Your curated collection of saved posts and media
When the AI haters get laws written to regulate AI it probably will be an AI that writes them. Irony of all ironies. At @hf0 in San Francisco I saw a company that does exactly that, and already has been used to write a bill. For our free newsletter this week, we cover machine written laws: when legislators start from AI drafts. @IrenaCronin and I write this newsletter every week. AI drafted laws could make legislation faster, cleaner, and more consistent, but they also shift power to whoever controls the models and their training data. Without strong transparency, oversight, and human responsibility, these systems risk quietly embedding lobbyist priorities, economic metrics, and hidden biases into law while pushing aside messier human judgments about justice and democracy. Read for free: https://t.co/HHwYy7MQKN and please subscribe.
NEWS: Waymo has just released an official statement following the massive power outage in San Francisco this week and their robotaxis shutting down. “We want to share an account of our operations during the outage and how we are evolving to better serve the city. Navigating an event of this magnitude presented a unique challenge for autonomous technology. While the Waymo Driver is designed to handle dark traffic signals as four-way stops, it may occasionally request a confirmation check to ensure it makes the safest choice. While we successfully traversed more than 7,000 dark signals on Saturday, the outage created a concentrated spike in these requests. This created a backlog that, in some cases, led to response delays contributing to congestion on already-overwhelmed streets. We established these confirmation protocols out of an abundance of caution during our early deployment, and we are now refining them to match our current scale. While this strategy was effective during smaller outages, we are now implementing fleet-wide updates that provide the Driver with specific power outage context, allowing it to navigate more decisively. As the outage persisted and City officials urged residents to stay off the streets to prioritize first responders, we temporarily paused our service in the area. We directed our fleet to pull over and park appropriately so we could return vehicles to our depots in waves. This ensured we did not further add to the congestion or obstruct emergency vehicles during the peak of the recovery effort. The path forward We are analyzing the event, and are already integrating the lessons from this weekend’s PG&E outage. Here are some of the immediate steps we’re taking: • Integrating more information about outages: While our Driver already handles dark traffic signals as four-way stops, we are now rolling out fleet-wide updates that give our vehicles even more context about regional outages, allowing them to navigate these intersections more decisively. • Updating our emergency preparedness and response: We will improve our emergency response protocols, incorporating lessons from this event. In San Francisco, we’ll continue to coordinate with Mayor Lurie’s team to identify areas of greater collaboration in our existing emergency preparedness plans. • Expanding our first responder engagement: To date, we’ve trained more than 25,000 first responders in the U.S. and around the world on how to interact with Waymo. As we discover learnings from this and other widespread events, we’ll continue updating our first responder training.”
The coolest thing Apple has launched in quite some time. But what is a movie? 22 pictures a second, displayed one after another. Sometime in next 18 months this will do 22 frames a second. In glasses. And then you will be in the movie. Except it will be created by a real time world model, like the one @olivercameron showed me with https://t.co/cBrAlkB91U And then you will be able to not just move around the movie, but talk with the actors in the movie. And they will be able to hand you things and take you on adventures. Will this technology save Hollywood? Maybe. Why do you think OpenAI hooked up with Disney? It is building a real time world model too, just like Oliver is building. So I predict that sometime in 2026 we will have an OpenAI pair of glasses that will let you have a light saber battle with Darth Vader. Or be able to ride in a race car going 200 mph. How can I state this with so much certainty? I've already done it. People missed what I was doing in Abu Dhabi. The @A2RLeague is building exactly that and it's sick. And it also is racing autonomous drones: https://t.co/PCOYkdcUbi At the Formula One track I discovered that you can't see most of the track, even from the expensive seats. When people figure out this is a way better way to watch the race than looking at the 30-foot-wide TV screens on the stands they all will bring their devices to the track. And who owns @F1 racing media rights? Apple. So my prediction is that in 2027 I will buy an Apple headset to experience all this. Thank you @3duaun for all the secret, non-recorded, audio spaces you and I have done together the past four years where we predicted just all this going down. Thousands of them. Our dreams will come true in the next 18 months and people have no idea how profound the change will be on not just sports, but education, manufacturing, health, real estate, shopping, and literally everything humans do. All year next year I'll be posting "the Holodeck is coming" and as the year progresses please DM me more examples so I can share them with everyone. And Dylan's team showed me football in the holodeck. The camera takes you right onto the football as it leaves the quarterback's hands. Sick.
She just keeps getting the #ratio. She is stupid enough to actually hashtag “NativeVote” which immediately let’s people know she’s pandering. My gosh. https://t.co/c0G5Wkms4S
https://t.co/ly8vxkqcM7
https://t.co/ly8vxkqcM7
10 years of residency and fluency in Japanese. Japan First Her approval rating is 75% in Japan. https://t.co/MibaMeP0AE
@scottbelsky A trend I’m seeing more is immigrant-founded restaurants hosting cooking classes It’s a tangible way to share your culture and ensure your family recipes/techniques become a part of your lasting legacy I took a tamale class last week and learned a few different regional styles https://t.co/vslh0CmF1a
You can now fine-tune LLMs with Unsloth then deploy them in @LMStudio! 🦥👾 We made a free notebook to fine-tune FunctionGemma (270M) so it “thinks” before calling tools, then export the model to GGUF for deployment in LM Studio. Notebook: https://t.co/3XoGxbDAuA
We worked with @UnslothAI on a new beginner's guide: How to fine-tune FunctionGemma and run it locally! 🔧 Train FunctionGemma for custom tool calls ✨ Convert it to GGUF + import into LM Studio 👾 Serve it locally and use it in your code! Step-by-step notebook: https://t.co/JkR
This AI agent just surpassed humans for the first time ever Simular open-sourced Agent S, and scored 72.6% on OSWorld bench vs 72.36% humans, automating your desktop and complex workflows 6 wild examples + how to try: 1. Turn research paper to X thread with assets https://t.co/jdB1TRHa3m
Open NotebookLM is INSANE! Fully Free Local NotebookLM Alternative with Gemini Integration: https://t.co/t4mgF3ryCv https://t.co/IB69nEEyJd
Open NotebookLM is INSANE! Fully Free Local NotebookLM Alternative with Gemini Integration: https://t.co/t4mgF3ryCv https://t.co/IB69nEEyJd
What if an 8B model could out-reason GPT-5 on math? Introducing PaCoRe. It breaks the sequential bottleneck. Instead of thinking step-by-step, it launches hundreds of parallel reasoning "threads" in each round, compresses their insights, and synthesizes them to guide the next round. The result? It scales compute to millions of tokens without hitting context limits, outperforming frontier systems. The 8B model scored 94.5% on HMMT 2025, beating GPT-5's 93.2%. PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning StepFun, Tsinghua, Peking Paper: https://t.co/qn8ShtFAwR GitHub: https://t.co/MHanDNRBYB Hugging Face: https://t.co/aN4hgyO7bE Our report: https://t.co/08pf1rxOug 📬 #PapersAccepted by Jiqizhixin
One of the things I love about @united is how good their iOS app is. The latest version has real-time boarding with a passenger progress bar, smart lounge guidance based on your gate and capacity, and more detailed live bag tracking. Lots of thoughtful little details ✈️ https://t.co/IrsVSgH36O
🚨Scientists just reprogrammed leukemia to self-destruct – and it worked. In a major breakthrough, scientists at Institut Pasteur have developed a therapy that forces leukemia cells to self-destruct—and alerts the immune system to wipe out the rest. The team targeted malignant B-cell leukemia with a triple-drug combination that reprograms cancer cells to undergo necroptosis, a form of inflammatory cell death. Unlike the silent shutdown of apoptosis, necroptosis creates an immune alarm, drawing in the body's defenses. Using real-time imaging, researchers watched immune cells swarm the cancer, leading to total tumor elimination in lab models. The challenge was that B-cell cancers typically lack a key protein, MLKL, needed for necroptosis. But the team cleverly sidestepped this using three existing clinical drugs. Together, they bypassed the missing protein and reactivated necroptotic pathways. The result: not just tumor shrinkage, but complete disappearance in multiple preclinical models. While human trials are still to come, the findings hint at a new kind of cancer therapy—one that doesn’t just kill tumors, but trains the immune system to join the fight. And because the drugs are already approved, the road to real-world use could be much shorter. Source: Le Cann, F., et al. (2025). Reprogramming RIPK3-induced cell death in malignant B cells promotes immune-mediated tumor control. Science Advances.
You were right @Scobleizer https://t.co/zuSMreiBwY
Qwen-Image-Edit-2511-Lightning https://t.co/myPYmIQ5ET
Capability overhang means too many gaps today between what the models can do and what most people actually do with them. 2026 Prediction: Progress towards AGI will depend as much on helping people use AI well, in ways that directly benefit them as on progress in frontier models themselves. 2026 will be about frontier research AND about closing this deployment gap — especially in health care, business, and people's daily lives.
Because most of my fellow Europeans haven't had the time nor stamina to fully acquaint themselves with the complete spectrum of insanity that is the current U.S. administration, I've taken it upon myself to offer a brief, honest introduction to its main characters.🧵 https://t.co/kGqm6LY9N5
HUGE BREAKING: Data Set 8 of The Epstein Files shows that a victim stated that Trump & Epstein raped her, and that a Limo driver heard Trump in the back speaking to “Jeffrey” about “abusing girls.” The victim then “committed suicide” in January of 2000 after reporting the incident. Police didn’t believe it was suicide. Pls share far & wide!
Along for the ride in unsupervised FSD testing https://t.co/AsyBHnndj8
the spirit app wasn’t loading but poke was https://t.co/PZagjAT5ER
WorldWarp Propagating 3D Geometry with Asynchronous Video Diffusion https://t.co/7CvB39UGvG
ULMFiT was really ahead of its time, complete with the pre-train -> mid-train -> SFT pipeline we use today https://t.co/ntOnsX6R6P
2017 Pre-training (ULMFiT) 😊
https://t.co/GEOZunWoXj alums and ULMFiT users like Jason have been fine-tuning LLMs since 2017/18. The rest of y'all are years behind, sorry. 🤷
@jeremyphoward 2018 Fine-tuning ULMFiT on song lyrics (DeepLyrics, grad school project) Watching it learn to rhyme was when I knew scaling was going to work
I bike everywhere in SF. I barely ever take a taxi/uber/waymo. But if you want to ban Waymo it means you don’t care about cyclists like me. https://t.co/BaQUdVpMjG
https://t.co/OrXsuh4Mrs
It was the closest race yet when it came to the top programming language of the year. 👀 https://t.co/GKDBVjiI0S
von Neumann gave us "zero-sum" in 1944 and it's one of my favorite examples of linguistic compression. Two words that mean “strictly competitive, fixed pie, stop looking for win-wins.” The term's real gift: it reveals the default human assumption that life is chess and value can’t be created—only transferred. Once you have the label, you start noticing how often that’s wrong
If you support Native American people’s, history & culture 🥰Say.. “Yes https://t.co/DmIpIrRUUn
☕️ @LatentLabs_, Outpost Bio and #AWS welcome you for breakfast on Jan 14 at #JPM2026. Hear from @biogerontology & discover how #AI is transforming drug discovery. 👉 https://t.co/aRQtFz2scX 🧬 Connect with leaders & VCs to explore the next generation of translational tools & biological FMs.
What if the safest driver on the road isn’t a human, but Tesla’s AI? LikeFolio's @AndySwan says Tesla FSD is already far safer than human drivers, in some cases multiple times safer! The problem? Tesla barely markets it. If people understood that pressing a button and letting AI drive is safer than driving yourself, it could be a game-changer by 2026–2027. $TSLA