Felipe Such
@felipesuch
Followers
2,597
Following
125
Media
2
Statuses
50
Joined January 2015
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Vance
• 1654136 Tweets
Ohio
• 148596 Tweets
Vivek
• 147923 Tweets
America's Hitler
• 132013 Tweets
Pence
• 103532 Tweets
Jack Black
• 89123 Tweets
#WWERaw
• 77551 Tweets
プライムデー
• 62910 Tweets
NCAA
• 59933 Tweets
連休明け
• 59830 Tweets
Tim Scott
• 44173 Tweets
Romney
• 42313 Tweets
National Anthem
• 41440 Tweets
Rhea
• 30484 Tweets
Hillbilly Elegy
• 27967 Tweets
Lester Holt
• 27735 Tweets
Grammy
• 26785 Tweets
#Number_i_ナゲットおトク夏
• 20018 Tweets
Gunther
• 16912 Tweets
#HomeRunDerby
• 16680 Tweets
Gunnar
• 13733 Tweets
#Renascer
• 13616 Tweets
Devin
• 13421 Tweets
Pete Alonso
• 13063 Tweets
Appalachia
• 11464 Tweets
Alec Bohm
• 10223 Tweets
❤️ sometimes we have to take a break from building the most awesome thing on the planet and just appreciate the people we build it with
9
32
756
Sometimes I forget that the milliseconds I shave off translate to incredible experiences.
Announcing GPT-4, a large multimodal model, with our best-ever results on capabilities and alignment:
2K
17K
64K
20
55
494
I replicated Ape-X and finally got around to open sourcing it (). On a single machine it gets human median performance on Atari in about 2h. Very impressed with the algo development by Horgan et al ().
1
35
126
Others: waiting patiently for openai news
Me: waiting anxiously but with ❤️
1
1
76
If you liked the low-key research preview, just wait until you get the real deal
3
0
67
If you are curious about our DALL-E 2 engineering stack, I'll be showing some of the challenges and how we tackle them at OpenAI
DLCT today: second part of the DALL-E 2 talk series!
After a wonderful overview of the science & research effort last week (recording here: ), we are covering the *engineering* efforts next.
Brought to you by
@felipesuch
Starting in 2 hours.
1
9
45
0
2
48
Proud to say I had a small part on this project. Awesome to see how excited the community is.
My experience so far has been incredible when using copilot for golang, unit testing, and scripting.
0
0
10
@francoisfleuret
I think what is specially mind blowing is that you can ask it not once, but hundreds of times and get meaningfully different answers. Not something you can easily do with a human let alone by yourself
0
0
8
Months of work to make it trivial for users. Glad to see it is finally out!
Developers can now create a custom version of GPT-3 for their applications with a single command. Fine-tuning GPT-3 on your data improves performance for many use cases. See results👇
34
239
1K
1
0
9
If you want to do something no one has ever done, you should work on solutions no one has ever (really) tried before.
“one must work on things like cryogenic tanks, turbopumps, etc”. Is a great example of working on preconceived solutions rather than the actual problem.
0
0
8
1
1
6
It’s odd how much time I can spend optimizing things. It’s one of my favorites part of my job. I’ve said “alright, this is as fast as it gets” so many times and I was wrong every time.
When I let myself spend a day doing CUDA work, I'm usually quite happy afterwards. Optimization is a grand puzzle game with an objective score function! I have to make myself stop at a wise point, despite all the intriguing rabbit holes left to pursue. \
11
17
436
0
0
6
Accelerating Deep Neuroevolution: Train Atari in Hours on a Single Personal Computer via
@ubereng
0
2
6
Come check out some cool stuff we have been working on!
Live demo of the next generation of OpenAI Codex, our AI that writes code, on Twitch at 10am PT, Tues Aug 10th. 👀
37
201
897
0
0
4
1
0
3
@savvyRL
@Miles_Brundage
@OpenAI
DALL-E is just internally yelling "there isn't enough head space for a 4th one, quit asking"
0
0
3
@bensprecher
@kenneth0stanley
@jeffclune
We will definitely update the repo if we find a solution worth sharing
0
0
3
@vanschneider
@sama
I really like this direction. It’s how cybersecurity is formulated. Do you know of any work in that direction for general information?
0
0
3
@Zergylord
@gsautiere
@sherjilozair
Not to mention Bowling is a game RL "fails" because most algos use reward clipping ( fixes that). Can't optimize your strikes if you can't tell the difference between knocking 1 pin vs. all of them.
0
1
3
Awesome work by
@AdrienLE
and
@Joost_Huizinga
!
Montezuma’s Revenge solved! 2 million points & level 159! Go-Explore is a new algorithm for hard-exploration problems. Shatters Pitfall records too 21,000 vs 0 Blog: Vid By
@AdrienLE
@Joost_Huizinga
@joelbot3000
@kenneth0stanley
& me
18
188
484
0
0
4
@savvyRL
@jasonyo
@steveddev
@ml_collective
I felt that. 😂. I used to have nightmares about filling out forms by hand.
0
0
3
@jon_perl
I wanted to push the limit on a single machine and made everything into custom Tensorflow ops. Using the same trick from our "deep neuroevolution on GPUs" work () I was able to get around 80k frames/second in a single machine (~1.5k frames/s/core)
0
0
2
@_danielgibbons
@kenneth0stanley
@jeffclune
Our approach is orthogonal to distributed AI. We maximize throughput by utilizing CPUs and GPUs in a single instance. Nothing stops you from using Ray to group many machines.
0
0
2
@rachelim
I had the same experience!! I waited 4 years thinking it was going to take forever. But then it took 20 min. The Uber took longer
0
0
1
TFW you do so well the other team accuses you of cheating!
Open AI gets GPT-3 to work by hiring an army of humans to fix GPT’s bad answers
4
7
47
0
0
2
@jon_perl
I checked out that implementation a while back when it came out. It is very clean and easy to understand. Ray's implementation (and the original) were distributed implementations of the algorithm
1
0
1
@bensprecher
@kenneth0stanley
@jeffclune
The CPU implementation supports unreliable slave instances, but, in my opinion, isn't a great solution. Ideally we would have something that supports many persistent masters and many preemptible slaves. We've tried a few solutions, but haven't found a definitive one yet.
1
0
1