Shuchao Bi
@shuchaobi
Followers
2K
Following
435
Media
10
Statuses
98
Research @ OpenAI, multimodal, product oriented research; previously: Engineering Director @ Google, co-founder of YouTube Shorts. opinions are my own
Mountain View
Joined December 2009
Today, we are rolling out Search in Advanced Voice mode. You can now get real time information while speaking with ChatGPT. It has been a very fruitful collaboration between the Search and multimodal product research team.
Day 8: ChatGPT Search Day .
14
20
435
Reinforcement fine-tuning is here after 5 months of amazing work. You can build your own reasoning model in your domain now. Would love your feedback and use cases!.
Day 2: Reinforcement Fine-Tuning.
13
10
191
10x cheaper realtime voice API. The internet went from text only to multimodal over the last 25 years: blogs + Google → instagram → short-form videos (YouTube Shorts, TikTok). Think about how many human hours are spent on writing / reading text vs talking or watching videos.
Day 9: DevDay Holiday Edition.
5
12
168
Today we started rolling out video, screensharing, image uploads and Santa Voice in ChatGPT Advanced Voice. This is an important step forward my team drove on the research side toward building a truly multimodal interface to AI. ChatGPT can now see, hear, and speak in real-time.
Day 6: A gift for everyone who has been nice this year 🎅.
10
12
135
we built internet, kernels, software and the whole virtual world as environment for virtual AI to learn.
1
2
61
This may sound fatalism but I believe in it more and more: it feels the whole human race has been preparing for this AGI/ASI moment. 1. We invented languages and printing to record our intelligence. 2. We invented computers, internet, cameras, mobile phones to make it easier to
1
1
22
love these demonstrations. there is also a lot of headroom for future improvements.
A new AI-powered vision tool created an anatomy quiz, reviewed hand-drawn answers on a chalkboard, and provided feedback — all in real-time. For a fun challenge, it also wrote and sang a rhyme about finding the area of a triangle in a British accent.
2
0
23
One evidence that video speech to speech will be the future of human<—>AI interface is that I found the most effective way of getting information from someone or getting some thing done via someone is by chatting with them face to face.
Advanced Voice is rolling out to all Plus and Team users in the ChatGPT app over the course of the week. While you’ve been patiently waiting, we’ve added Custom Instructions, Memory, five new voices, and improved accents. It can also say “Sorry I’m late” in over 50 languages.
2
0
16
It feels like human is in this escape room (earth or the universe) that is created by some superpower. Figuring out AGI and ASI is a key step toward solving the puzzle.
3
0
16
the best way to predict the future is to create it. join us to continuously improve our video speech to speech models to power her:
0
0
16
good prediction.
0
0
13
Achieve AGI together with our customers. 🤟.
new stuff for developers today, including o1 in the API with new features like structured outputs and function calling, realtime API updates, preference fine-tuning, and more.
1
0
10
Come join our team if you are excited about this direction and would love to work with these amazing researchers:
0
0
9
Hot takes: .1. when you have a policy that works decently well, the key to success RL is a high quality grader, that’s almost all you need. Given most of times grading is easier than generating, this can form a flywheel. 2. when the current policy is too bad to generate good
0
0
9
it just got started.
Introducing Operator. Our first agent. It uses a cloud browser to do tasks on the web. Available to all Pro users in the US starting today.
0
1
7
Reading an old paper (from Andrew Barto’s talk) on reward prediction error hypothesis: the phasic activity of mesencephalic dopamine signals the TD error, rather than the absolute reward. This agrees with Munger’s philosophy that the key to happiness is to lower your
0
0
7
The future workspace may be different due to people increasingly speaking to their AI to get things done.
Finally, the dream demo of talking to ChatGPT about stuff on your computer. Shipping today on macOS:.- Advanced voice works with your apps.- New tools like @warpdotdev .- Extending beyond devs with Apple Notes, @NotionHQ, and more!. Next:.- More apps & capabilities.- Windows
1
0
6
now you can call our models, even on non-smart phones.
You can now talk to ChatGPT by calling 1-800-ChatGPT (1-800-242-8478) in the U.S. or by sending a WhatsApp message to the same number—available everywhere ChatGPT is.
1
0
6
Hot take: death and birth is one key way of keep human race’s intelligence entropy from collapsing
0
0
6
On the research side, shout out to the researchers that prepared the data, trained/ iterated / debugged numerous video speech to speech models with Search natively integrated. @SoniaPhene, @ehsanmasdar, @ChengxuZhuang, Damian, @rown.
1
0
5
If you are curious how we make advanced voice so real, check this out
1
0
5
Multimodal will become a major interface to intelligence.
excited about the future too.
0
0
5
"simple" but easily hackable rewards are root cause of so many problems in the society.
0
0
4
@rown @ChengxuZhuang Come join our team if you are excited about this direction and would love to work with these amazing researchers: We would love your feedback of the advanced voice as well.
0
0
4
Watching a visually impaired woman using AVM with video to put on makeup is touching and motivating. Moments like this keep reaffirming our belief of AGI can and will benefit all of humanity.
0
0
3
@ema_dagostino That would a few levels higher than the current AI we have but I am personally optimistic that these self evolving agents are possible.
1
0
3
@Luke_Metz @barret_zoph @LiamFedus @johnschulman2 Thanks for being an awesome onboarding buddy for me, Luke. You achieved something extraordinary here at OpenAI . The success of ChatGPT, past and future, will always be part of your story.
0
0
3
@gdb The next token prediction is not very efficient at modeling the arithmetic attention. It predicts the most difficult part first.
0
0
1
This is the team that contributed most evidence to “there is no data wall”.
We are also hiring top researchers/engineers to keep breaking the data wall and find out ways to pretrain both frontier models & extremely cost/performance efficient models. If you are interested in working on this, apply & drop me an email.
1
0
2
absolutely.
Talking to a model is a new access to the internet, distilled or real-time. If an alien talked to the model it would have no mental model of what humans understand as the internet.
0
0
2
@ehsanmasdar it has been super awesome working on this together. you are a rocking star @ehsanmasdar.
0
0
2
@barret_zoph Thanks for bringing me in and the opportunity working with you and the legendary team. You built something that marked the new era of AI. I learned a lot from you. Best of luck in future endeavors.
0
0
1
Congrats to @karinanguyen_ and the team.
We’re rolling out an early version of canvas—a new way to work with ChatGPT on writing & coding projects that go beyond simple chat. Starting today, Plus & Team users can try it by selecting “GPT-4o with canvas” in the model picker.
0
0
1