link this with besides sauna access/ ?
Thank you rockstar teams @
@protomaps
@observablehq
@bdon
for enabling this, all open source + free to reuse! Please RT!
Wow
@Kyle_L_Wiggers
must be one of the few journalists that reads the arXiv every day 😮👀 (our work on ClinicalBERT got covered immediately after hitting the arXiv! )
🙇♂️ humbling to meet
@hfmuehleisen
,
@mraasveldt
,
@__AlexMonahan__
in-person and get to deliver in order to "quack this way" at
#duckcon
. Rare to talk about HIPAA, capital stacks, 2 petabytes, AI/RAG, cochleograms!
.
@porterairlines
Still stranded in NYC (here at 8am, now 2pm and you have delayed my flight 6 times). No Christmas dinner for me. 🎄😥
#MerryChristmasEve
@KyleCranmer
@blei_lab
Thank you so much
@KyleCranmer
!!! Couldn't have done it without you and your labs' support, and I'm pumped to continue our work together 🙏😅
Want to reliably compare pre-trained representations for a downstream task?
Proud to have been a small part of work to address this question led by
@wfwhitney
How many papers come with code that runs on both
@jax
AND
@pytorch
??? 🦠🔬
New paper! We propose to measure the quality of learned representations using the complexity of finding a nearly-optimal predictor on a downstream task.
Blog:
Paper:
Library:
$70-90k/month is the typical going rate for these billboards. Excited to build
@paylesshealth
to make these million dollar a year investments less appealing to hospitals.
@harpriiya
Do you remember the article? That’s a great quote. I was lucky to work at
@GoogleDeepMind
when young and see the same concept, realized thru golden handcuffs and a Byzantine leveling system for performance reviews :( and other dopamine induction and behavioral reinforcement
I need help:
building the next version of this map to visualize the price of health care in underserved communities.
Know anyone who wants to help?
Send em this/have them DM me.
This slide from a talk I gave last year at the Open Visualization Summit hosted by
@openjsf
STOKED about this chapter of my thesis! Optimal recall + food (content) recommender systems
✅ map of science (ok, ) according to how people actually read science
✅ blog post by the incred
@olaya_gayol
-
📜
To GOOG friends: I feel guilty after not seeing ads; for double dipping into your P&L; I know you have family. I can’t just read the AI result and leave, burning your electricity without viewing the ads that help pay for it so I scroll past the AI result and click ads. For you.
Hey
#NYC
peeps - need help with understanding gentrification:
Select areas you lived in '05–'22, see whether the income distribution and types of work people do aligns with your experience of seeing a neighborhood change!
What should we look at next?
Love this
#Miyazaki
clip from Nausicaä of the Valley of the Wind.
Helps practice loving-kindness meditation and visualize the tentacles–the holy appendages–of compassion in suffering.
h/t
@TaraBrach
Life path optimization via stochastic gradient ascent
Reduce the variance (increasing the accuracy) of your next step in life by 'collecting datapoints' from people who have done what you want to do.
#machinelearningkoans
🙏💫 h/t
@cassandraxia
's talk @ :
panama canal is man-made, &
@TWSemicon
and
@nvidia
have man-made supply side bottlenecks :)
-
@WhiteHouse
did the CHIPS act () and it's going well
- created thru lots of input from AI safety/EA orgs
- i deprogrammed some high schoolers about this shit
Pro large language model interface = Jupyter AI!!
Highly recommend trying this, feels magical and way better than the
@OpenAI
ChatGPT interface or the
@AnthropicAI
interface I've been using for the past years.
Wrote a mini quickstart here in 3 steps:
1. Install Jupyter AI:
.
@reggiewatts
you'll dig this room-size looper my brother built (u have to jump to use it; partial insp your 2880 setup). Check it if you're ever in Toronto!!
My teams (, , ; 850+ people total) use
@zulip
and agree. It's better in every way.
A *homework* assignment enabled by Zulip -- visualize the class' team chat logs with help from GPT:
Statutory holiday service project - income distributions in
#nyc
vary across time, space, and race.
Does this data align with your anecdata & lived experience in these neighborhoods?
Need feedback/thoughts - RT/forward to a friend that might care!
Rare sneak peek into future of health, wellness, wealth — and longevity:
Highly recommend.
Set aside a coffee for this (especially the audio version), especially if you can’t gawk at enough Huberman / Attia / Ferriss / Johnson dark triad content 🙇♂️🙇♂️
I have a new project called Care Culture, a newsletter and membership community focused on mental health, wellness subcultures, and integrative practice.
my first post is live! I document my field trip to Vitalia, the longevity-themed city in Honduras.
Does this use query-key-value caching? Or other methods - in health care
@onefact_org
we need to consider the downstream impacts different methods can have (building a benchmark with
@duckdb
and
@kuzudb
and
@neo4j
for RAG slowly :)
#MachineLearningKoan
In (in silico) reinforcement learning, we use a model of the world to learn a model to control the world. Why can't we just use the model of the world in the first place? 🤖🏴
h/t
@sirrice
Need feedback:
This visualization - all 106,000+ prices at a hospital in NYC. Crazy (!!) variation by maximum and minimum negotiated rate.
This is NOT an efficient market (y=x).
Open to collaborate on this if anyone is interested; dms open!
Can Transformers extrapolate from short training sequences to long ones? Our new work shows that they display surprising “length generalization” capabilities on many algorithmic tasks: addition, multiplication, and even in-context simulation of SGD!
Yeah,
@othership_us
is wild - my partner and I attended their first ketamine event in NYC with
@JoinCurio
. Cray cray to see social wellness and culture get monetized :) and I think this is a much needed replacement for night drinking / unhealthy stuff. Way to go!