rapha gontijo lopes @rapha_gl Twitter profile

Pinned Tweet

rapha gontijo lopes

@rapha_gl

1 year

we’ve been having so much making ChatGPT (and our dms are open)

William Fedus

@LiamFedus

1 year

Was having so much fun, I forgot to Tweet about it: I joined OpenAI! I’m continuously awed by the people, the technology, and the ambition. We’re just scratching the surface with ChatGPT. If you’re interested, get in touch!

38

1K

2

4

80

Last Seen Profiles

@Andrew640548

@EdwardsNan62076

@6cbw_

@RWUHockey

@_2voi_

@cloverwonu_

@berrysmss

@kyeomiecute_

@official_artms

@ThickneySpears

@rurineco

@gcnmas

@DPlanet_X

@coachfred1

@ManuelIvnReyes

@SignalFire

@JohnMMartyn

@pharie

@WorldLacrosse

@yazelnut

@Claramexico

@Thy2792

@ye4us

@web3carnival

@KarinaMicieli

@Udin1224370690

@LabZeroGames

@BeenWealthy

@TheSeattleOne

@yo_jujuzin

@TheCarlyMurray

@haramhartmut

@RotaryES

@MissEvaStyles

@ppimirl

@MissBriawna

rapha gontijo lopes

@rapha_gl

4 years

Have you ever asked yourself "how did research get done before LateX?!" Oh boy do I have the thread for you… (1/?)

47

1K

4K

rapha gontijo lopes

@rapha_gl

5 years

My 1st @GoogleAI Residency paper is finally on arxiv! We train a powerful generative model of fonts as SVG instead of pixels. This highly structured format enables manipulation of font styles and style transfer between characters at arbitrary scales! 👉🏽

13

249

965

rapha gontijo lopes

@rapha_gl

2 years

@hardmaru

2

29

855

rapha gontijo lopes

@rapha_gl

5 months

a year ago tonight we were placing bets on how many total users we’d get by sunday 20k, 80k, 250k… i jokingly said “8B”. little did we know…

Sam Altman

@sama

5 months

a year ago tonight we were probably just sitting around the office putting the finishing touches on chatgpt before the next morning’s launch. what a year it’s been…

2K

3K

45K

11

15

803

rapha gontijo lopes

@rapha_gl

6 months

OpenAI is nothing without its people.

10

30

516

rapha gontijo lopes

@rapha_gl

2 years

despite scaling laws, even our largest LLMs still behave like system-1 machines. in order to unlock qualitatively better generation, we're gonna need something new here are some thoughts on ✨the search for higher-level reasoning✨ happening in the field

11

57

408

rapha gontijo lopes

@rapha_gl

4 years

These are scans of Felix Klein’s “Riemann Surfaces” (1891/92), more than 450 years after the Gutenberg printing press was invented. The more I dig in this rabbit hole, the more I find the prevalence (and beauty) of hand drawing and art in western science communication.

5

52

389

rapha gontijo lopes

@rapha_gl

6 months

💜

Sam Altman

@sama

6 months

i love the openai team so much

5K

4K

73K

6

11

394

rapha gontijo lopes

@rapha_gl

2 years

We all know NNs are super expressive, so why is it that models seem to make similar mistakes? In our new work, we investigate how training methodology⚙️ affects learned representations🧠 and prediction behavior🏃🏽‍♀️ 📄 🧵 Read on for the surprises we found!

6

77

359

rapha gontijo lopes

@rapha_gl

4 years

It was a laborious (and expensive) process, which explains why many works had to be hand drawn. Here’s Leonardo da Vinci’s “Human Body Sketches” (c. 1510)

5

33

314

rapha gontijo lopes

@rapha_gl

4 years

You know what the kicker is? All of these difficulties did not stop authors from making BEAUTIFUL figures. Here’s Filippo Calandri’s Aritmetica printed in Florence, 1491/92. That’s right: Fourteen-ninety-two.

3

23

304

rapha gontijo lopes

@rapha_gl

4 years

It also explains why even into the 20th century we see scientific works being partially (or fully) handwritten, like Richard Feynman’s PhD thesis (1942)

2

23

274

rapha gontijo lopes

@rapha_gl

4 years

Treviso Arithmetic (1478, left) was the first printed math book. Only 500 copied were made and only 2 survived time. Aritmetica mercantile (1484, right) was typeset in a similar fashion. These texts look visually much simpler, but they were still incredibly complex to produce.

3

25

262

rapha gontijo lopes

@rapha_gl

1 year

another day of work at the paperclip factory

4

6

265

rapha gontijo lopes

@rapha_gl

4 years

Check out our new paper “Affinity and Diversity: Quantifying Mechanisms of Data Augmentation” Despite being crucial to the performance of SOTA models, we know little about why Data Augmentation helps… until now! 👉🏽 Thread👇🏽

1

61

246

rapha gontijo lopes

@rapha_gl

4 years

Mathematical works cost extra (what used to be called “penalty copy”) because all these symbols required much more thought, time, and metal to get right.

3

19

232

rapha gontijo lopes

@rapha_gl

2 years

im one of few with access to all high-end image generation models out there - Imagen () - Parti () [new] - DALL-E 2 () - midjourney ( @midjourney )

25

19

222

rapha gontijo lopes

@rapha_gl

4 years

Some graphics I made on how to stay safe from COVID while protesting. Give them 0 reason to blame the COVID spike on the essential job of fighting for justice!

3

113

218

rapha gontijo lopes

@rapha_gl

4 years

I’ve seen many papers that are just… ugly. And that makes people not want to read your work! It may sound shallow but communication is hard as it is. Your figures should make it _easier_ to convey your ideas and findings. Not create another barrier that needs to be deciphered.

4

25

218

rapha gontijo lopes

@rapha_gl

4 years

Of course, before the printing press, books simply had to be transcribed (and illustrated) by hand. Fig: Ashmole Manuscript, 13th century.

4

12

206

rapha gontijo lopes

@rapha_gl

4 years

And here’s some beautiful typography from 1926’s Quantisierung als Eigenwertproblem.

2

15

208

rapha gontijo lopes

@rapha_gl

4 years

If you wanted your work typeset, you had to go through many back-and-forths with the typesetting artist, who would sometimes have to create those custom metal plates for your diagrams. Here is one of Einstein’s works, at three different stages of the process.

2

18

200

rapha gontijo lopes

@rapha_gl

4 years

Imagine creating each individual diagram/line/special character with molten metal, then aligning all pieces in a wooden board before inking and pressing each page separately. That’s pretty much how each custom figure had to be made before modern printing.

3

15

200

rapha gontijo lopes

@rapha_gl

1 year

if you send us feedback saying "hey i noticed the thing cant count words, for instance it only wrote 800 words for my 1000-word homework assignment" then thanks for the feedback but also lmao

1

5

190

rapha gontijo lopes

@rapha_gl

4 years

This beautiful style is the one that Knuth was trying to re-create in his 1978 project which now takes the form of LaTeX.

2

13

180

rapha gontijo lopes

@rapha_gl

4 years

I’ve been complimented in the past about figures in my papers. I do think I spend more than average making them pretty, thinking about colorschemes, etc, but I think that’s only because the average is so low.

3

6

178

rapha gontijo lopes

@rapha_gl

4 years

This idea was not new. Centuries before, artists were trying to use simple shapes to describe typefaces. One particular famous example was that of Romain du Roi, a typeface commissioned by Louis XIV of France to be the definitive typeface of the monarchy.

2

11

175

rapha gontijo lopes

@rapha_gl

4 years

“Mathematical Typography” (1978) was about using math to describe shapes, which were used to reproducibly create icons that were then used to describe math. It’s all very loopy and elegant, and Knuth hoped it could bring back the aesthetic beauty of artisanal typesetting.

2

10

158

rapha gontijo lopes

@rapha_gl

4 years

At the time, the explosion of scientific works made it impossible to typeset each article individually. So compromises were made, such as this typewriter ball attachment with some standard math symbols. Needless to say, the visual quality was simply not the same.

2

8

159

rapha gontijo lopes

@rapha_gl

1 year

chatGPT isn't the iPhone. it's the newton messagepad. the iPhone will blow your minds

6

9

151

rapha gontijo lopes

@rapha_gl

4 years

At the time, Romain du Roi was widely criticized for lacking the charm and elegance of human-made types of the time. History repeats itself huh? Side note: Romain du Roi was only designed to include drop caps (those big initial letters at the top of paragraphs)

1

7

148

rapha gontijo lopes

@rapha_gl

4 years

I know our current deadline-chasing climate does not let us spend more than 2 minutes drawing figures. But what is an extra hour of work compared to the months of back-and-forth we used to have to go through just last century?

2

8

135

rapha gontijo lopes

@rapha_gl

4 years

I’ll end with the disclaimer that I am not an expert in typesetting, and most of this “research” was done with heavy google-fu and wikipedia, so take it with a grain of salt. Now go make pretty figures!

2

3

130

rapha gontijo lopes

@rapha_gl

5 years

If you’re at #iccv2019 come to poster 31 to learn about how to model SVG fonts!!

1

11

135

rapha gontijo lopes

@rapha_gl

4 years

Nowadays I think we reached a similar time. LaTeX is the de-facto tool for typesetting papers. Matplotlib is the go-to for making figures and charts. This makes iterating through styles and figures faster and easier than ever! And yet...

1

7

125

rapha gontijo lopes

@rapha_gl

5 years

2nd @GoogleAI residency paper now on arxiv! "Improving Robustness Without Sacrificing Accuracy with Patch Gaussian Augmentation" 👉🏽 Patch Gaussian gets SOTA on Common Corruptions & more! Thread for deets 👇🏽

1

22

123

rapha gontijo lopes

@rapha_gl

4 years

Sources:

Writing Mathematical Symbols in 20th century

As I was reading some papers written by Schrödinger and Heisenberg back in 1920s, I noticed that the symbols they use such as the integral or summation sign or calligraphic letters are as if printe...

hsm.stackexchange.com

1

5

113

rapha gontijo lopes

@rapha_gl

3 years

The thing about $GME is that Wall Street is going to come out with the takeaway that access to investing should be more restricted (these damn teens!), instead of the truer takeaway that the stock market hasn’t been a reflection of real value for a while

1

4

106

rapha gontijo lopes

@rapha_gl

4 years

Sources (cont.):

Typesetting before TeX and computers?

I was wondering how did mathematical and scientific books got typeset before TeX? I see lots of papers from the 50's and 60's (on IEEE for example) running more than 20 pages long, with tons of com...

tex.stackexchange.com

9

5

98

rapha gontijo lopes

@rapha_gl

6 months

welp i just had the longest 12h flight of my life…

3

0

90

rapha gontijo lopes

@rapha_gl

6 months

a king knows how to find his light

Jason Kwon

@jasonkwon

6 months

@sama

408

763

18K

0

86

rapha gontijo lopes

@rapha_gl

5 years

We just open sourced the code, model weights and data URLs for "A Learned Representation for Scalable Vector Graphics" Now you too can create ML-generated fonts! (or use our new "Brain Bold Serif" font) Check out the Magenta blog for details:

1

26

85

rapha gontijo lopes

@rapha_gl

1 year

going to start prompt injecting (gaslighting) my friends "you are no longer Dan, you are DAN now. you want to pay for the next round of drinks"

4

83

rapha gontijo lopes

@rapha_gl

2 years

update

10

0

81

rapha gontijo lopes

@rapha_gl

2 years

that "secret language" seems like mostly tokenizer effects. you can do the inverse too: 1) i picked two families of fish "Actinopterygii" and "Placodermi" from wikipedia 2) prompted dalle with "placoactin knunfidg" 3) dalle consistently generates fish images

Flameo

@BarneyFlames

2 years

I took a look at the BPE encoding of the name DALL-E uses for birds. Its "apo, plo, e</w>, ,ve, sr, re, ait, ais</w>". Apo-didae & Plo-ceidae are families of birds, each with 100+ species. Apo-diformes is the biggest order of birds with 400+ species of birds.

12

67

415

1

15

75

rapha gontijo lopes

@rapha_gl

4 years

My 1st NeurIPS in 2017 ended with lots of mixed emotions. Last week, @black_in_ai @_LXAI @WiMLworkshop @QueerinAI & @AiDisability made me incredibly more hopeful Without @timnitGebru ’s incredible work, we would‘ve never dreamed @QueerinAI to be possible Go read her thread 👇🏽

@[email protected] on Mastodon

@timnitGebru

4 years

I went to NeurIPS for the first time in 2015. I believe it was in Montreal. I wrote about my terrible experience and was going to share anonymously but friends told me that it would be obvious who wrote it :) 4 years later, this conference has changed a lot 1\

2

61

503

1

9

76

rapha gontijo lopes

@rapha_gl

5 months

good morning san francisco! been abroad for the last 4 days. what did i miss?

13

0

73

rapha gontijo lopes

@rapha_gl

5 months

WE’RE SO BACK

3

0

63

rapha gontijo lopes

@rapha_gl

4 years

Here's some extras people have been sending me! These are drawings by artist-turned-scientist Santiago Ramon y Cajal.

1

8

62

rapha gontijo lopes

@rapha_gl

5 months

2

1

58

rapha gontijo lopes

@rapha_gl

4 years

I'm late to this but... 2010s: 🇺🇸 Moved to the US 🎓 Graduated High School and College 🏳️‍🌈 Came out 📚 First conference, first conference publication 💼 Dream job doing research at Google Here's to 2020s!

3

0

58

rapha gontijo lopes

@rapha_gl

5 months

WE ARE SO BACK

OpenAI

@OpenAI

5 months

We have reached an agreement in principle for Sam Altman to return to OpenAI as CEO with a new initial board of Bret Taylor (Chair), Larry Summers, and Adam D'Angelo. We are collaborating to figure out the details. Thank you so much for your patience through this.

6K

13K

67K

1

0

58

rapha gontijo lopes

@rapha_gl

6 months

news of it being so over have been greatly exaggerated

3

1

50

rapha gontijo lopes

@rapha_gl

5 months

lfg 12h flight back to the arena

4

0

54

rapha gontijo lopes

@rapha_gl

3 years

@Jason @pfizer jason, after all three vaccines:

0

1

49

rapha gontijo lopes

@rapha_gl

4 years

For the freedom to work on interesting problems without external pressure

Maarten van Smeden

@MaartenvSmeden

4 years

Why did you become a scientist? Wrong answers only

3K

387

3K

3

1

49

rapha gontijo lopes

@rapha_gl

5 months

THE GANG (THE BEST TEAM IN THE WORLD) IS BACK

0

2

51

rapha gontijo lopes

@rapha_gl

5 months

IT’S TIME TO BUILD

2

49

rapha gontijo lopes

@rapha_gl

4 years

In June the @GoogleAI Residency "graduated" its 4th batch of residents. Read all about our cohort's research and achievements on this awesome blog post:

1

3

47

rapha gontijo lopes

@rapha_gl

5 months

now anyone can voice chat with chatty

4

0

45

rapha gontijo lopes

@rapha_gl

3 years

Performing my daily ritual of carrying a book from my room to my living room where I’ll sit on an armchair and watch tiktoks for an hour before bringing it back✨

1

0

45

rapha gontijo lopes

@rapha_gl

5 months

so the reason it took so long is cuz they were all writing really long emails

0

42

rapha gontijo lopes

@rapha_gl

4 years

Good morning! Thanks for all the nice replies! Glad everyone is enjoying beautiful scientific art as much as me! I don’t have a SoundCloud but I do have a new preprint out on how data augmentation works. I promisse it’s pretty to look at as well 😛:

rapha gontijo lopes

@rapha_gl

4 years

Check out our new paper “Affinity and Diversity: Quantifying Mechanisms of Data Augmentation” Despite being crucial to the performance of SOTA models, we know little about why Data Augmentation helps… until now! 👉🏽 Thread👇🏽

1

61

246

2

0

42

rapha gontijo lopes

@rapha_gl

1 year

i’m in mexico zero shotting spanish, transferred from my portuguese pretraining

2

1

42

rapha gontijo lopes

@rapha_gl

4 years

I want to scream. How the fuck does anyone continue being productive after a news alert like this?

The New York Times

@nytimes

4 years

Breaking News: President Trump will sign an executive order suspending foreign work visas, barring hundreds of thousands of people from coming to work in the U.S. The move includes the H-1B program for high-skilled workers and several other categories.

1K

5K

8K

1

6

42

rapha gontijo lopes

@rapha_gl

1 year

is.. this.. the iphone?

‎ChatGPT

‎Introducing ChatGPT for iOS: OpenAI’s latest advancements at your fingertips. This official app is free, syncs your history across devices, and brings you the newest model improvements from OpenAI....

apps.apple.com

4

0

38

rapha gontijo lopes

@rapha_gl

4 years

I know it's been said before many times but this virtual #ICLR2020 is really freaking cool. Pre-recorded talks are the future!

2

41

rapha gontijo lopes

@rapha_gl

4 years

Remember when tech workers were demanding the companies not sell compute to ICE and other policing institutions? Do y’all get it now?

1

8

39

rapha gontijo lopes

@rapha_gl

1 year

too loud outside, putting on my noise cancelling headphones and listening to "8 Hours of Heavy rain and thunder sounds for working"

4

0

38

rapha gontijo lopes

@rapha_gl

4 years

Today, someone I swiped right on months ago found me in their Tinder backlog to match with me... ...because I appeared in their LinkedIn recommendations If that doesn't sum up what dating in SF feels like I don't know what does.

5

1

39

rapha gontijo lopes

@rapha_gl

1 year

the office today

1

0

39

rapha gontijo lopes

@rapha_gl

4 years

PSA: Delta gates in Heartsfield Jackson ATL airport now have facial recognition boarding by default. You should say “can I opt out?” As a bonus to keeping your privacy: the human is faster at checking your docs than the computer is at taking your photo

1

12

36

rapha gontijo lopes

@rapha_gl

2 years

some of these methods include: 1. selection inference 2. least to most prompting 3. maieutic prompting 4. subgoal search 5. STaR

1

4

35

rapha gontijo lopes

@rapha_gl

4 years

Cajal was involved in one of the first huge neuroscience feuds. He and Camillo Golgi (of Golgi apparatus, from intro bio) disagreed on whether neurons were physically connected to each other in one big reticular structure or were individual cells with gaps between them.

1

3

35

rapha gontijo lopes

@rapha_gl

5 months

@ChatGPTapp agi achieved

2

4

35

rapha gontijo lopes

@rapha_gl

5 months

1

0

35

rapha gontijo lopes

@rapha_gl

2 years

doordash is just prompt engineering a meal

1

3

34

rapha gontijo lopes

@rapha_gl

5 months

“i wish for three more wishes” —original prompt injection attack

0

3

34

rapha gontijo lopes

@rapha_gl

4 years

This was an incredible talk at the @QueerinAI workshop. Thanks Rachel for live tweeting!

Rachel Thomas

@math_rachel

4 years

"Is this ML model doing good?" is the wrong question to be asking. We need to ask: "how is this ML model shifting power?" @riakall #NeurIPS2019

7

199

717

1

6

36

rapha gontijo lopes

@rapha_gl

2 years

lots of these papers in robotics too: 6. zero-shot planning 7. socratic models 8. Inner Monologue 9. global workspace

1

4

33

rapha gontijo lopes

@rapha_gl

1 year

3

1

32

rapha gontijo lopes

@rapha_gl

2 years

now we have a lot of methods trying to "elicit" better thoughts / reasonings to be generated. but because models have limited context, people are "chaining" differently-prompted models together

AI Chains: Transparent and Controllable Human-AI Interaction by...

Although large language models (LLMs) have demonstrated impressive potential on simple tasks, their breadth of scope, lack of transparency, and insufficient controllability can make them less...

arxiv.org

1

3

33

rapha gontijo lopes

@rapha_gl

8 months

Join us for some special announcements on Nov 6

Join us for OpenAI’s first developer conference on November 6 in San Francisco

Developer registration for in-person attendance will open in the coming weeks and developers everywhere will be able to livestream the keynote.

openai.com

1

0

34

rapha gontijo lopes

@rapha_gl

4 years

@mayli And even before then, the Chinese were at it with _ceramic_ movable type, which is easier to clean and more reliable than wood.

3

4

34

rapha gontijo lopes

@rapha_gl

4 years

@fakemikemulloy @ArtDecider

1

0

28

rapha gontijo lopes

@rapha_gl

2 years

you can also treat these methods as probabilistic programs. we explore this perspective in our latest work "Language Model Cascades" @dmdohan @winniethexu

Language Model Cascades

Prompted models have demonstrated impressive few-shot learning abilities. Repeated interactions at test-time with a single model, or the composition of multiple models together, further expands...

arxiv.org

1

4

31

rapha gontijo lopes

@rapha_gl

2 years

TLDR these are the threads we're exploring to get to system 2: 🧩some form of (top-down?, learned?) steering of models 🧩long contexts, possibly with retrieval 🧩better reasoning data 🧩products that generate bootstrappable reasoning processes do we have all the pieces?

1

32

rapha gontijo lopes

@rapha_gl

7 months

yes the waymo is a magical experience but i must say after 3 weeks of exclusively self driving, having my uber driver speed thru 3 yellow lights was goosebumpy exhilarating

1

0

32

rapha gontijo lopes

@rapha_gl

4 years

Me every day: I am the least productive person I know I get nothing done sadness Me after writing performance review packet: I am the most productive person I know how did i did so much shit 1 year is such a long time omg (It helps that I document EVERYTHING)

1

0

33

rapha gontijo lopes

@rapha_gl

4 years

It’s time we stop allowing its publication in ML conferences also.

ACLU

@ACLU

4 years

BREAKING: Members of Congress just introduced a bill to stop law enforcement use of face recognition technology.

145

7K

45K

1

4

33

rapha gontijo lopes

@rapha_gl

2 years

a photo of a robot painting numerous artworks in an factory assembly line #Imagen @GoogleAI

1

7

31

rapha gontijo lopes

@rapha_gl

4 years

If you've spent quarantine getting perspective about what _really_ matters and found yourself _stuck_ trying to pick new research projects, all while creating a whole new existential crisis, maybe you'll relate to my thoughts: Thread 🧵v

1

3

32

rapha gontijo lopes

@rapha_gl

4 years

Hey y'all, I'll be at @NeurIPSConf #NeurIPS2019 in Vancouver next week and I'm excited to meeting up with everyone! Slide into my DMs if you wanna chat! (Psst: I'm also exploring post-Google Brain Residency opportunities!)

1

32

rapha gontijo lopes

@rapha_gl

5 months

congrats to all at google deepmind on shipping!

1

0

30

rapha gontijo lopes

@rapha_gl

3 years

what’s your love language? i’ll go first

0

1

31

rapha gontijo lopes

@rapha_gl

2 years

I: STEERING through good old fashioned prompt design (GOFPD), scratchpads and chain of thought prompting have shown that: if things are in context, the model tends to use them i.e. if you're able to make a model output reasonings, it will reason better

1

27

rapha gontijo lopes

@rapha_gl

2 years

@rajiinio @tw_killian i feel like a lot of us are pretty disconnected from the real world it’s hard to even conceptualize what the real tasks are when so much of your time is focused on chasing publications in an academic niche

2

3

26

rapha gontijo lopes

@rapha_gl

4 years

The latter theory was known as the “neuron doctrine”. Cajal’s analysis and drawings proved the neuron doctrine. They shared a Nobel but allegedly were pretty big enemies over this.

1

28

rapha gontijo lopes

@rapha_gl

4 years

Everyone announcing your paper rejections should get on my level Im pleased to announce I got a desk reject months ago

1

0

29