President & Co-Founder @OpenAI
Joined July 2010
- Tweets 4,815
- Following 65
- Followers 495,334
- Likes 1,386
GPT for learning a new programming language:
So this is pretty amazing. @OpenAI’s playground is actually helping me learn Swift.
I started using it to ask specific questions about my code or when I need a second opinion. It actually breaks it down with clear explanations and reasoning. AI collaboration = the future.
9
10
99
Science optimizes for paradigm shifts; engineering optimizes for steady progress.
Both like to think they are the only way—eg that engineers don't do novel work, scientists are not pragmatic.
But a 10x win matches 25x 10% optimizations, and in practice, each unlocks the other.
6
18
2
151
GPT-oetry is getting pretty good too (and just wouldn't be the same without the DALL-E-provided illustration):
5
14
2
107
Hard to internalize that this image was drawn largely by doing some matrix multiplications & some other vector math (with parameters found by doing some hill climbing in a billion-dimensional vector space):
12
20
5
266
Despite many valid drawbacks, it's hard to beat a monorepo for building a long-term project that involves many tightly-collaborating teams.
4
11
122
3 million DALL-E images generated so far, and have been building out our safety systems in parallel. We are now ramping to 1,000 invites per week:
Update on the DALL·E 2 waitlist: We'll be onboarding up to 1,000 people every week as we continue to enhance our safety systems. openai.com/blog/dall-e-2-upd…
15
11
1
151
zsh plugin using Codex to write your shell commands from a comment: github.com/tom-doerr/zsh_cod…
4
15
3
111
GIF
GPT-3 summarization has gotten pretty good:
HOT OR NOT?? 🔥
I made a quick @airtable script that uses @OpenAI API to summarise any text entered in the first field and adds the "tl;dr" to another field. And it works like a charm!! There are like TONS of #nlp #ml #automation things to be done with the same logic...
1
7
1
67
A reliable system is just an unreliable system whose failure modes have been repeatedly encountered, studied, understood, and fixed.
10
19
2
254
Hard to imagine personally working anywhere else, or working on any other problem besides AGI.
Building next-generation AI systems is some of today's hardest (and sometimes most frustrating) engineering & scientific work, but the reward is well worth it.
Send Sam a DM!
Why work at OpenAI?
1. You'll work on hard & important problems, with the most resources in the field and a small-ish group of the most talented colleagues, with minimal bullshit and maximum agency, be held to very high expectations, and be compensated at the top of the market.
2
6
2
99
Fun article on a Ruby DSL to interpret natural language-like syntax.
Also interesting to realize that these days, it's actually easier to just use full natural language to generate Ruby code.
Making the Ruby interpreter run a program written in a natural language: dmitrytsepelev.dev/natural-l… Comments: news.ycombinator.com/item?id…
2
3
18
My wife & I just said goodbye to our foster dog. He was rescued from a shelter hours before being put to sleep; we cared for him for a month but were unable to save him. He was incredibly loving; a gentle giant who loved our rose garden and the quiet life. Sad but glad we tried.
14
4
1
225
Most stressful part of college was being told it would be the best four years of my life. Completed only two good-not-great years and was always worried I’d passed peak happiness & forgotten to enjoy it.
A decade later, am far happier and still yet to peak. Chart your own path.
11
22
1
326
Every so often I stumble across some emails I sent a decade ago, and it strikes me how little I knew back then. Very exciting to think I might be able to learn so much in the next decade that in retrospect I will feel the same way about my emails from today.
4
4
159
Building software is an iterative game of increasingly deeply understanding a specific problem and balancing the fundamental tradeoffs you discover.
4
41
2
284
ML bugs are so much trickier than bugs in traditional software because rather than getting an error, you get degraded performance (and it's not obvious a priori what ideal performance is).
So ML debugging works by continual sanity checking, e.g. comparing to various baselines.
46
229
39
1,903
Biggest mental shift while switching from classic to scientific programming: explicit “for” loops are now extremely expensive.
Instead, you express as much as you can by chaining hyperoptimized lower-level “for” loop primitives, eg matrix multiplies.
Fun & new way of thinking.
18
18
2
288