Your curated collection of saved posts and media
No More Typing: I've Been Talking to 5 AI Chatbots. This One Really Stood Out https://t.co/Eg9cCQ8aEQ @lancewhit @pcmag
Opinion | The AI cheating panic is missing the point https://t.co/xdJw6NpyuW @washingtonpost
ChatGPT isnβt just for cheating anymore https://t.co/euM3QVlqEC @adamclarkestes @voxdotcom
How We Chose the @TIME100 AI 2025 https://t.co/XThlyQlcsE @sampjacobs
MIT researchers develop AI tool to improve flu vaccine strain selection https://t.co/7Ud8cpxhXp @mit @MIT_CSAIL
China Warns Against βDisorderly Competitionβ in Booming AI Race https://t.co/fEV4FP7za4 @business
βThe Wizard of Ozβ Is Getting an A.I. Glow Up. https://t.co/s90atpESRA @brooksbarnesNYT @nytimes @SphereVegas @google https://t.co/K6dOpn4qxs
βAI psychosisβ: could chatbots fuel delusional thinking? β podcast https://t.co/Xy2BEm1v5x
In case a full day of splines isn't enough to make you book your trip to JMM26... Perhaps a 90min session with @ylecun (+ TBD speakers) on self supervised learning/world models WITH math will close the deal! No matter your background, come in number to discuss research together! https://t.co/dh1aJLzMgN
Interested in splines for AI theory (generalization, Grokking, explainability, generative modeling, ...)? Wait no more! We are organizing a dedicated session at JMM26 (the largest math conf.): https://t.co/ccCNDvQIfM Consider submitting your abstract! Deadline in 2 weeks! https:/
Single vector models have **fundamental** limitations You can try to make the embedding dimension larger and larger if you want, you'll eventually hit a wall Rerankers mitigate those issues but cannot scale for full databases ColBERT does both. https://t.co/fLgyqzPFvV
On the Theoretical Limitations of Embedding-Based Retrieval @orionweller et al. at Google DeepMind demonstrate that vector embeddings have fundamental limitations in representing all possible document combinations. πhttps://t.co/jpV9MpegUY π¨π½βπ»https://t.co/aTA8Aiz4UH
This should be a wake-up call for US policymakers: https://t.co/Q2QZoRvfeC
This should be a wake-up call for US policymakers: https://t.co/Q2QZoRvfeC
Thank you for all the support, folks! If there's one thing some people seem to want me to do differently it's going into more depth. I'll think about how to do this in future videos. The format precludes the same level of depth that I have in my newsletter with @sayashk but I might include additional notes in the text transcript that I publish along with the videos. https://t.co/7zC9o4GrG9
Iβm excited to announce Iβve started a YouTube channel. I plan to publish videos regularly explaining my views on AI and its present and future impacts. My first video asks: What happens if thereβs an AI crash? https://t.co/7zC9o4GrG9 This is my first foray into video (beyond
Video: In a talk at the Neuro-Symbolic AI summer school, I argued that research fields progress in explore/exploit cycles, and that while the AI community is very good at the exploit phases, it is not so good at the explore phases and tends to get stuck in local minima. Progress toward AGI (though I don't use that term) will be helped if we have strong sub-communities that have different progress criteria from mainstream AI/ML and are big and powerful enough to support scholars' career advancement. https://t.co/q1mdjCHK91 (link goes to timestamp starting at 51:10) The above is Part 1 of the talk; Part 2 is based on this essay with @sayashk and is about the application of AI to other scientific disciplines. https://t.co/aycitbnKzW Thanks to the Centaur AI institute for inviting me and for the participants for a great discussion.

Testing Vega on @baseapp Blackjack straight from your wallet Dangerous π¬ https://t.co/JUJop5N1zw
investment research analysts are cooked. this research agent took 5min to build (demo for a client) and probably costs 1 latte per day link below https://t.co/dmPAq8vT0U
@apoorvasriniva https://t.co/qRk2Fs5kvU
I have no inside knowledge but here are my uninformed guesses about the @Tesla roadster many billionaires have been waiting for for years. It is a controllable wing. People guess that it is for flying. It is, but usually will be in upside down mode. The air jets feed a device and the wing. Faster flowing air on top of a wing increases lift. Put it upside down and it increases downforce. Tesla has a patent that shows the device and how a turbine fan in the frunk sucks so much air out, increasing downforce that it sticks to the ground, even at high speed, like no other automobile can do. NASCAR legend Brett Bodine gave me a ride at high rate of speed to demonstrate how he thinks. A lot of that is about downforce. I guarantee you that if he gives you a demo of his skills you will remember the ride fondly for the rest of your life. But when he is in the trailer he is talking to his team mostly about downforce. What are AI Agents for? To get more downforce so your tires will be pressed strongly into the asphalt they are on. The more downforce, the better. The more fun you are having. The more likely you are to win a race. I got to drive behind this 1959 Jaguar. It was the first one to have disk brakes and won many races, its owner told me. Current value $30 million. All the other billionaires were jealous. So letβs start with jealousy. When I get me first ride in one. I expect the same experience. But on the second ride Grok will say βwhy donβt you drive, I will help you go faster.β βOr you can tell me to drive this like we were at Daytona.β What does @elonmusk get right? He goes after the little kid in us who wanted to race our Big Wheels with our best friend. The real secret sauce is the variety of different styles of talking a path down the road. Tesla AI team will be showing off a bunch, I predict. Be aware of the Formula 1 mode. It will push the tires to the edge of traction the whole ride around the track the way Brett Bodine did for me. The NPR radio personality in the front seat was screaming the entire time. Since it is a vehicle that will compete with a $450,000 Ferrari. A Ferrari makes love to you at about 120 miles per hour and live gets better every mile above that. The Tesla can do that in its steer by wire steering system. But then you can take your hands off the wheel and the experience turns into a Disneyland ride. But one that is 100x more fun than even Tron in the new Shanghai one is. Inside it will be quiet with a new kind of audio system. One that turns the vehicle into an audio Holodeck. It can turn every centimeter into a virtual audio speaker. I first heard this at Virginia Tech in a cube building its augmented reality department built. It is a very large single room building with 1,600 speakers inside. They put me into a football game and if I closed my eyes I swore I was in a real football game. The audio team has AI that gets about 30 speakers much closer to that than any other vehicle. I wonβt be shocked if Elon says that only people who wear glasses in it can drive it. Same rules for F-35 fighter jets. They are wearing @LumusVision displays in front of their eyes. Same as Meta will announce with Qualcomm, who built the chips inside, and Essilor, who owns Oakley that did the design and the production/manufacturing. At least there should be an API so developers can make it all work. If they do the car will turn invisible. You will only see the forest/road around you. OK, OK, the Metas arenβt good enough to do that all today. But they are a big step up and will turn the inside of the cabin into a visual Holodeck as well. So now you have an amusement park ride inside an amusement park ride. I have been shoved off the deck of an aircraft carrier by a steam-driven catapult. Tesla AI will do the same in the Roadster. If you had something like an Apple Vision Pro running Google Genie 3 it could make you believe you were controlling an F-18 instead of driving a robot.
However big you think solar will be, itβll be bigger than that. https://t.co/Z1O4enrOPG
This should be a wake-up call for US policymakers: https://t.co/Q2QZoRvfeC
However big you think solar will be, itβll be bigger than that. https://t.co/Z1O4enrOPG
500+ hours of real world manipulation data, covering residential, kitchen, retail and office settings. A important step towards generalized manipulation models! Great work Galaxea team! https://t.co/zgRFLxO5DM https://t.co/vZtFfDWr3r

We shipped 3 quality-of-life updates: - Markdown rendering in Playground, more useful for looking at markdown outputs when comparing LLMs. - New Trace Graph View for visualizing trace trees. - Time-to-first-token metric to track LLM latency. All aimed at faster debugging. https://t.co/kpG8hyd2cB

π Assign tasks to GitHub Copilot coding agent and track its progress using our new @raycast extension! π Now available on macOS. π» Get more done, more efficiently. Learn how.π https://t.co/GWecwTl4dY
Game on at #GitHubUniverse πΎ Want to play at home? You might want to Inspect the GitHub Universe website π https://t.co/Wbnkex9695 https://t.co/4boVCcwekD

Behind every AI breakthrough and launch, there are hundreds of Googlers working to build AI thatβs helpful for everyone. This year, our very own @JoshWoodward VP of Gemini and Google Labs, @JeffDean Chief Scientist at DeepMind and Google Research, and Hartmut Neven founder and lead of Google Quantum AI have been recognized on @TIMEβs #TIME100AI, a list recognizing the 100 most influential people in artificial intelligence. Find the full list and learn how weβre thinking about the future of AI at https://t.co/qDi7b2FRBQ

80% of all crimes are committed by people with 3+ prior arrests. Most people are not criminals. A tiny minority are antisocial psychopaths who commit most crimes. This is some of the lowest-hanging policy fruit there is. https://t.co/6N7Zj0VpbC
the actual reason we "can't incarcerate our way out of crime" is because literally every American commits dozens of crimes every single day. Speeding, jaywalking, drinking in public, etc. The people who think they are most hardcore about "crime" are people who think that their cr
Liftoff of Super Heavy, the most powerful launch vehicle in history, on Starshipβs tenth flight test https://t.co/dnQQAqKWUf

In 2027, we WILL return American astronauts to the Moon. Β We won yesterday's space race. We'll win today's space race against China, and we'll always win tomorrow's space race. Β @SpaceX's Starship test flight success moves us one step closer toward achieving that goal. https://t.co/VaYVN6x3sI
CNN also fails to report shooter was trans. When news organizations routinely omit relevant facts about an incident to protect sacred ideological cows, that is not journalism. People don't trust the media for good reason. https://t.co/9YoPQMXlnP
Mainstream media's review of @Tesla's FSD (Supervised) after using it public roads in Australia: "To have this car drive me around Brisbane for an hour, we're talking in the city, motorway, spaghetti bowl of on-ramps, it handled it so well. It was mind-blowing." https://t.co/Rjsh6yXk93