President & Co-Founder @OpenAI
Joined July 2010
- Tweets 4,815
- Following 65
- Followers 485,194
- Likes 1,386
A pretty good hypothesis on how this arises (as an artifact of how the text input space is tokenized):
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.
3
3
49
Turns out DALL-E can read the seemingly gibberish writing it produces. Built its own mini-language that is consistent between its text input space and image output space:
DALLE-2 has a secret language.
"Apoploe vesrreaitais" means birds.
"Contarra ccetnxniams luryca tanniounons" means bugs or pests.
The prompt: "Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniounons" gives images of birds eating bugs.
A thread (1/n)🧵
8
23
9
132
A well-architected software project is one where the difficulty of adding a feature is proportional to the delta in functionality you want to add.
Small tweaks should be easy, and big new features that were not part of the design should not be too scary due to the existing code.
11
58
4
331
The hardest task for a senior engineer is to learn a new area where they will need to become a beginner again. But at some point, this is the only way to grow, and doing it successfully unlocks unique impact. Especially true for great engineers switching into machine learning.
37
102
17
1,074
DALL-E empowered a 4-year old to design these red chicken slippers, which now exist in reality. Best kind of Sim2Real result:
My 4 year old designed cute fuzzy red chicken slippers with #dalle a couple weeks ago. 🤩
I had them custom made and they just arrived. Soooo cute 🥰
She named them Henry and Wendy 🐓🐓
This is the future of creativity, young entrepreneurship, fashion design and ecommerce.
7
22
1
157
There's nothing quite so freeing as the feeling of starting a greenfield project from scratch.
3
6
102
DALL-E as a new medium of artistic expression:
been playing with dall-e for ~24hrs and as someone who has never felt traditionally “artistic”, being able to see my ideas in such visual fidelity is incredible!!! im completely obsessed
here’s some stuff ive made so far (will keep adding to this thread):
2
65
Nice result: hinting to a language model that it's ok to show its work (by adding "Let's think step by step" to the prompt) rather than directly outputting the final answer massively boosts results on reasoning benchmarks. True of humans too.
Large Language Models are Zero-Shot Reasoners
Simply adding “Let’s think step by step” before each answer increases the accuracy on MultiArith from 17.7% to 78.7% and GSM8K from 10.4% to 40.7% with GPT-3.
arxiv.org/abs/2205.11916
2
8
1
87
ML systems debugging rewards a willingness to dig across the entire stack, to chase a slightly suspicious signal back to its source, and to derive chains of failures from surprising end results. High cognitive burden but also some of the most exhilarating work upon success.
12
26
5
279
On a day-to-day basis, certainty in one's own beliefs is something that the scientist cannot afford; but on the scale of decades, certainty in correct contrarian beliefs is what differentiates the great careers from the rest.
3
7
89
In most startups, the median business person has less impact than the median engineer, but the top business person has more impact than the top engineer.
To have truly outsized impact as an engineer, need to find the rare business where sustained tech velocity is high order bit.
6
15
171
The most productive engineers also tend to be the ones who create the biggest absolute number of bugs (even if their relative rate is low). Don't be afraid to make mistakes, but instead be afraid to let problems lurk beneath the surface or linger once they have been discovered.
5
18
220
Greg Brockman retweeted
I joined OpenAI with a primarily engineering background. Just last year I got to ship code for such wide-ranging efforts as our interpretability research, the Codex launch, and evals for DALL-E 1.
@sama's right: we have so much work to do that engineers can help with. Join us!
In particular, this is a great time in the field to have a lot of impact as an engineer.
Everything we release--GPT-3, Dall-e, Codex--is just a mile marker and we have a roadmap to make them 10x better. Please come help us do it.
DMs open!
3
9
77
There are now 70 production Codex apps: openai.com/blog/codex-apps/
1
25
1
97
DALL-E 2 is stunningly better than DALL-E 1, but what's interesting (and underappreciated) is the fact that the newer model is much smaller & amount of training compute is similar.
Improvements are essentially all due to algorithmic innovation.
22
64
6
943