founder @GeneralAgentsCo | previously autopilot @tesla, deep learning @googledeepmind, phd mila.quebec, cs @iitdelhi
Joined August 2009
- Tweets 2,068
- Following 3,699
- Followers 12,207
- Likes 6,261
Pinned Tweet
Today I'm launching my new company @GeneralAgentsCo and our first product.
Introducing Ace: The First Realtime Computer Autopilot
Ace is not a chatbot. Ace performs tasks for you.
On your computer. Using your mouse and keyboard.
At superhuman speeds!
I feel like captain obvious but LLMs obviously predict actions AND the consequences of those actions, and thus are world models, but in the world of text.
LLMs don't have goals. But chatbots and coding agents do. Their goal is to be useful to their user, as judged by the user.
over the past few months i've seen things that make me confident we have an alternative path to agi..
if you're an extremely cracked engineer and generalist passionate about computer use,
reply with your most impressive work and i'll interview you this week
Sherjil Ozair retweeted
A while back, I was talking with a friend who'd moved to the US a few years prior
At some point, I said something like "... we Americans..." and he froze with shock. He asked, "Wait, do you think of me as American?" Without much thought, I said "Of course!"
He paused for a while and then said, "Wow. Back home no one would ever refer to a recent immigrant that way. Heck, they wouldn't refer to any immigrant that way, even if they'd lived there for decades"
While the US has a lot of problems (many of which are our own making), this moment made me swell with joy and pride for this special place
Happy birthday, America. I'm so proud to be the citizen of a country where, if you choose to build your life here and contribute to your community, you can become an American too 🇺🇸
Coding models basically don't work if you're building anything net new. Vibe coding only works when you split down a large project into components likely already present in the training dataset of coding models.
Pretty much the entire team of @GeneralAgentsCo was hired via Twitter DMs. If you are an exceptional engineer and want to build computer AGI, slide into my DMs!
This also negates a lot of AGI x-risk concerns imo.
Typical safety-ist argument: RL will make agents blink past human-level performance in the blink of an eye
But: the current paradigm is divergence minimization wrt human intelligence. It converges to around human performance.
A lot of people presume we use reinforcement learning to train Ace. The founding team has extensive RL background, but RL is not how we'll get computer AGI.
The single best way we know how to create artificial intelligence is still large-scale behaviour cloning.
PSA: agents acting in an environment is *not* reinforcement learning
reinforcement learning is about having a reward signal for which you get reinforcement (hence the name)
if your "reward" is 99.99% "predicting accurately the consequence of an action" (which is just regular unsupervised learning) and 0.01% some additional specific goal (which is actual RL), then calling the training procedure "reinforcement learning" is technically accurate but is very much a sin against the truth
RL always was and will always remain "the cherry on the top" due to fundamental information reasons
a sparse 1d reward signal just doesn't have enough information to train complex agents with trillions of parameters in complex enough environments, whereas just predicting the outcome of every action is a maximally dense feedback signal from the environment in terms of the information it provides
I really find it somewhat offensive that RL people try to bucket everything that is about agents acting in environments into the RL bucket, because if you are slightly less in the weeds then you buy this and you eventually have an incorrect and imprecise understanding of the world
The best thing about supervised learning is that you can clone any static system arbitrarily well. The downside is that supervised learning only works for static targets.
This is an AI koan and also career advice.
I can't believe that we'll soon have drop-in remote workers, indistinguishable from a human worker, and it will be trained purely on behavioral inputs/outputs. Total behaviorism victory!
Sherjil Ozair retweeted
this thing is so fast. custom trained model for computer use.
Today I'm launching my new company @GeneralAgentsCo and our first product.
Introducing Ace: The First Realtime Computer Autopilot
Ace is not a chatbot. Ace performs tasks for you.
On your computer. Using your mouse and keyboard.
At superhuman speeds!
Part of the inevitable future 👀
Today I'm launching my new company @GeneralAgentsCo and our first product.
Introducing Ace: The First Realtime Computer Autopilot
Ace is not a chatbot. Ace performs tasks for you.
On your computer. Using your mouse and keyboard.
At superhuman speeds!
Sherjil Ozair retweeted
1/ Last month i riffed with @ShaneAParrish of @farnamstreet on
the ‘speed limits’ of biology
-vs-
the ‘speed limits’ of technology…
“…In biology you’ve got all kinds of reactions, nature figured out, evolution figured out, enzymes and catalysts and things that can speed up reactions, but you can’t move faster than the speed of biology.
Now you think about AI. Totally different field, but the same sort of underlying philosophical principle. If you’ve tried ChatGPT Operator, it can only move at the speed of the web. And at that, it’s sort of with latency a little bit slow. So we’re thinking about, What are the technologies that can accelerate these things that have these natural almost-physics limits, even if those limits are biological or digital? That is something that at the moment I’m obsessed with, in part because I have ignorance about it. And when I have ignorance about it, my competitive spirit says, How do I get smart about this? How do I get some differentiated insight? How do I know what everybody else thinks? And then find the white-space answer. So that’s one big thing….”
"I think the most valuable thing that AI is going to do, when you ask it questions and it comes up with the answers, assuming those answers are accurate and cross-correlated and double-checked, [is going to be when] they actually say, “Here are the five questions you didn’t ask.” That is going to unleash real insight."
Episode 217 of The Knowledge Project with @wolfejosh is out now and jam-packed:
(3:27) AI and User Interface Evolution
(6:10) AI, Memory and Inference
(9:43) Future of @nvidia
(11:42) Human Intelligence Limitations
(13:42) AI's Threat to Elite Professions
(16:32) AI and Coding
(17:55) AI and Business Margins
(20:02) AI and Data Monopolies
(26:24) AI and Science
(30:22) AI in Experimental Design and Automated Labs
(31:47) AI in Patenting and Prior Art
(33:47) @elonmusk and Tesla: A Critical View
(39:10) AI in Investment Analysis
(42:17) AI and Information Advantage
(43:42) Privacy
(46:56) Potential Changes in Market Structure
(48:00) Shift Towards Retail Investment and New Financial Vehicles
(50:09) AI's Impact on Creativity and Art
(54:31) AI's Influence on Future Professions
(55:12) The Importance of Human Connection
(1:03:33) The Possibility of Extending Human Lifespan
(1:12:45) The Use of Technology in Filtering Information
(1:16:58) AI in Personal Life and Privacy
(1:24:19) AI and Military Technology
(1:30:35) AI, Ethics and International Competition
(1:31:13) Weaponization of Social Media and Information
(1:33:52) Vulnerability of Infrastructure to Attacks
(1:36:42) Views on #DOGE Initiative
(1:38:25) Policy Suggestions for Global Competitiveness
(1:42:56) Attracting and Retaining Talent
(1:45:40) AI, Copyright, and Intellectual Property
We're hiring for research engineers, product engineers, software engineers, and agent trainers at @GeneralAgentsCo!
Join us if you believe that computers should work for us, and the way to make it happen is by training neural networks to use computers like we do!
generalagents.com/careers
We're hiring! Check out our open roles here: generalagents.com/careers/
Good morning. This is the future we're building.
Today I'm launching my new company @GeneralAgentsCo and our first product.
Introducing Ace: The First Realtime Computer Autopilot
Ace is not a chatbot. Ace performs tasks for you.
On your computer. Using your mouse and keyboard.
At superhuman speeds!
Absolutely blown away by the magnitude of engagement and interest in using Ace.
Proves our thesis that desktop computer control is something a lot of people really want.
We're working hard to give access to Ace to as many people as we can.
Today I'm launching my new company @GeneralAgentsCo and our first product.
Introducing Ace: The First Realtime Computer Autopilot
Ace is not a chatbot. Ace performs tasks for you.
On your computer. Using your mouse and keyboard.
At superhuman speeds!