josh
@j_mcgraph
Followers
549
Following
829
Media
57
Statuses
1,375
founding member of @datologyai | PhD student @dsg_uwaterloo | poodle enthusiast
San Francisco, CA
Joined November 2012
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
$GME
• 590532 Tweets
Cohen
• 252724 Tweets
Aziz Yıldırım
• 83722 Tweets
Villa
• 73477 Tweets
GameStop
• 66866 Tweets
Martinez
• 63947 Tweets
Diaz
• 58540 Tweets
Ali Koç
• 58518 Tweets
Gomez
• 48780 Tweets
Roaring Kitty
• 40719 Tweets
Mourinho
• 38569 Tweets
gracie
• 36823 Tweets
सुशील मोदी
• 36421 Tweets
GPT-4o
• 34137 Tweets
Atiku
• 29614 Tweets
$AMC
• 29416 Tweets
LEE KNOW AT GUCCI LONDRA
• 27941 Tweets
Tuberville
• 23094 Tweets
DIAMOND HANDS
• 21307 Tweets
सुशील कुमार मोदी
• 18627 Tweets
Vitor Roque
• 16672 Tweets
#KızılGoncalar
• 15201 Tweets
#AVLLIV
• 14829 Tweets
Aleaga
• 13362 Tweets
Meek
• 13076 Tweets
Lamine
• 12767 Tweets
Bronny
• 12653 Tweets
Assassin's Creed Shadows
• 11502 Tweets
John Fury
• 11204 Tweets
Watkins
• 10198 Tweets
One problem with finding ML talent since the bitter lesson is that ML is now distributed systems and statistics and our university system wasn't setup to produce people who are good at both.
8
14
263
@mattyglesias
Xi is obviously concealed carrying the weights to the tiktok recommender algorithm in there.
2
0
144
@BlueBir75555922
Nothing is perfect but I like this reading list
Grounding yourself in examples is important too!
the book high performance python chapter 8 on concurrency is also great
3
3
60
Shouldn't we be measuring by the number of downloads instead of uploads?
Numbers of public models on
@huggingface
:
-
@Meta
: 689 including MusicGen, Galactica, Wav2Vec, RoBERTa,... -
-
@Google
: 591 including BERT, Flan, T5, mobilnet,... -
-
@Microsoft
: 252 including DialoGPT, BioGPT, layoutLM, uniML,
22
163
755
6
5
40
@typedfemale
This is the TikTok length 🍿 I was looking for. The vector DBs from not DB people is taking too long.
1
0
34
@thenabanana
Tweetable proof sketch
1. The product of two normally distributed zero mean variables is a central chi square distribution
2. Dot product is a sum of these products
3. Expectation is zero by LOE
4. Variance goes down as you add terms (dimensions) to the sum
2
0
32
Excited to share our recent work, AriGI. We had our CEO
@arimorcos
manually complete BIG-bench, the LSAT, HumanEval... (1/n)
1
0
31
One problem with finding ML talent since the bitter lesson is that ML is now distributed systems and statistics and our university system wasn't setup to produce people who are good at both.
8
14
263
0
0
28
@IDoTheThinking
My commute from SF to RWC takes 1hr 45min if I take Caltrain+Muni, and 45 min if I drive. And I have to plan when I leave around the hourly departures.
1
0
25
@whyvert
60% of the time for a binary prediction problem is terrible performance. This isn’t doing much better than flipping a coin.
I wouldn’t take much away from this
0
1
24
@telmudic
Rationalists are a status improvement on the akshually meme without any substantive difference
0
0
20
The key to compute efficiency? training data
The key to fairness? training data
The key to better models? training data
training data, training data, training data, training data
0
1
20
Who’s going to help me cofound a thrift shop but we only buy defunct startup swag? This is the tourist attraction SF is missing
2
1
18
@TrungTPhan
Anons there is nothing stopping you from just walking up to the cute ones and asking them on a date. Don’t fall for this psyop
1
0
16
This is real AI risk, not some convoluted alignment scenario out of a sci-fi movie.
DAIR is a refreshing break from MIRI larpers.
I can't get over that the maximum sentence for "shoplifting snacks" could have been 25 years.
What is the maximum sentence for governments that fail their people to the point that they'd need to shoplift snacks?
(from a while back by
@kharijohnson
)
3
26
90
1
4
16
@Noahpinion
It’s hard to see the UCSD news in the feed while also seeing CS PhDs talk about how hard it is to find an academic job 😭
0
0
15
@yoavgo
Regardless of however they feel about you, your point is not even a little audacious.
Knowing your target and what’s around it is literally like the most basic part of using any weapon as safely as possible. Being able to do that better is a good thing.
1
0
11
@john__allard
When you find the functions that will bring about AGI but the dependency graph lookin like a 19 year old boys closet
1
0
12
@grumpysavior
@Aella_Girl
I feel like people are less willing to express their vulnerability around wanting friends than a romantic partner
Or bumble BFF is just bad because the app is buggy
0
0
13
@zestular
As with using adderall instead of coffee, the modern tech bro needs a stronger downer than alcohol.
0
1
11
If you're a Josh in AI looking to join more Josh's in AI, we have the place for you!
Some personal news: yesterday was my first day at
@datologyai
! I will be working on what I consider to be the most interesting problem in data engineering: curating training datasets for machine learning models.
21
2
176
4
0
12
Sometimes I get irrationally mad that we force kids to take precalc. We should replace that formula memorization garbage with a year of epsilon delta definition of limits, set theory and basic topology of Rn
1
1
10
Super grateful today for Alireza for being such a great colleague and
@thodrek
,
@ihabilyas
for being wonderful mentors. Excited for my first paper, too!
#HoloClean
#HoloDetect
#SIGMOD2019
#SIGMOD2019
Is weak supervision the answer to data cleaning? Alireza and
@joshuawmcgrath
show how learning data augmentation policies from unlabeled data (without supervision) helps you train deep learning models for error detection (w.
@ihabilyas
)
#HoloClean
#HoloDetect
2
9
40
0
2
10
@ronawang
- Doing systems level SWE stuff
- HPC optimizations for distributed training jobs especially with enroot/kube
- Taking on linux admin-y tasks
A lot of this is easier to learn at work than just by reading! Helpful to see the problems companies run into by being at one
0
0
10
@cauchyfriend
It’s stunning to see such a large company use just 100 samples to estimate such a business critical statistic as DAU.
This is not an endorsement of Musk. This is a critique of how twitter is spending money to understand their business’ environment.
1
0
8
@lorisdanto
I heard instead of having these, they make the kids debug shape errors in PyTorch code
0
0
9
@tomcritchlow
Database is just the colloquial term for a database management system. Excel/sheets are databases in that you can store data in them.But they don’t have the same things a DBMS does. Examples below.
1
0
9
Excited that we're finally out of stealth! If you want to work on *efficiently* scaling models, join us!
Hello world! We are incredibly excited to come out of stealth today to help make better data accessible to everyone, automatically.
Hear from our founders about our mission and vision for DatologyAI:
0
9
65
0
0
8
@Aella_Girl
I think we're still farther than most people realize, there's no evidence that scale is all you need
Alignment will get fixed along the way
Prediction markets are about as efficient at predicting the future as regular markets (a lagging indicator)
About the raging debate regarding the significance of recent progress in AI, it may be useful to (re)state a few obvious facts:
(0) there is no such thing as AGI. Reaching "Human Level AI" may be a useful goal, but even humans are specialized.
1/N
65
574
3K
1
0
8
0
0
7
People are bad at cleaning data, you need to do it with a model!
0
0
7
@yoavgo
We should definitely go back fleets of B-17s burning entire cities to the ground. This whole JDAM thing saves far too many lives.
I think most people live in an ahistorical fantasy land.
1
0
7
@tomcritchlow
The most notable difference is that a DBMS almost always has a component called query optimizer. This will look at the structure of your query and selects which algorithm to use to execute that query.
1
0
6
Realizing that I must like dogs more than people because I absolutely would not pick up anybody else’s shit but I literally congratulate my dog while I pick hers up.
1
0
7
@leavittron
This is like a more general form of the bitter lesson
The inductive bias in the model is less important than compute+data
Transformers efficiently utilize the GPU, so we’ve hit diminishing returns on compute
For data, we can do better than unexamined piles of web text
0
0
7
@IDoTheThinking
I also find it hilarious to think that anybody visiting SF won’t have seen all the articles about a city cleanup
0
0
7
being a waterloo prof/grad student is goated because the undergrads are cracked waterloo kids
2
0
7
@chairsign
@growing_daniel
Podcast —> vibe w author —> buy book
Better than any Goodreads or friend rec I’ve ever had
1
0
6
@thenabanana
I know there’s one in the lecture notes that sets one of the variables to the zero vector but that feels less satisfying
0
0
5
Thankful for
@ihabilyas
and
@thodrek
for their mentorship,
@Mina_Farid
and
@TheBiggOh
for being amazing colleagues. I've benefited immensely from being able to work with some amazing people who happen to be non-Americans.
I’m a US citizen and have worked here my whole life.
Along the way, I have studied with 5 research mentors. Each has been incredibly thoughtful, dedicated, interesting, and world-class at what they do.
They also all were immigrants. What are we doing ...
7
40
626
1
0
6
I really want a demographic study of people simultaneously believe
“AI is going to cause inequality”
And
“Not every student needs algebra 2”
Editorial: Not every student needs algebra 2. UC should be flexible on math requirement (via
@latimesopinion
)
163
4
40
0
2
4
@dezren39
@typedfemale
Vector indices are a real value. But exposing them as raw constructs is like if relational DBs asked you to know about the B+ tree or something.
People want robust text retrieval, not a wrapper around HNSW.
1
0
6
@davidchalmers42
Current systems won’t get us there — but the next things we’ll find are probably going to be just as simple.
0
0
5
@datologyai
is all about being in the Pareto frontier of compute per nat
"To be frontier you first need to be be pareto-frontier". ~ First law of LLM training. 😃
21
39
413
0
0
5
@HotForMoot
Then explain this
The laziest grossest thing a man can do is have a beard or long hair
83
7
256
0
0
4
@Aella_Girl
I grew up Jewish and I thought this and the Easter Bunny were so odd. Also I don't really know of another religion that does it? Like just lie for.... fun?
6
0
5
@tszzl
lean startup is non-software ppl trying to apply toyota methods to software. the worst part is that there are relevant TPS principles that don't get disseminated because so many software ppl have an aversion to the branding
0
0
5
Everybody come to the chipotle in chan to support robotics!!!!
#beepboop
#doitforthechildren
#OMGrobots
0
2
5
@ConradBastable
I’m not sure about E&M, but I’m here to advocate for government mandated Civics and Linear Algebra.
1
0
5
@amasad
@robcobbable
I’m guessing he’s talking about Gödel numbers?
He basically embedded every well formed formula (read: statement in a formal language) into the natural numbers. This embedding is a key tool in proving the Gödel incompleteness theorems.
1
0
5
@inerati
“They’re pretty trad” referencing something like a belief in the nuclear family always gets me. Nuclear families is only trad on like a 2 generation time scale.
0
0
5
@mylordcod
@tszzl
@telmudic
I’m just agreeing with tel that it’s a bad grift.
Good grift swindles the high impact population. Bad grift just swindles people who didn’t care in the first place.
1
0
4
@leroidecapitale
@BroductManager
@vgr
His guide is written to/for people skeptical of the BSE premise. Bullies are doing these things naturally for better or worse. The skills listed in BSE are ingrained in them regardless of whether or not their target has defenses.
0
0
4
@bayeslord
We’re moving back to seeing “move fast and break things” as good/acceptable as we should.
Just try to design in some reversibility, and have humility when you’re wrong.
fb didn’t have the latter.
2
0
5
@thecaptain_nemo
For most people having an accurate mental map of the world won’t (at least quickly) effect on their wellbeing and so they don’t try to acquire one
0
0
4
@DistractedAnna
Haven’t we all watched people stan their side too hard, stay in bum relationships, and believed the loudest person in the room by that time ???
1
0
4
@mattparlmer
Strategic bombing has been a euphemism for killing civilians since we started doing it
0
1
4
@carajanefoster
ya biology does explain why people die. I don't think that will be on the test though.
1
0
4
“LLMs aren’t as efficient at learning as humans”
My brother in christ, we have spent a millennia of compute hours designing curricula for humans.
0
0
4