We have reached an agreement in principle for Sam Altman to return to OpenAI as CEO with a new initial board of Bret Taylor (Chair), Larry Summers, and Adam D'Angelo.
We are collaborating to figure out the details. Thank you so much for your patience through this.
ChatGPT can now browse the internet to provide you with current and authoritative information, complete with direct links to sources. It is no longer limited to data before September 2021.
ChatGPT can now browse the internet to provide you with current and authoritative information, complete with direct links to sources. It is no longer limited to data before September 2021.
We’ve added initial support for ChatGPT plugins — a protocol for developers to build tools for ChatGPT, with safety as a core design principle. Deploying iteratively (starting with a small number of users & developers) to learn from contact with reality:
annual confession that I think most of my buds are from interact and am still very grateful for the 2015 retreat for convincing me to move out to sf 🥲 apps are open now!!
one of the main issues with language models is making stuff up. allowing ChatGPT to use a web browser to research isn't a full solution to this problem but having citations and access to up-to-date info makes it easier to fact-check replies
spent the holidays in Buenos Aires and I think my favorite bit was that ice cream shops were open till 2am. a much better alternative to bars (bars closed even later fwiw)
This is one of my favorite use cases of the chatgpt API - to find OSS malware/vulnerabilities (even with attempted prompt injection in the comments)! Congrats
@feross
+
@SocketSecurity
on the launch!
We even saw an instance of "prompt injection" where a malware author added a comment to convince the reader (human or AI) that their code was not malware.
And...
🤖 Socket AI: "While the author claims it is for bug bounty purposes, this behavior can still pose a privacy risk."
bought a 1970s 35mm film camera off eBay and I’ve been v much enjoying intentionally slowing down and composing shots. it’s also very cool to me that the camera has no digital bits and still works 50 years later. meanwhile I’m swapping my phone every few years
Why I’m excited: I see dialog as critical interface for AI. With prompt engineering, you get one shot at getting your answer. With dialog, you’re able to iterate with the model through back and forth to get your answer 🤝
read The Library of Babel by Jorge Luis Borges because
@iantbd
said it was related to large language models🙄and am still thinking about it a few weeks later!!
🔍looking for more short story recs (does _not_ need to be related to models)
We’re excited to announce the first whole genome screening for embryos is available!
Parents can get 100x more data about their embryos’ genomes, empowering them to make an informed decision and give their baby the best chance at a healthy start.
Which is actually how we come to answers or work with others in real life — we constantly edit, go back and forth, refine (eg me drafting this tweet 🙃)
Twitter is cool bc you get rewarded for having thoughts. Same with any internet presence. I’m more likely to think someone is thoughtful bc I’ve read their blog. Even if the blog posts are just OK. Anyways the tldr is output is good and getting rewarded for outputs is good
I deeply regret my participation in the board's actions. I never intended to harm OpenAI. I love everything we've built together and I will do everything I can to reunite the company.
Got LASIK last week, and WOW I can't believe I didn't do this sooner. Also recommend laser hair removal. Basically, I am pro lasers as a panacea for all your problems.
Thrilled to see my longtime friend
@Mtclai
stepping up to run for SF DCCC on March 5th, 2024!
I've known Michael for years and trust him to bring positive changes to our city
I’m running for office for San Francisco DCCC on March 5th, 2024!
In the city where ppl are building AI & longevity drugs, it's crazy that the basics don't work:
Streets are unsafe
New housing is blocked
We don't teach 8th grade algebra
I'm running to change SF 🧵
accidentally locked myself out of the bedroom of the place I’m currently at. it’s an automatic electronic door - ended up unlocking it by turning off the electricity. turns out turning it off and on is still the best debugging method 😅
Highly highly encourage you to apply to interact. It's been a treat to grow up with these folks and can't wait to add more thoughtful humans to this special community. I currently work and live with interacters 🤗 DM me if you're curious/interested
#joininteract
Hello twitter I’m back for a good cause:
@joininteract
has been a deeply meaningful community to me and hundreds of other technologists. If you’re under 24 I highly encourage you to apply for the seventh Fellowship Class.
The last time
@noor_siddiqui_
and I were in a limo was for a high school dance 😂
Congrats to two of my favorite schemers on pulling off the most ridiculous and tender proposals 🥰
I wanted to ask Noor to marry me in a cute and grand way, so I planned a treasure hunt for her.
It led her to 7 Bay Area locations, looking for clues with her best friends in a limo!
I was waiting for her at the end :)
Here's the full video:
2/5
In the low data regime, pre-training is helpful across model sizes, but especially in large model sizes. When using pre-trained models, model performance is parameter limited.
Followed a interwebs rabbit hole last night on stoicism and came across Musashi Miyamoto. This line from Dokkōdō, written right before his death, resonated with me: “think lightly of yourself and deeply of the world.”
*really* hoping that everyone uses up all their fireworks tonight in an effort to out-do their neighbors so I can stop wondering if the mission is under siege every night
it's been a glorious 365 + 1 days at Sourceress :) can't believe that just a year ago it was the 4 of us (+ our remote folks) working out of a room in our house. And now we work above our house and have grown to 12 people. It's been a magical ride and can't wait for the next year
It is the 5 year anniversary of me truly believing I was teaching my siblings something shocking. In my defense, I had only seen persimmons in the Asian grocery stores growing up. Simpler times 🙈
Noor has been scheming about this since we were in high school. So excited for Orchid to add more science and agency to one of the biggest decisions humans make!
Be like Noor. Don’t let your dreams be dreams!
Excited to finally share what I've been working on at
@OrchidInc
✨
Some context:
I grew up watching my mom lose her vision to a degenerative retinal disease.
I remember learning assembly in my operating systems class and thinking wow SO glad we’ve abstracted this away and can use higher level programming languages - Codex represents the next stage of programming
Everyone says they want ‘community’, but do they also want the endless social obligations, 3am crisis calls, constant attention to the trust between two people who aren’t you, hours spent on the phone hearing why one person is mad at another? And this is in a healthy community!
Whenever I learn something new, I’ve been visualizing myself as a 2D video game character punching the air, receiving 3 more stars, and proceeding to the next level 🎮 Definitely unnecessary but I think noticing when I feel like I’m learning actually helps me remember it better
my project builds upon "Scaling Laws for Transfer" by
@Hernandez_Danny
et al. () and continues the study of transfer between distributions and looks at scaling between three other languages and English
Scaling laws for transfer are important because the scaling relationships explain how to work in limited data regimes. Ideally, you'd have an infinite amount of data for a model to learn from but getting a large quantity of high quality data is nontrivial
I found that a) pre-trained English models help most when learning German, then Spanish, and finally Chinese and b) transfer from English to Chinese, German, and Spanish scales predictably in terms of parameters, data, and compute
Listen to my sweet, sweet pipings about hiring, remote work, and Sourceress! Thanks for having me
@andrewlinfoot
- was a blast 😊
And if you liked what I said, we're hiring ;)
Pytorch-bert v0.6 is out with OpenAI's pretrained GPT-2 🦄 small model & the usual accompanying example scripts to use it.
Now... can you guys wait that the ACL deadline has passed to release any crazy new transformer? 😅
Thanks, you are the best! 🥰
my project tries to the answer: how much does pre-training on English help when transferring across different languages as we vary the dataset size and model size?
@eriktorenberg
Threw a world record breaking birthday party -
@jasoncbenn
and my attempt of catching the most lemons caught blindfolded in 30 second is currently pending with the Guinness World Records :)
Read this great thread by
@j_asminewang
on why you should donate to bail funds.
I've been volunteering with a group to help friends/family find arrested protestors and connect them to bail funds/lawyers and bail funds could use the help! Stop pretrial detention!
more details are on the blog post:
this thread is also an
#ad
for more folks (especially when compute-limited) to consider scaling laws flavored experiments! Also s/o and thanks to way too many people for their advice and feedback🥰
my friend tells me this tweet is missing the context that I’m in quarantine in Taipei, I don’t speak mandarin, my phone+laptop were in the locked room AND I didn’t have pants on. always take your phone to the bathroom is the 2nd moral of this story
you could build a communal living/working/eating space in the center and everyone could still have their own separate, private spaces. the issue with coliving houses now is that many of the spaces aren't conducive for coliving. intentional design of spaces matters!
Yesterday Beth introduced us to a very fun game - Human Reinforcement Learning. You pick a secret task for someone to do and snap to give them rewards as they try to figure what they're supposed to do. TLDR;
@jasoncbenn
is a very good agent.
Noticing the meta level of what’s happening and processing feels like putting a sticky note or highlighting text in book. Makes it easier to find again later.
TIL: that most berries are not berries and things you think are not berries are berries! Watermelon, canteloupe, and grapes are all berries. But blackberries, cherries, and raspberries are not berries.
#fruitfacts
“The story of human intimacy is one of constantly allowing ourselves to see those we love most deeply in a new, more fractured light. Look hard. Risk that.”
I compared the test loss for transformers trained from scratch v.s. fine-tuned on different sizes of the target dataset. Dataset size spanned six orders of magnitude, and model size spanned two orders of magnitude trained on three different languages: Spanish, Chinese, and German
Just used
@Upwork
to collect data for a new project I'm trying and wow they got back to me within an hour with everything I wanted and I am delighted!! Crowd computing is just as magical as ML
#NotAnAd
FiDi‘s really eerie at night and it’s a bummer to see a neighborhood die after work hours. This is why I love the mission - there are schools, offices, restaurants, and bars making the streets feel alive during most hours of the day and night
Went on a run to ocean beach and once I got there realized I had zero desire to run the 5 miles back home - and hopped on a jump bike instead. The future is now! We don’t need flying cars, we just need more bikes.