CTO of Technology & Society
@Google
, working on fundamental AI research and exploring the nature and origins of intelligence. Order my book 'Who Are We Now?' ⬇️
We’ve posted an analysis of the (now infamous) “AI gaydar” paper showing that the authors’ conclusions were probably wrong. Wrong in the 1800s, still wrong today. Bonus: (attempted) straight and gay selfies from us.
Just published: Do large language models understand us? It’s sometimes claimed that ML is "just stats" and AI can't "understand". I'm arguing that LLMs have a great deal to teach us about language, understanding, intelligence, sociality, even personhood.
Our first two AIY Projects, Voice Kit and Vision Kit, are now available at Target stores, with everything included in the box-- yay for their new STEM aisle!
#AIYProjects
Advances and Open Problems in Federated Learning:
from 22 Googlers and 36 academics at 24 institutions! (a) improving efficiency, (b) preserving privacy, (c) defending against attacks, (c) addressing fairness. Here’s to years of future progress on these :)
After years of development, excited we have finally launched Clips-- Google's first product relying fundamentally on on-device AI. Privacy win. Happy with the use case, and proud of the team.
Google Clips, a hands-free camera that learns to help capture spontaneous moments of family, friends, pets and yourself. Available now — learn more at:
One of the cornerstones of privacy-preserving machine learning is the ability to run models on a chip in situ, such that the data don't need to go anywhere. Not all AI should run in the cloud! Proud our AI chip is now public:
#AIYProjects
#EdgeTPU
Artificial General Intelligence is Already Here, from
@NorvigPeter
and me on
@NoemaMag
"Today’s most advanced AI models have many flaws, but decades from now they will be recognized as the first true examples of artificial general intelligence."
Proud of our second
#AIYProjects
effort, Vision Kit— bringing on-device deep learning to
#RaspberryPi
for speed, robustness and privacy... for under $50. Can’t wait to see what gets made :)
Beautiful work from colleagues Johannes von Oswald and crew-- randomized algorithms can be learned by transformers. Randomness is key not only for classic algorithms in computer science but for cybernetics and social brains.
@bratton
and I describe 7 problems:
Machine Majority Language Problem
Ouroboros Language Problem
Apophenia Problem
Artificial Epistemology Confidence Problem
Fuzzy Bright Line Problem
Availability Bias Problem
Carbon Appetite Problem
@NoemaMag
🧵
Very proud of the team's work on a number of AI features in Pixel 4, including realtime transcription in the Recorder app. Useful and beautifully done.
The new Recorder app uses speech recognition and AI to transcribe lectures, meetings, interviews and more—and makes them easy for you to find later. (English only right now, with more languages to come.)
#madebygoogle
From
@dhanya_sridhar
@tyrell_turing
@g_lajoie_
and I, our tl;dr on AI x-risk: Thanks
@TheEconomist
By Invitation for helping us get this centered take out there. Lots to do in AI ethics+safety, let's not get distracted by catastrophism+concern trolling.
Beautiful new work from the SelfOrg team elaborating on Neural Cellular Automata: how, under biologically realistic constraints, arbitrarily complex morphogenesis can happen using purely local and isotropic signaling! Differentiable "DNA", so learnable :)
New article: "Isotropic Neural Cellular Automata", a review of our previous work on making NCAs rotation and reflection equivariant
@RandazzoEttore
@eyvindn
@cwpfautz
Sam, thank you for a great interview! Hope this is of interest to anyone thinking about neuroevolution, post-optimization paradigms, Alife, and sociality in AI.
Today we’re joined by Blaise Aguera y Arcas (
@blaiseaguera
), Distinguished Scientist at
@GoogleAI
. In our conversation, we discuss his talk 'Social Intelligence' from
#NeurIPS2019
, his team's role at
@Google
, and much more!
@WorldSummitAI
@Kantrowitz
@frossi_t
Beyond the binaries of FACT and FANTASY we also have the wide provinces of “we don’t know” and “we may need to expand our vocabulary to talk intelligently about this”. Hype vs cynic narrative is pretty limiting.
Thank you
@NoemaMag
, this was a really nice collaborative effort to raise awareness of the underreported but far more moderate / heterodox / I believe majority perspective of AI researchers re x-risk. Shout out to coauthors
@dhanya_sridhar
@tyrell_turing
@g_lajoie_
, +
@latams
It's great to see the press recognize that we have been working on foundational technology for privacy for years-- not just in response to recent developments.
Fun little piece in bite-sized chunks on Leibniz, AI prehistory, speech synthesis, and nonhuman intelligence (featuring talking dogs!).
1011 theses on the concept of nonhuman intelligence
#audiolm
My novella, Ubi Sunt, launched at the San Francisco Art Book Fair yesterday. Available for free online at ! Buy the beautiful paper edition from Hat & Beard at .
@practise
@hatandbeardbks
"AudioLM: a language modeling approach to audio generation", w/
@zalanborsos
and colleagues.
AudioLM achieves high-quality and long-term consistency, from audio only. Meaningful speech w/o text, consistent piano w/o MIDI.
📃:
🔊:
I'm so proud of how far Federated Learning has come over the past 7 years. The Field Guide will be helpful to everyone working with or advancing the state of this basic technology. Privacy and large scale machine learning: we can have both <3
Originating from conversations at the 2020 Google Workshop on Federated Learning and Analytics (), 26
@GoogleAI
researchers and 27 academic researchers began a year-long collaboration resulting in A Field Guide to Federated Optimization
Excited to see our on-device AI platform, Coral, starting to scale up. This started as prototyping for makers, but I believe it’s on its way to becoming big. Running models locally on specialized hardware opens avenues to stronger privacy and less energy use.
@GoogleAI
Coral products are now out of beta. Coral is a set of products that provide high performance, low power use of machine learning in many different kinds of computational environments, enabling new ways of using machine learning in the world.
Proud of the team’s work on Federated Analytics-- foundation for analytics and telemetry to be done without compromising privacy. Part of the bigger project of solving simultaneously for individual agency and collective intelligence.
#GoogleAI
#privacy
Early in 2021, between covid shots 1 & 2, I wrote a (literal) fever dream novella, Ubi Sunt. Very excited it's now available for preorder from awesome indie art press Hat & Beard. Huge thanks to JC & Sybil of
@hatandbeardbks
and to design genius James Goggin
@practise
Ubi Sunt, the 1st title in a new literary series called Hat & Beard Editions, due out this spring.
@blaiseaguera
's imaginative mind and
@practise
's bold graphic design are on full display here. for more.
1/2 Excited to see in Wired:
“Most AI giants [...] rely on the continuous collection of personal data from their users [...] This practice, however, comes at a cost. [...] In 2019, we will see an alternative [...] emerging in the form of AI at the edge--
This is to our knowledge the first time Federated Learning with a rigorous differential privacy guarantee has been shipped in production. An exciting milestone! Go team :)
Introducing a proof-of-concept production deployment of a next-word-prediction model for Spanish-language
#Gboard
users that was trained using
#FederatedLearning
while providing a rigorous
#DifferentialPrivacy
guarantee. Learn how it was done ↓
How do we envision our future?
Who are “we,” anyway?
Does “we” include collective identity groups as well as individuals? Nonhuman animals and plants? Governments and corporations? Will it soon include robots?
2/n: 10/28, 7pm EST: Perspectives on consciousness: Michael Graziano on how consciousness may emerge from motor control and sociality; and Frederick Bennett on psilocyin and what psychedelics can reveal about the mind; discussion moderated by Blaise Aguera y Arcas (Google AI).
Join us next Wednesday, December 13th for a
#WhoAreWeNow
reading, presentation, and Q&A at
@ElliottBayBooks
.
I’ll be signing some of the first copies at the event, before the book is released on December 19th :)
Brilliant dreamer, lovely human, and fellow Googler Henri Astre
@dddexperiments
has recreated Photosynth2 for the Internet Archive, and for the ages. A small piece of internet history and an epic bit of hacking. Hats off
Last week on the
#savagelovecast
,
@fakedansavage
and I had a rich, wide-ranging conversation about human identity and my upcoming book, Who Are We Now?
Thanks so much for having me, Dan.
Listen to the full episode:
Come see Melanie Mitchell and I discuss the state of AI at Town Hall tomorrow eve! Or watch on livestream.
Is it intelligent yet?
7:30pm PT.
Discussion will be moderated by Lili Cheng.
@MelMitchell1
@lilich
Had the pleasure of speaking with
@mudede
about the future and impact of AI. Thanks for the feature,
@TheStranger
! Read the piece and watch the full discussion at the link here:
My latest interview about identity, language, biases, and "Who Are We Now?" for
@frontier_is
. Thank you for the thoughtful questions
@briansholis
!
Read the full conversation:
"the relentless unforeseen was what we schoolchildren studied as “History,” [...] where everything unexpected in its own time is chronicled on the page as inevitable. The terror of the unforeseen is what the science of history hides, turning a disaster into an epic."
“... the ongoing double-helix relationship between AI and the philosophy of AI needs to do less projection of its own maxims and instead construct more nuanced vocabularies of analysis, critique, and speculation based on the weirdness right in front of us.”
In Dialogues,
@Google
's James Manyika,
@Kent_Walker
,
@demishassabis
, and experts from a variety of backgrounds explore some of the biggest unanswered questions about AI, society, and what comes next.
Explore the digital magazine here.
15/15 Neither should they remain bulleted lists of desiderata posted on the web pages of standards bodies, governments, or corporations, with no direct connection to running code. They should, instead, become the legible and auditable “operating handbooks” of tomorrow’s AIs.
Exciting news! My latest book "Who Are We Now?" is now available online, with interactive graphs, pictures, podcasts, and more.
Tap the link here to dive in:
Had a great time discussing "The Gender Question" in Aspen alongside my fellow panelists
@fakedansavage
, Lucy Sante, and
@ahylton26
.
#AspenIdeas
Watch the full panel:
It was a pleasure to bring "Who Are We Now?" to Austin for
@FastCompany
's
#FCGrill
and
#SXSW2024
last week. Thanks to everyone for your kind words after the panel and presentation!
#FCGrill
photo by Maggie Boyd.
1/6) In my experience, all the furore over Instagram & Facebook causing teenage depression is overlooking the bigger issue: beauty filters
I mean just look at this absurdity
Huge congratulations to
@thekaransinghal
and the whole Med-PaLM team on the
@Nature
piece! I'm certain foundation models will both advance medicine as a field and improve quality of care for billions of people in the very near future.
New paper! 📜
We introduce BLUR (Bidirectional Learning Update Rules), a novel optimization framework that parametrizes forward and back-propagation via a set of very low-dimensional meta-learned matrices.
🧵 to follow...
“I’ve actually been very confused by why more people don’t agree with me that we now have AGI.”
@blaiseaguera
explains why he thinks AI models have already hit that milestone, on our “Babbage” podcast 🎧
8/15 Such doomsday scenarios were more credible in 2014, but they’re less so now that we’re starting to understand the landscape better. Language modeling has proven to be the key to making the general-purpose AI technology of the 2020s.
Had a great time on the
@TEDxManchester
stage last weekend! Thanks for having me and for all of the warm comments about "Who Are We Now?"
📸
@WeAreInevitable
Excited to present my new book "Who Are We Now?" at
@skylightbooks
this month when I'm back in LA :) February 15th at 7:00pm. More details at the link here:
@jomc
6/15 Today’s AI models can engineer drugs (or poisons), design proteins, write code, solve puzzles, model people’s states of mind, control robots in human environments, and plan strategies. Whether for good or for ill, these things are hard to dismiss as mere babble.
2/2 “-- machine learning that will take place “near” the user, on their device or home hub [...]. This will take different forms, including [...] distributed or federated learning”
@bratton
@NoemaMag
“The real lesson for philosophy of AI is that reality has outpaced the available language to parse what is already at hand. A more precise vocabulary is essential.”
Many of us are finding ourselves bewildered, unable to understand large segments of our society. Most of us do not understand, either, how alien the recent past was. This book is an opportunity to listen, learn, and place our own views in context.
12/15 For an AI model to behave according to a given set of values, it has to be able to understand what those values are just as we would— via language. By sharing language with AIs, we can share norms and values with them too.
NYC! Join Lawrence Weschler and me at
@mcnallyjackson
next month on Wednesday, May 1st for a discussion around my book "Who Are We Now?". RSVP at the link below. I hope to see you there!
5/15 On the other hand, if, as I argue in this paper, language models *can* understand concepts, then they’ll have far greater utility— though with this utility, we must also consider a wider landscape of potential harms and risks. Big social and policy questions arise too.
9/15 The result isn’t an alien with inscrutable goals. Anyone can chat with a language model, and it can respond in ways so familiar that concern has shifted from worrying about AI’s alienness to worrying about our tendency to anthropomorphize it. It’s all *too* human-like!
@AINowInitiative
@techreview
Many of us from Google were there, but more importantly are passionate on this topic + working hard on it. Frustrating headline!