Tessa Profile Banner
Tessa Profile
Tessa

@tessybarton

Followers
2,236
Following
938
Media
25
Statuses
288

Pretraining team at @Databricks and @MosaicML . Prev: @NYTimes .

San Francisco CA
Joined October 2011
Don't wanna be here? Send us removal request.
Explore trending content on Musk Viewer
@tessybarton
Tessa
3 years
@jurijfedorov My doula tucked her howitzer into her bandeau bra. She applied milliamps of electricity to my chignon with her jacquard katana causing my piezos to buzz, my servos to actuate my espadrilles, as my thermoresistors began to strafe the sateen at an azimuthal angle in my taffeta...
18
29
683
@tessybarton
Tessa
2 months
Not a Claude in the sky @caseychu9
Tweet media one
6
7
588
@tessybarton
Tessa
1 month
Blazingly fast fashion🔥
@farhatchristina
Christina Farhat
1 month
Tweet media one
234
1K
15K
22
16
432
@tessybarton
Tessa
3 years
@jurijfedorov ... peplum top. A femtosecond later the yakuza burst though the verbena brandishing neodynium pessaries.
10
5
340
@tessybarton
Tessa
1 month
The only purse in the world subject to ITAR restrictions.
Tweet media one
4
18
270
@tessybarton
Tessa
27 days
hey @nvidia now even the press is after you about those graphics drivers
Tweet media one
14
7
268
@tessybarton
Tessa
6 months
LLM evals are a mess! They are noisy, inconsistent, and contradictory. Scaling laws on the other hand have consistently held up to increasing scrutiny. Can we use the reliability of scaling laws to predict the quality of our eval benchmarks?
Tweet media one
9
40
255
@tessybarton
Tessa
2 months
Feeling an uncontrollable urge to lasso and domesticate a Waymo
8
12
207
@tessybarton
Tessa
2 months
Laser focused on the problem of intelligence. Congrats on the PhD @akhil_bagaria ! @RichardSSutton
Tweet media one
2
6
123
@tessybarton
Tessa
2 years
Tweet media one
0
12
118
@tessybarton
Tessa
3 months
i bet a d_model that is a multiple of 128 feels so good for the flashattention kernel
2
3
102
@tessybarton
Tessa
3 months
Go with the mlflow~
Tweet media one
5
1
83
@tessybarton
Tessa
12 days
My fellow Americans, I need your vote: Can I go 2 weeks after the US Presidential Election without finding out who won?
Tweet media one
15
0
83
@tessybarton
Tessa
7 months
TFW the code eval metrics are so good your boss does his hair blue
@jefrankle
Jonathan Frankle
7 months
Tweet media one
5
0
40
1
6
68
@tessybarton
Tessa
7 months
DBRX is a magical new open source LLM! ✨ All the 3am NCCL timeouts in the world couldn’t stop us from training you 💪 🥰
Tweet media one
4
6
53
@tessybarton
Tessa
1 month
The rumors are true. We have an H100. And no, you cannot wear it to China
1
5
52
@tessybarton
Tessa
1 year
ML FASHION WEEK: GPU purse. NVIDIA Fall 2013.
Tweet media one
@ml_hardware
Abhi Venigalla
1 year
🎉 ML FASHION WEEK🎉 Day 2: Scaling Laws, circa 2023. You can buy this one online! Kudos @antimatter15 for the design, and @GoogleDeepMind for the math
Tweet media one
1
1
34
3
1
52
@tessybarton
Tessa
2 years
I keep meeting young people who are not only better than I was at their age but better than *anyone* my age was at their age and it makes me super hopeful!
7
1
51
@tessybarton
Tessa
6 months
All this Celsius and nothing to grind
Tweet media one
7
1
51
@tessybarton
Tessa
1 month
8
2
43
@tessybarton
Tessa
3 months
In Paris for the Olympics, already earning reputation as bit of a crepeist! @caseychu9
Tweet media one
Tweet media two
2
0
41
@tessybarton
Tessa
3 months
Not enough people are talking about how cool Claude Artifacts are
1
0
35
@tessybarton
Tessa
2 years
\textbf{} is a Latex reminder to text your bf
2
1
34
@tessybarton
Tessa
7 months
The best part of DBRX is it’s still a little bit unhinged
@andersonbcdefg
Ben (e/treats)
7 months
me: can i get a recipe for banana cream pie that is bad dbrx-instruct: Sure, here's a recipe for a not-so-delicious banana cream pie: Ingredients: * 1 pre-made pie crust * 3 overripe bananas * 1 cup of mayonnaise * 1/4 cup of mustard * 1/4 cup of vinegar * 1/4 cup of soy sauce
18
15
190
1
0
32
@tessybarton
Tessa
2 years
The existence of neurodivergence implies the existence of neurocurl
2
1
30
@tessybarton
Tessa
2 years
People don’t like ghosting because people want dating to be differentiable. If they know your loss function and the score you gave them they could more easily improve
4
0
29
@tessybarton
Tessa
6 months
Mooses Mooes… MOEs!
Tweet media one
0
0
25
@tessybarton
Tessa
2 years
MANY of the new emoji are bondage related: 🪢 ⛓🗝🥽🧤🪶🚬📿🩺🏳️🧎🏻‍♀️🧎🏼🕯🧊. Has the Unicode Consortium been a naughty Consortium? 😈
1
2
24
@tessybarton
Tessa
2 years
The recession will cause men to prefer women with hearty, stockier body types. Swarthy broads who could swing an axe and unkink a supply chain. Our time has come
1
1
23
@tessybarton
Tessa
2 years
i’m 28 but I still get too excited reading college course catalogues urgh
0
0
23
@tessybarton
Tessa
6 months
We tested out 39 different LLM benchmarks to see how they score a series of models with a progressively larger FLOP budgets. We segmented them according to how monotonically they ranked the models. Here are the well behaved benchmarks:
Tweet media one
1
0
22
@tessybarton
Tessa
2 years
I didn't realize lim inf was different from lim, so my summary above is incorrect, its still cool though:
@Scientific_Bird
Inquisitive Bird
2 years
@tessybarton To be clear, it's not that two consecutive primes are always bounded by 70 million, but only that *at least on occasion* you're guaranteed a relatively small gap between two consecutive primes. (Lim inf vs just lim)
2
1
81
2
1
21
@tessybarton
Tessa
2 years
Stop asking for my New Years’ resolution. It’s 512 by 512 just like last year
1
0
20
@tessybarton
Tessa
2 years
After massive tech layoffs, programmers are returning to their ancestral villages, to a life of subsistence click farming
0
2
21
@tessybarton
Tessa
2 years
Why is it always “raise your standards” or “lower your standards,” never “project your standards into infinite dimensions and find the closest point using the dot product and use its valence”
1
2
20
@tessybarton
Tessa
2 years
Every document is a dating document.
1
1
21
@tessybarton
Tessa
1 year
Me: 'I will share my screen, just gotta close some windows haha' The windows in question:
Tweet media one
0
2
20
@tessybarton
Tessa
3 years
@_RodolfoOcampo "Topological sort was first discovered by Jesus. He was sitting on a hill, thinking about how to sort the people into groups."
Tweet media one
0
3
18
@tessybarton
Tessa
3 years
@jurijfedorov I can keep going I know all the words.
2
0
16
@tessybarton
Tessa
2 years
I do not need another self help book. I need a physical jackboot thug holding a katana to my neck while I code
2
0
16
@tessybarton
Tessa
2 years
What I think of when I hear ‘ill-conditioned Hessian’
Tweet media one
1
1
15
@tessybarton
Tessa
2 years
A guy just called me old because I was regularizing using dropout!
1
0
16
@tessybarton
Tessa
2 years
@MichaelTrazzi I was at Pixar in 2019. I think it’s hard to imagine any film to be fully AI edited. There is a lot of alpha in tiny tweaks by humans to procedural content. An iterative process. AI will absolutely open up studio-quality animation to the masses. But I’m more bullish on teams.
2
0
14
@tessybarton
Tessa
1 year
When horse girls play chess 🐎
2
0
12
@tessybarton
Tessa
6 months
These benchmarks were monotonically related to model scale at some few-shot settings but were unrelated to model scale at other few-shot settings.
Tweet media one
1
0
14
@tessybarton
Tessa
6 months
We trained the 5 models with larger amounts of training data: from 20-500 tokens per parameter. Each model had 3 billion parameters, so the total FLOPS ranged from 2.1e22 to 5.4e23 FLOPS. This FLOP budget was too small to get signal on these benchmarks:
Tweet media one
1
0
13
@tessybarton
Tessa
6 months
Some benchmarks actually got worse with scale. These benchmarks might mislead researchers about what decisions to make with their experimental results:
Tweet media one
1
0
13
@tessybarton
Tessa
2 years
Machine learning pickup lines🧵
1
2
12
@tessybarton
Tessa
3 years
@jurijfedorov The lead yakuza pulled off his ruched voile. Turns out its that she-male mathematician Gauss. His hair is died in the chambray style, and kohl lines his eyes. In a whipstitch he reties the taffeta around his face, cause he must be covered in accordance with the bushido. Showing..
1
0
11
@tessybarton
Tessa
2 years
Boys AREN’T looking for love. They are looking for a DARK PRIESTESS to bring them into the desert, to slice into their ripe arteries and make JELLO SHOTS of their blood.
2
0
11
@tessybarton
Tessa
3 years
@jurijfedorov his face was just a checksum that his teraflops of brain processing had succeeded and he had woven a Higgs boson into the freesias’ damask on the wall. He uses the remote control software parsec to log into his windows machine at home to make sure the boucle covered aileron
1
0
11
@tessybarton
Tessa
2 years
“He’s not just a beta male, he is the entire beta distribution”
0
0
11
@tessybarton
Tessa
3 months
@ArtirKel Training models and training models!
1
0
10
@tessybarton
Tessa
2 years
Visual trends recur, linguistic trends do not, why?
1
0
10
@tessybarton
Tessa
2 years
They make music you like, until you turn 30. Then you have to listen to that music only until you die.
1
0
10
@tessybarton
Tessa
2 years
People who think they can quit social media are the same lockean muttonheads who think they can just ‘not sign’ the social contract
1
0
9
@tessybarton
Tessa
2 years
I am entering my wave era. no, my particle era. no, definitely wave era. jk, particle era.
0
0
9
@tessybarton
Tessa
2 years
The stoic british slang I was expecting to learn here was certainly not everyone (EVERYONE) greeting me with “Hiya”
1
0
9
@tessybarton
Tessa
2 years
@midjourney a dog with the face of friedrich nietzche
3
0
8
@tessybarton
Tessa
2 years
Man is born free but everywhere he is in shades of navy blue or khaki
0
1
9
@tessybarton
Tessa
3 years
Say what you will about ADHD but this one time I saw my mom break her toe and then forget about it and chop down a whole tree
0
0
8
@tessybarton
Tessa
3 years
Vibecamp fun fact party favor for you: the female colon is usually 10 cm longer than the male colon ::::::::::
0
0
8
@tessybarton
Tessa
2 years
@arctanno Jingle (bar)bells
0
0
8
@tessybarton
Tessa
3 years
“I had to go to the dentist. Fortunately I had two anesthetics: I had just discovered books on vector analysis. Second my dentist had a very pretty nurse.”
1
0
8
@tessybarton
Tessa
2 years
I have never clicked a link so quickly
Tweet media one
0
0
8
@tessybarton
Tessa
3 years
I’m worried about all the data that is not on the cloud or owned by corps but is instead in my closet in a stack of increasingly thick laptops
0
0
8
@tessybarton
Tessa
2 years
Are you the Adam optimizer? Cause I am gonna need more than just one moment with you.
0
1
8
@tessybarton
Tessa
2 years
@FreshMangoLassi The Three Body Problem by Cixin Liu is so awesome, the progress of generations is simulated by one of the characters being cryopreserved and woken up at intervals between 1990 and very far in the future. Oh also the first book is kind of bad but 2&3 make it worth it
0
0
8
@tessybarton
Tessa
2 years
I found a box in my parents’ basement just labeled ‘Tessa’s head.’
1
0
7
@tessybarton
Tessa
3 years
How to tell an algorithm is from the 90s. It will have a name like ‘the BLAST Algorithm’. It will use a ‘MATROID’. It’s runtime? GALACTIC.
0
0
7
@tessybarton
Tessa
6 months
@00aarti I like it but I like a lot of bad things
1
0
7
@tessybarton
Tessa
3 months
@b__feldman @AnthropicAI Hi! I was wondering where this photo was taken? I am in Paris and would love to go see this ad.
1
0
6
@tessybarton
Tessa
2 years
The software on Jeff Bezos’ computer should be called bezOS
0
0
6
@tessybarton
Tessa
2 years
The woman of my dreams has a Picasso face, a Renoir body, and a single, solitary, van Gogh ear
0
0
6
@tessybarton
Tessa
1 year
@AgnesCallard @robinhanson Hi! This was me. I'm very glad I got to sit and chat about Pessoa and math!
0
0
6
@tessybarton
Tessa
2 years
Like we may dress like 1975, use nostalgic 90s typography, but nobody goes back to saying ‘golly’, or ‘groovy’, or ‘by Jove’
3
0
6
@tessybarton
Tessa
2 years
@michael_nielsen As a graphics/vision person, the most amazing AI advance is how powerful language has been in improving image understanding. I used to be postmodern, ‘language doesn’t mean a lot’. Turns out language is a powerful technology and results like CLIP and VQGAN prove it.
0
0
6
@tessybarton
Tessa
2 years
@sashachapin If only I could ‘find the time’ 🫠
1
0
6
@tessybarton
Tessa
3 years
Nabisco: National Biscuit Company
1
0
6
@tessybarton
Tessa
2 years
Where can you go to actually fight someone in nyc
5
0
6
@tessybarton
Tessa
1 month
0
0
6
@tessybarton
Tessa
2 years
They have a grocery store aisle just for puddings
0
0
6
@tessybarton
Tessa
2 years
50% of people in NYC have a mildly consumptive smoker’s cough from allergies or long covid and it’s obviously cool and aesthetic but nevertheless alarming
1
0
4
@tessybarton
Tessa
2 years
@Liv_Boeree @willmacaskill The ‘happy birthday as dirge’ is a legitimate infohazard of this episode. Once you hear the song’s true nature you cannot unhear it
2
0
5
@tessybarton
Tessa
2 years
This morning I smelled the puke-scent of NYC and thought: “That’s not so bad.” I need to get out of here
0
0
5
@tessybarton
Tessa
2 years
What people don’t realize about 10x engineers is that if you hire one you also have to hire 3-5 normal people for them to recreationally torment
1
0
5
@tessybarton
Tessa
3 years
@mechanical_monk When I met you at the Airbnb I didn’t know your twitter handle, and my first thought was ‘I probably have nothing to talk about with this normie’ when in fact I have been bookmarking your tweets for months
1
0
5
@tessybarton
Tessa
2 years
@kyzcralotSxaM It gets weirder along weird dimensions that would be unpublishable today, an interesting if not perfect ending
0
0
5
@tessybarton
Tessa
2 years
I lost 45 pounds and kept it off for 2+ years and cannot abide the ludicrous diet advice of naturally thin people
3
0
5