Zain Shah
@zan2434
Followers
2,660
Following
1,958
Media
119
Statuses
1,507
teaching machines previously: @samsung @opendoor @openai @ycombinator S13
San Francisco, CA
Joined June 2009
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Birmingham
• 506930 Tweets
#LaCasaDeLosFamososMx
• 302923 Tweets
#LCDLFMX2
• 234351 Tweets
Protecting
• 201635 Tweets
Gala
• 169158 Tweets
TATI
• 149777 Tweets
原爆投下
• 113564 Tweets
Adrián Marcelo
• 104500 Tweets
Ricardo
• 98263 Tweets
Mariana
• 88402 Tweets
ファミチキ
• 50892 Tweets
Sian
• 46920 Tweets
Stanowski
• 43818 Tweets
Goyang
• 43496 Tweets
Vargas
• 41492 Tweets
Briggitte
• 31547 Tweets
#BangladeshCrisis
• 25520 Tweets
はだしのゲン
• 20451 Tweets
सुषमा स्वराज
• 17738 Tweets
#FileninSultanları
• 14550 Tweets
Zehra
• 12601 Tweets
消しゴムマジック
• 11155 Tweets
Hande
• 10246 Tweets
Last Seen Profiles
Pinned Tweet
I'm annoyed with voice AIs that interrupt me before I'm done speaking 😠
So here's a demo (sound on! 🔉) where it only speaks when I'm done talking, and stops when I interrupt it.
It's crazy how much more natural this feels!
Deets below 👇🧵
65
97
877
PSA that
@perplexity_ai
is actually incredible and has completely replaced both Google and ChatGPT for me. Sources for every query result, conversational search is great, no hallucinations
41
49
471
Super simple recipe:
1) Speech transcript is formatted like a screenplay w user and AI as speakers
2) LLM continuously predicts the future of the transcript
3) if LLM predicts user will keep speaking, it waits
4) if LLM predicts AI will start speaking, it speaks prediction aloud
6
4
96
Now you can make your own cool neural net dream images!
http://t.co/vG7pAdo8CI
#deepdreams
http://t.co/zxbEEI04Z0
20
34
56
Core ingredients:
✅ real-time speech recognition w diarization
✅ a fast LLM w completion API
✅ interruptible speech synthesis
5
1
55
Finally pulling back the curtain on some work stuff today! Proud to have had a hand in designing "Circle to Search" and "Note Assist"
4
2
50
This is an insane discovery! 🤯 Shows just how far we are from understanding what's going on. DALLE-2 generates gibberish text because of how they implement text encoding, but by using that gibberish as a prompt it turns out it refers to categories like "birds", "bugs", "seafood"
DALLE-2 has a secret language.
"Apoploe vesrreaitais" means birds.
"Contarra ccetnxniams luryca tanniounons" means bugs or pests.
The prompt: "Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniounons" gives images of birds eating bugs.
A thread (1/n)🧵
204
2K
9K
1
7
41
Join our cult! We have onigiri 🍙
The line vibes were also amazing. ✨ Idk why this hasn’t received local press coverage yet. Journalists, please interview me or other people who were there during 7-Eleven San Francisco onigiri launch day. This is very good news!
32
29
269
3
1
41
IMO this is exactly the wrong set of compromises bundled together. Yes, it's advanced tech, but it's in service of making an unobjectionable device rather than something actually useful. Opposite of bold. 1/4
1
0
38
We’re in 🌱 life on Earth’s 🌎 fifth and final chapter 🎬, and we don’t talk about it enough. In 1B years our sun ☀️ will run out of hydrogen ⌛️, it'll get real hot 🥵, all liquid water on Earth will evaporate♨️, and life will cease to exist ☠️ (1/7) 🧵
2
1
32
PS: One further implication - I fear the “metaverse”. It is the most dangerous and irresponsible idea of all time. It can tear us away from reality into our delusions at precisely the time when engaging more deeply with reality and exploring it is most pressing.
#defendIRL
(8/7)
3
1
28
To rephrase - predicting the next word is the objective, not the capability. Just as learning to throw a ball may be accomplished any number of ways, including learning physics and calculus. There is no bound on capabilities acquired to accomplish the objective.
I absolutely do not understand how "predicting the next word" implies any bound on the cognitive capabilities of a system.
Quite the opposite since, if it was indeed able to do it, it'd be able to answer *any question*.
115
40
700
1
0
20
Still, we may be alone. In this cold, ruthless, chaotic universe, a self-sustaining phenomenon lasted long enough to see its own fragility, and to be able to do something about it. You, me, we - are that. It is truly the most incredible, beautiful miracle I can imagine. (3/7)
2
3
22
There’s a chance we aren’t just the stewards of Earth, we are the stewards of life itself. (7/7)
2
1
19
This incredible work illustrates the power of good generative models. Image editing software powered by GANs:
0
7
18
My greatest fear is that the “great filter” which prevents life from pervading the universe is not the obviously dangerous - nuclear war, biological weaponry, artificial intelligence gone awry - but actually the innocuous allure of virtual reality. (9/7)
3
4
18
Still, it's possible that isn’t true. That this did all just - happen. I'm responsibility-mongering, not fear-mongering. We should spend our energy appreciating reality deeply, caring for each other, figuring out if we’re alone, and doing what we need to in case we are. (6/7)
1
1
18
We should enjoy it as it is. This recognition gives even the most difficult times so much rarity and beauty. There’s much we don’t know, but at some point the life we know will probably still need to end. Some acceptance of impermanence is necessary. (4/7)
1
2
15
In light of Hereticon in Miami, it's Interesting to consider just how many counter-cultural ideas were incubated (though not necessarily started) in the SF Bay Area. E.g. psychedelics, gay rights, burning man, startups, farm-to-table cuisine... what are others?
2
0
15
I just published “Building a Deep Learning Powered GIF Search Engine”
0
4
13
I cannot overstate how special and formative this community was for me. I truly don't know where I'd be without all the incredible people I met through Interact. I highly encourage anyone interested to apply!
Applications for the Interact Fellowship close at midnight PST! If you are an 18-23-year-old mission-driven technologist—designer, engineer, founder, engineer, scientist, writer, artist—apply here: 💫
0
8
32
0
2
13
@newcomputer
big moves! Came for the next gen computer, stayed for the punctuation motifs
@sjwhitmore
@jasonyuandesign
1
0
12
The bright side of these awful layoffs is the end of the FAANG hegemony. They have been hoarding talent for decades.
I worked at Facebook when we had 200 employees
I’ve never understood why FB needs 20,000 employees. And less so the 70,000 employees Meta has now
@elonmusk
’s Twitter layoffs are appropriate
Unfortunate, but appropriate
These businesses just don’t require that many people
3K
6K
83K
1
0
12
I couldn't be more excited to support my friends
@patriciamou_
@adimelamed_0
@iamrosewang
and
@jasoncbenn
building the vibrant, nerdy, cozy hangout spot of my dreams. Come join the fun!
🚨 We’re excited to share that we’ve found a space for The Commons!
ft. a co-working space, philosophy salon space, communal library, cafe, meditation studio, and lots of cozy reading nooks.
We need your help in securing the space!
👉
🧵on our vision
11
40
282
1
2
12
I hung a sheet of Mylar outside my window to get real light from my lightwell. Sadly necessary, but it works
#cityliving
0
0
12
How are Apple and Microsoft still the two most innovative tech companies in the world, bringing the future of computing to the masses? What year is it? 1983?
2
1
12
The stars will dwindle, the universe will be only black holes, with no energy left. Maybe life exists at many levels, the universe itself an enormous biome at scales of time and space we couldn’t possibly conceive. Of which we are just a self-important part. I hope so! (5/7)
1
0
11
Maybe there’s other life out there - another proverbial torch carrier - I really hope so. Latest estimates with basic distributional assumptions put it at about a 50% chance, which is pretty good! (2/7)
1
0
11
This is actually a mindblowing result buried in the new Code LLaMa paper. Great find. Use a ton of artificial LLM generated instructions to fine tune LLaMa2 and we get almost GPT-4 level performance.
They don't want you to know that synthetic data is the future.
LLMs generating synthetic data to train on drives a huuuge boost in "unnatural" code llama -- the one model they aren't releasing. Surpasses gpt-3.5 and gets close to gpt-4 performance on a 34B model
65
218
2K
0
0
9
I expected Rome to be pretty. But I was not prepared for it to be so breathtakingly beautiful everywhere I looked at all times. The city is a gallery, a monument to beauty itself - manmade, modern, ancient, and natural. Photos do not do it justice; I’m in awe.
0
0
10
I'm grateful to have known Michael Lai for many years and I've seen how fully he devotes himself to the causes, things, and people he believes in: as a founder, an educator, a public servant, and as a friend. Michael believes in SF and is exactly the kind of leadership this
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 🧵
103
148
1K
1
0
10
Actually, I know who it's for. It's not for the user, that's for sure. It's for everyone else to not be creeped out 🙃 5/4
0
0
10
So proud of
@noor_siddiqui_
, what an enormous accomplishment and gift to the world! Full genome sequencing is huge for reducing risk of genetic disease. The next generation will suffer less thanks to you. Couples considering IVF - use
@OrchidInc
!
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.
32
59
378
1
1
10
Awesome work. Like Sora, it’s a “learned world model”, but unlike sora it is dynamically controllable and interactive - literally a learned video game engine
Google presents Genie
Generative Interactive Environments
introduce Genie, the first generative interactive environment trained in an unsupervised manner from unlabelled Internet videos. The model can be prompted to generate an endless variety of action-controllable virtual
79
532
2K
0
1
10
It's not ambient AI; it won't do anything without explicit interaction. It's not screenless; it just has a projector instead of a screen because you can't see it where it is. It's not private, it's not in your ears or on your mouth, so you have to talk with it out loud. 3/4
1
0
9
It is not so many things. What is it? Who is it for? No clear answers. 4/4
1
0
9
This essay I found posits that the Bay Area has actually been so productive in this regard because of the storied tension between two opposing and prominent cultures: the hippies and the techies
1
2
9
For anyone wondering what this "GPT-3" thing is - I gave a talk to my housemates contextualizing it in decades of language modeling research! My hope is to convey not only where we are now but how it differs from what came before.
0
1
9
same energy
Exclusive: In a December meeting w/ top execs, Satya Nadella laid into his head of research.
"OpenAI built this with with 250 people," he said. "Why do we have Microsoft Research at all?"
My deep dive on how $MSFT swallowed its pride to back OpenAI:
48
181
2K
0
0
9
It has a camera and mic and is located conspicuously on your chest, so it’s creepy.. but it doesn’t actually run all the time, so not that useful. You have to touch it to interact with it, so it’s not actually hands free, but it’s not located somewhere touching is convenient. 2/4
1
0
8
Found an old pedometer in my room. Does this count as a "wearable"? Am I cool yet?
0
1
7
🛫 En route to Sydney for ICML! Hit me up for ☕️🍻 if you'll be around
#icml2017
gotta
#shutdowntheai
0
0
7
h/t to
@nickcammarata
@joshalbrecht
@willwjack
@serenabian1
@christinahkim
for helping me develop these thoughts.
Sources:
0
0
8
I love the dedicated device form factor. I don't care if it *could* be an app. Before the iPhone bundled everything we had a deluge of devices for all our different needs and consumers actually had choice, they could choose to buy only what they need.
2
0
8
Self-play is what made AlphaGo capable of superhuman play. Self-play is only possible when there are objective evaluation criteria. Typically not the case for language models, so we’re stuck imitating humans. But it’s totally possible for language models if they’re writing code!
"Language Models Can Teach Themselves to Program Better"
This paper changed my thinking about what future langauge models will be good at, mostly in a really concerning way. Let's start with some context: [1/11]
44
654
3K
1
1
8
Hmm so LLM reasoning ability comes from code data, LLMs can also generate their own code training data, and we can filter that generated code for correctness by just evaluating it.. we could bootstrap our way to superhuman reasoning with this recipe alone!
The reason why open LLMs like LLaMA are still lacking? Probably because they didn't see enough code during pretraining
24
54
414
1
0
8
Check us out on ProductHunt! Clara is a virtual employee that schedules meetings for you:
http://t.co/ao1ZgSedcu
0
2
8
"The way you truly win big in software is to take something deeply abstract, weird, and confusing, then put an interface on top that makes it look like the most normal thing in creation."
0
1
8
Getting another opinion helps, even if it's less accurate overall! Learn that and more about ensembling from my colleagues on the data science team
@Opendoor
cc
@Chris_Said
0
0
8
Today’s the day to vote!
I’ve never been more optimistic and excited about local politics than I am for this election. It’s all due to
@Mtclai
and the phenomenal DCCC slate he’s running with. They truly love this city and have concrete plans to make it way better. Margins can be
voice is back today!
i had to say this: the margin for victory for the last seat on the DCCC in 2012 was 32 votes (!!).
with two days until the election can you text 3 friends about me and
@SFDemsForChange
?
7
15
91
1
1
7
Nobody exists on purpose. Nobody belongs anywhere. Everyone's gonna die. Come watch tv?
0
0
7
A “web browser” where the search box is GPT3 and the pages are generated rather than found. Just the right level of skeuomorphism imo. What do y’all think?
Today the OpenAI API was very stable, I didn't have any timeout so here is a new demo of my latent web browser!
There are a lot of fails, but it is easy to close the offending tab and click on another link to throw the dices again 🎲
(note: video accelerated for brevity)
19
66
482
0
0
7
Why does it come as a surprise that tools which enable us to share information, but only with others who share the same views, would undermine democracy?
0
0
7
@TrungTPhan
While I think it's true that LLMs are valuable enough we'll have them running on-device, and it's totally doable, by the time this is shipped in iOS we'll already have a much smarter LLM in the cloud (think Alpaca running on your phone, but GPT-5 in the cloud)
1
0
7
I'm annoyed with voice AIs that interrupt me before I'm done speaking 😠
So here's a demo (sound on! 🔉) where it only speaks when I'm done talking, and stops when I interrupt it.
It's crazy how much more natural this feels!
Deets below 👇🧵
65
97
877
1
0
7
Inspiring. Reminds me of
@simonw
's analogy of LLMs as "calculators for words". Points to foundation model cos like OpenAI and Anthropic having the most revenue but not highest margins. Who will the Apple / MS of this new wave be? ~An LLM in every home~
0
0
6
This is very cool! So much less boilerplate, and these days cloud storage is so cheap this is a no-brainer. Yay for reproducibility:
0
0
6
@andy0___
@perplexity_ai
I agree that part is very surprising! It may be because I'm asking it fairly specific questions. But yes I'm making queries that afaict are very susceptible to hallucination, like whether a certain piece of software exists for Mac OS / whether it has a certain ability or feature
0
0
6
@CoffeeVectors
It’s just using the web speech api here. On Chrome this means it’s actually synthesized on Google Cloud
0
0
6
Such a creative product! Can't wait for AI to fill our lives with more bite sized delights just like this.
I made an AI clock for my bookshelves! It tells the time with a new poem EVERY MINUTE composed by ChatGPT. And it has... weird vibes??
tl;dr it's on Kickstarter as of right now
I want to tell you the story because I didn't mean to do this...
55
129
1K
1
0
6
the multiplayer math game
@__ishaan
and I have been working on is done! Play Axis:
http://t.co/TIUcojqGPv
1
3
6
Trump is getting the full 2020 experience:
☑️ Got COVID
☑️ Lost his job
Next up:
⬜️ Being evicted from his house
0
0
6
@danielgross
I think speculative decoding with the smaller proposal LLM local and the larger LLM remote would be an interesting instance of this
0
0
5
Such a cool, audacious, curiosity driven project and it actually worked! Glad to live in a world that supports it. More of this plz
Ten months ago, we launched the Vesuvius Challenge to solve the ancient problem of the Herculaneum Papyri, a library of scrolls that were flash-fried by the eruption of Mount Vesuvius in 79 AD.
Today we are overjoyed to announce that our crazy project has succeeded. After 2000
2K
16K
70K
0
0
5
@colinfortuner
No headphones necessary! Diarization does double duty and also eliminates the need for acoustic echo cancellation of the ai speaking
0
0
5
Something I've been waiting for
@lyft
to do for a while, and is going to be huge. Ride sharing with other riders:
http://t.co/fY4dBO8pa9
1
1
5
Who is going to fine-tune Code LLaMa w a vision encoder so we can finally get that sweet sweet napkin sketch to code GPT-4 demo they teased us with 5 months ago? 👀
You knew that was coming: Code LLama !!!
- Llama-2 tuned for code generation, debugging, etc
- base models, Python-specific, and instruction-tuned.
- 7B, 13B and 33B params models
- Available with the same license as Llama-2
- Blog:
- Paper:
46
366
2K
3
0
5