President & Co-Founder @OpenAI
Joined July 2010
- Tweets 4,815
- Following 65
- Followers 487,944
- Likes 1,386
The mathematics of measuring how well our true objective (e.g. helpful or accurate responses) is being optimized when training an AI systems on a proxy metric (e.g. the output of a reward model). Pretty cool & not obvious that you can measure this at all!
openai.com/blog/measuring-go…
1
4
1
44
Software engineering: 50% understanding requirements, 40% complexity management, 9% debugging, 1% solving "interesting" algorithmic problems.
You'll enjoy software engineering a whole lot more if you instead think of the first 99% as the interesting part.
33
214
20
1,466
Big challenge in ML engineering: finding aggregate views that let you quickly understand the micro details of everything happening in your system. Often the biggest problems are simple problems obvious from looking at one specific computation. ML rewards attention to detail.
6
32
1
195
It’s time to become an ML engineer: blog.gregbrockman.com/its-ti…
17
134
15
725
As a bonus, here's a DALL-E generation for "a program running on a universe-sized quantum computer":
9
18
2
232
Reality is a program running on a universe-sized quantum computer. So the best way to solve a hard problem is to maximize contact with reality — lets you tap into an unfathomable amount of compute.
28
53
12
389
I also really like this one, created by the prompt "DALL-E dreaming of becoming an AGI":
66
126
58
1,466
Whenever I cowork with my colleagues in person, I realize that the pandemic has completely dulled my sense of just how energizing & productive good collaboration can be. Months of remote work have resulted in many tiny unshared assumptions, which we only notice in person.
2
16
2
192
Get an onsite interview, get early access to DALL•E. Good way to show rather than tell what kinds of projects you could work on here.
(Attached photo: "A walrus typing at a keyboard doing a programming interview at an office".)
We just decided we'll give early access to DALL•E for any applicants to @OpenAI who make it to the onsite interview stage.
Should help candidates get a better sense for where the technology is and how to think about the opportunity.
6
16
6
125
So meta:
8
9
3
123
Sign up for the DALL·E 2 waitlist here: labs.openai.com/waitlist
6
5
2
99
DALL·E 2 — generate any image from a text description. Imagination is the limit.
"A Shiba Inu dog wearing a beret and black turtleneck"
"A photo of a quaint flower shop storefront with a pastel green and clean white facade and open door and big window"
openai.com/dall-e-2/
53
211
184
1,319
All evidence indicates that we are on track to create AGI before reliable video conferencing.
24
52
7
705
AI engineering is poring through logs puzzling out why your run crashed, profiling to increase perf by a more few percent, and figuring out which code version was running during that weird blip.
Only when the model is finally trained do you realize it was magic all along.
7
15
1
120
In a startup, you should seek out activities that seem hard, boring, annoying, and unscalable. The highest-value tasks are often hiding amongst them, and no one else has noticed because they seem unappealing on the surface.
45
285
35
2,155
.@Replit can now automatically find bugs in your code using OpenAI Codex under the hood.
A concrete step towards transforming programming to be more about expressing your intent, and less about getting the incantations exactly right:
8
86
11
600