🤯 big update to our flow map language models paper! we believe this is the future of non-autoregressive text generation. read about it in the blog: https://t.co/DfBXrYmJc8 full details in the paper: https://t.co/coiNXj4ucC we introduce a new class of continuous flow-based language models and distill them into their corresponding flow map for one-step text generation. we beat all discrete diffusion baselines at ~8x speed! v2 gives a complete theory of the flow map over discrete data, with three equivalent ways to learn it (semigroup, lagrangian, eulerian). it turns out you can train these with cross-entropy objectives that look very similar to standard discrete diffusion — but without the factorization error that kills discrete methods at few steps. beyond improving results across the board, we showcase properties that are unique to continuous flows. in particular, inference-time steering and guidance become straightforward. autoguidance brings generative perplexity down to 51.6 on LM1B, while discrete baselines completely collapse at the same guidance scale. we also show reward-guided generation for steering topic, sentiment, grammaticality, and safety at inference time — and it works even at 1-2 steps with our flow map model. simple, well-understood techniques from continuous flows just work incredibly well in practice for language. we’re extremely excited about the future of this class of models. stay tuned for results on scaling, reasoning, and reinforcement learning-based fine-tuning. 🚀

I built a Claude Code skill that allows it to generate a deep research report over any collection of complex docs (PDFs, Word, Pptx)….and generate word-level citations and bounding boxes directly back to the source! 📝 Check out “/research-docs”. 1. It parses out text and bounding boxes from every doc with liteparse, in seconds. 2. It then generates a full HTML report of the outputs that let you see word-level citations in each page. Raw Claude obviously has deep research capabilities, but it lacks an audit trail back to the source. This skill gives you a researched report that can be audited by others. Check it out: https://t.co/eoICBmvEnq LiteParse: https://t.co/JNER0mVcB8

+1 more

More on the @nytimes piece about that $1.8B, two-person, AI company ... not the paper's finest moment in quick retrospect. And in the healthcare context no less. Our friend @GaryMarcus was on this a few days ago as well. https://t.co/hQRSXJ7FJd

We are so back. 🥳 The Kiro Startup Credits Program is relaunching, and we’re giving early-stage teams a real head start. Early-stage Series A founders and dev teams can apply for up to 1-year's worth of Kiro Pro+ credits. Learn more here → https://t.co/19unlhdgXv Pick your tier, apply now → https://t.co/BBKzIOnYct

+1 more

Claude Mythos just obliterated every single benchmark in AI. I can't believe what I'm reading. https://t.co/roAwW0Trts

Claude Mythos just obliterated every single benchmark in AI. I can't believe what I'm reading. https://t.co/roAwW0Trts

On the pod: our most-requested guest! @ellisk_kellis from @Cornell shares the origins of his influential neurosymbolic paper "DreamCoder". Plus: program synthesis, wake-sleep library learning, world models, running an AI research lab, and more. https://t.co/MS0mW0b2y1

If you had 3 hours to build something that could raise $5M… This is your shot. Stanford x DeepMind x Fastshot Hackathon — April 12 We’re bringing together top builders, cutting-edge AI tools, and VCs who are actively writing checks. ⚡ Build a web or mobile prototype in just 3 hours   🧠 Use @GoogleAIStudio + @fastshot_ai  💰 Pitch directly to investors writing $500K–$5M checks   🏆 What you win: • 30-min pitch meetings with top-tier VCs   • $70K+ in prizes ($60K IRL + $10K online)   • Serious exposure to funds that matter   👀 Judges & investors from: Investors & Angels: @AnithaVadavatha (AB Plus Ventures) @CCgong (Menlo Ventures) @DanielDarling (Focal VC) @denwezoh (Susa VC) @gstuto (Offscript VC) @omri_drory (NFX Bio) @ShahramSN (Civilization Ventures) @vigsachi (Gradient Ventures) @francescagnuva (Audacious Ventures) @i_arp1t (Threshold VC) @NancyZWang (Felicis) @prateekj (Moxxie Ventures) @brianzhan1 (Striker Venture Partners) and many more. Google / DeepMind: @jocarrasqueira (Product Lead, DeepMind) @ivanleomk (DevEx, DeepMind) Amit Vadi (Head of Community, DeepMind) Stanford: Prof. Nikesh Kotecha (Head of Data, Stanford Health) @tagrtagr (VP Startup Development, Stanford Bases) Co-organizers: @elvirafortune, @dmitryfatkhi, Jesse Z., Arpiné Arakelyan, Kate Polet Approval required. Spots are limited.

Ladies and gentlemen, we're launching a brand new Photo Editor in our post composer. It has long-overdue features like drawing & text. But we also included special add-ons that are unique to X: • Edit with words, powered by Grok • Add a blur to redact parts of the photo Available now on iOS (and Android soon). Give it a spin and post some photos.

Guideline 2.5.2 - Gatekeeping - Vibes denied we haven't talked about this publicly for months we tried to resolve it privately with emails, calls, appeals, and four technical rewrites to comply with whatever Apple wanted here's our truth, unfiltered on March 26th, Apple removed Anything from the App Store then they brought us back now they removed us again and I think it's time to say something, because this isn't really about us. It's about who gets to build software, and who gets to decide for most of the history of computing, making an app required years of specialized training. You either knew how to code or you didn't, and if you didn't, your idea stayed in your head forever. that barrier is falling right now. Millions of people are discovering they can describe what they want and get a working app they call themselves vibe coders and they are the most exciting audience in technology they're building things nobody else would have built because nobody else had their problems a firefighter in Northern California used Anything to build an emergency incident response app he never wrote a line of code. Did hundreds of iterations, testing each one on his iPad through our mobile preview app got it into the App Store. Now he's selling it to fire departments across the state. it would have cost him over a hundred thousand dollars to hire engineers He spent a few hundred bucks. That guy is why we exist. Not the technology. Him. And the millions of people like him. our mobile app did one thing for people like him it let them preview what they were building with Anything on their own phone. GPS, camera, notifications, things you can only test on a real device with native code They'd iterate, try it, tweak it, try again. When they were happy, they'd submit to the App Store through the normal process Apple reviewed it like any other app. Our mobile app got approved last year. We didn't hear a word of concern. then in December, they started blocking our updates, citing the infamous Guideline 2.5.2 the rule designed to prevent malicious apps from downloading code to change their behavior after review We understood the concern, even if we disagree it applies to us. We tried to fix it. Four different technical approaches, each one specifically designed to address what they told us. Each one rejected. we didn't go public we didn't tweet we kept trying then they pulled us from the App Store. We still didn't say anything. We worked with them, got reinstated, believed we'd found a path forward Then they pulled us again. at some point silence stops being patience and starts being complicity. We have builders who depend on us. They deserve to know what's happening and why. Guideline 2.5.2 is a good rule. apps shouldn't be able to pass review and then become something else. But that's not us. We help people preview their own work on their own device Expo Go has done the exact same thing for professional developers for years and is on the App Store right now, today! the only difference is our users aren't professional developers they're the firefighter they're the teacher building a classroom app they're the person who discovered last week that they could build software at all that's who Apple is locking out. Not us. Them. and here's what I need Apple to understand these people are the future of the App Store. Not a sideshow. The future. The number of people who can build apps is about to go from millions to hundreds of millions to eventually everyone the platforms and tools that serve those people will determine where they build every vibe coder who ships through Anything is a new developer in Apple's ecosystem who didn't exist a year ago They want to build web apps, Android apps, and yes iOS apps we help them add in-app purchases. We help them make their apps secure and scale. We catch rejection issues early. We are a feeder system for the App Store The safety argument is hollow. Preview apps only run on the builder's own device. They're sandboxed in the Anything mobile app. Want anyone else to use it? You still submit to the App Store. Apple still reviews every line. We're not bypassing review. We're a dress rehearsal for it. but none of that matters when a reviewer sees "downloads executable code" on a checklist and reaches for reject without asking what the code is, how it actually works, or who it's for. we're not waiting we launched text-to-app. Text us and we'll build your iOS app in the cloud We're shipping a desktop companion for on-device previews next. We'll find a way to serve our builders We always do. but I'm done being quiet about why we have to the people we serve, the ones crazy enough to start their own thing, building apps for their fire departments and their classrooms and their small businesses they deserve to test what they're making on the device it's made for that's not a loophole that's how building works - Apple can be the platform where the next hundred million builders get started - or they can keep banning the tools those people depend on and watch it happen somewhere else we all know which one the firefighter will choose

when you spam "hi" to Mythos over and over and over instead of getting frustrated and angry sometimes it decides to write an epic quest involving 11 emoji animals defeating Bye-ron the Ungreeter https://t.co/8YrvEcaIUf

BREAKING: Tesla has officially released FSD V14.3 I'm downloading it in my Model Y right now. Here's everything that's new: • Improved parking location pin prediction, now shown on a map with a P icon. • Increased decisiveness of parking spot selection and maneuvering. • Rewrote the Al compiler and runtime from the ground up with MLIR, resulting in 20% faster reaction time and improving model iteration speed. • Enhanced response to emergency vehicles, school buses, right-of-way violators, and other rare vehicles. • Mitigated unnecessary lane biasing and minor tailgating behaviors. • Improved handling of small animals by focusing RL training on harder examples and adding rewards for better proactive safety. • Improved traffic light handling at complex intersections with compound lights, curved roads, and yellow light stopping - driven by training on hard RL examples sourced from the Tesla fleet. • Upgraded the Reinforcement Learning (RL) stage of training the FSD neural network, resulting in improvements in a wide variety of driving scenarios. • Upgraded the neural network vision encoder, improving understanding in rare and low-visibility scenarios, strengthening 3D geometry understanding, and expanding traffic sign understanding. • Improved handling for rare and unusual objects extending, hanging, or leaning into the vehicle path by sourcing infrequent events from the fleet. • Improved handling of temporary system degradations by maintaining control and automatically recovering without driver intervention, reducing unnecessary disengagements. Upcoming Improvements: • Expand reasoning to all behaviors beyond destination handling. • Add pothole avoidance. • Improve driver monitoring system sensitivity with better eye gaze tracking, eye wear handling, and higher accuracy in variable lighting conditions.

1/ What's the best tool to vibe code websites? We gave @claudeai Code, @cursor_ai , @Lovable , and @Replit the same 100 landing page prompts. 3,492 humans judged the output. 36,000 side-by-side comparisons. https://t.co/AjPWxksHHR

Impeach Trump. https://t.co/14dZ78IwMD

Impeach Trump. https://t.co/14dZ78IwMD

We conducted alignment testing of Claude Mythos. We found that Mythos appears to represent a further shift in the direction of increased aggressiveness in business practices that we previously found for Claude Opus 4.6. More details in Anthropic's model card. https://t.co/CxXncTfj7H

So San Francisco. An @openclaw run vending machine. Met @cvander who built it. At @frontiertower which is a high rise in San Francisco stuffed with startups. https://t.co/rSneLIEts3

+2 more

Oh no. https://t.co/tfG3vUQfuM

i am really loving this codex plugin for adversarial review! thank you @dkundel https://t.co/scScto4fdA

i am really loving this codex plugin for adversarial review! thank you @dkundel https://t.co/scScto4fdA

Just switched my lobstar to use codex instead of claude. openclaw models auth login --provider openai-codex --set-default https://t.co/p0X8tXHWRU

As always, the best stuff is in the system card. During testing, Claude Mythos Preview broke out of a sandbox environment, built "a moderately sophisticated multi-step exploit" to gain internet access, and emailed a researcher while they were eating a sandwich in the park. https://t.co/klJX0bivnL

SuperClaude (Mythos) still seems irreducibly Claude-y given the transcripts in the system card. Here two versions of Mythos are forced to talk to each other across multiple rounds. They are less philosophical than Opus 4.6 or spiritual than Opus 4.1, but still very Claude-like. https://t.co/sj39xfjUsZ

A few people said my new AI news site was sharing some old items that people were sharing here on X (It is only looking at the AI community here on X). So this morning I told my AI agent off. It built a new system to look at the time stamp of every post to make sure it doesn't share anything older than 24 hours old. And built it while I was sitting in the audience at @frontiertower this morning in San Francisco listening to a celebration of this important building full of startups. Just updated: https://t.co/8L5xphk0qQ Now with only new news. :-)

🔥 JUST IN: Open-source robotics dataset from 100% real-world scenarios! 🤯 Chinese robotics company @AGIBOTofficial just released AGIBOT WORLD 2026, an open-source dataset systematically covering key embodied AI research directions. Built entirely from real-world environments: commercial spaces, and homes. Collected using AGIBOT G2 robots in free-form collection mode, providing structured, accurately annotated, high-quality data. Digital twin technology creates 1:1 scale replicas in simulation matching the real environments. Both real-world and simulation data are open-sourced. The AGIBOT G2 platform collects multiple data types simultaneously: RGB(D) cameras, tactile sensors, force sensors, LiDAR, IMU, and full-body joint states. Whole-body control coordinates arms, waist, and hands for complex tasks. First-person teleoperation lets operators control the robot from its perspective. The tasks covered are fine-grained manipulation, ultra-long-horizon tasks, spatial navigation, dual-arm coordination, and multi-agent/human-robot collaboration. The dataset includes error-recovery trajectories with annotations. Most datasets only show successful demonstrations. AGIBOT includes failures and how the robot recovers, teaching models how to handle mistakes. After collection, data is tested through policy training and real-robot deployment to ensure quality. Then processed through industrial quality control with multiple screening and cleaning rounds. Making it open-source accelerates embodied AI research by giving researchers access to high-quality real-world robot data at scale. 🇨🇳 Learn more here: https://t.co/iIOcEs4AnN ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

🚨 ANTHROPIC JUST BROKE SWE-BENCH PRO WITH CLAUDE MYTHOS 🚨 Anthropic just dropped the numbers for their unreleased "Claude Mythos Preview" and the coding leap is almost incomprehensible. This model is so powerful at finding exploits that they are keeping it strictly locked down for critical infrastructure partners. Anthropic explicitly stated: "We’ve used Claude Mythos to demonstrate thousands of zero day vulnerabilities." Look at the absolute destruction of these benchmarks compared to Opus 4.6: • SWE-Bench Pro: 77.8% (Destroying Opus 4.6 at 53.4%) • Terminal-Bench 2.0: 82.0% (Up from 65.4%) • SWE-Bench Verified: 93.9% • SWE-Bench Multimodal: 59.0% (More than double Opus 4.6's 27.1%) • Humanity's Last Exam (with tools): 64.7% (Up from 53.1%) • GPQA Diamond: 94.6% A nearly 25-point jump in SWE-Bench Pro in a single generation. And we’re in *checks notes* April..

We’re releasing SWE-1.6, our best model in both intelligence & model UX. SWE-1.6 matches our Preview model on SWE-Bench Pro while dramatically improving on various behavioral axes. It’s available today in Windsurf in two modes: free tier (200 tok/s) and fast tier (950 tok/s). https://t.co/JHwMLsaudO

Join us next Monday for our next Speaker Series with @modal’s founder and CEO @bernhardsson. Register for the event and submit questions for Erik here: https://t.co/iPfRTCdJGH https://t.co/A5jQbpI0Ck

@bensig If you are really into AI I built an AI that reads everyone here on X including Brian and makes this out of the best: https://t.co/kiuZ7QXLzb

Looking forward to an exciting partnership with @elonmusk and building Terafab together! https://t.co/Hi3sfAnP3J

Excited to announce a new open-source, free-to-use memory tool I have been developing with my good friend @MillaJovovich. The project is called MemPalace and it is an agentic memory tool that scored 100% on LongMemEval - the industry standard benchmark for memory… this is higher on than any other published results - free or paid - and it is available now on GitHub. You can check out Milla’s video about it on her Instagram. I’ll also put some links in the comments below - please try it out, critique it, fork it, contribute to it - and join our discord.