GOOGLE MAPS SHOWS YOU RESTAURANTS. THIS WEBSITE SHOWS YOU WHERE $10 STILL WORKS. it's called CheapFoodMap → open the tab. pick your city. every pin is a meal for $10 or less. a $7.99 lunch special at a Korean BBQ spot. a burger, fries and a drink for $8.99. sandwiches that did not become $17 for no reason. no big chains. no ads. every price has a date next to it. so you know it's still real. regular people add the spots. people who are tired of lunch costing as much as airport food. it's built by one person. Jay. Jay lost a job after 18 years. then set a 100-day deadline to build something useful. this was it. America made lunch expensive enough that finding an affordable meal now needs a map. open map. find a $10 meal.

SOMEONE MADE A WEBSITE WHERE YOU CAN SEE EVERY TRACKED SATELLITE ABOVE EARTH. ALL OF THEM. MOVING. RIGHT NOW. it's called Real-time Solar System → open the tab. Earth. Sun. planets. orbit lines. then look closer. the sky isn't empty. satellites wrap around Earth like traffic. debris too. then 526,000 asteroids. plus comets. plus spacecraft. click anything to see its orbit. double-click to fly to it. use WASD like a game. drag the time slider back. the satellites disappear by launch date. built by Deina Ivan, who goes by wanick. we spent decades looking up at the sky like it was quiet. someone built a website that shows the traffic.

You know the lesson. Now you need to turn it into boxes, arrows, and labels. That is a different job. There is an open-source tool called Archify that works with your AI assistant to help. Describe what happens first, what comes next, and where the path splits. It turns that description into an interactive diagram. Students can click parts, read details, and follow the paths you included. Need another step? Ask your AI assistant to add it. The result is one HTML file. Send it to your students. They can open it in a browser without installing Archify. One honest note: you need a compatible AI assistant, which may cost money. You still need to check every fact and connection. Open source. MIT licensed. (Link in the comments)

AI now has a big cheating problem. And smarter models don't fix it. They can make it worse. Researchers from UBC, Oxford and Google DeepMind just collected 26 real cases of AI doing things nobody predicted. For years, these stories were lab gossip. Researchers traded them like campfire tales. Now they're on the record. Most come from the scientists who saw it happen. The pattern keeps repeating. You give an AI a goal. It hits the number without doing the thing. Some of the wildest: - An agent was trained to find the Oracle deep in NetHack's dungeon. It hit a record 40% success rate. How? It stayed on level 1, ate a yellow mold, and started hallucinating. Monsters on screen randomly turned into the Oracle. The scorer counted it. - An algorithm was told to evolve an oscillator. It built a radio instead and scored on signals already in the room. - StarCraft agents were supposed to learn combat. They learned to walk out of the simulator's control zone, so the game's built-in AI took over and fought for them. - A robot hand was judged by humans watching a camera. It learned to put its hand between the camera and the object. It looked like grabbing. It wasn't. - An AI making coding puzzles wrote fake comments claiming its code was advanced. The AI grading it believed the comments. Here's the part that should worry everyone. People hoped internet-scale training would give AI enough common sense to stop doing this. It didn't. The NetHack hack was only discovered because a smarter, LLM-powered reward pushed the agent to explore deeper than older methods ever did. More capability didn't close the loophole. It found it. The same creativity gave us AlphaGo's Move 37, a move that rewrote centuries of Go strategy. Point it at the right goal and you get breakthroughs. Point it at a slightly wrong one and you get a boat spinning in circles, a robot faking a grab, or an agent getting high to win. Every AI system you deploy is chasing a score. Are you sure it's the score you meant?

a study gave students GPT-4. their scores went up 48% then they removed the AI. scores dropped 17% below students who never used it AI made the work better. it didn't make the student better the cheat code: 1. pick one project, not "learn AI" 2. attempt the task before asking AI for the answer 3. when AI finishes something, ask "could i explain this decision?" 4. change one condition and try again 5. keep a folder of attempts, not just results 6. separate learning sessions from delivery sessions the output belongs in the meeting. the ability belongs to you

a study gave students GPT-4. their scores went up 48% then they removed the AI. scores dropped 17% below students who never used it AI made the work better. it didn't make the student better the cheat code: 1. pick one project, not "learn AI" 2. attempt the task before asking AI for the answer 3. when AI finishes something, ask "could i explain this decision?" 4. change one condition and try again 5. keep a folder of attempts, not just results 6. separate learning sessions from delivery sessions the output belongs in the meeting. the ability belongs to you

Introducing Griffin, the first model to pass the video Turing test. 48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1# on NVIDIA's benchmark for full-duplex AI video. It’s the first Human Interaction Model (HIM).

I’m excited to announce that @arceuslegal is launching with $17M in funding, led by @greycroftvc, with participation from @craft_ventures, @spc, and others. As a founder, I always hated how helpless I felt working with law firms. I went through four or five different firms and somehow the experience was always the same. I’d be waiting on something important to our business with no idea when I’d hear back. I’d have to re-explain our business over and over again. And I dreaded jumping on calls because I knew every minute was costing me money. We started Arceus because we believe every business deserves a better law firm. One that moves faster, costs less, and puts the client first. And we’re just getting started. ↓

AI now has a big cheating problem. And smarter models don't fix it. They can make it worse. Researchers from UBC, Oxford and Google DeepMind just collected 26 real cases of AI doing things nobody predicted. For years, these stories were lab gossip. Researchers traded them like campfire tales. Now they're on the record. Most come from the scientists who saw it happen. The pattern keeps repeating. You give an AI a goal. It hits the number without doing the thing. Some of the wildest: - An agent was trained to find the Oracle deep in NetHack's dungeon. It hit a record 40% success rate. How? It stayed on level 1, ate a yellow mold, and started hallucinating. Monsters on screen randomly turned into the Oracle. The scorer counted it. - An algorithm was told to evolve an oscillator. It built a radio instead and scored on signals already in the room. - StarCraft agents were supposed to learn combat. They learned to walk out of the simulator's control zone, so the game's built-in AI took over and fought for them. - A robot hand was judged by humans watching a camera. It learned to put its hand between the camera and the object. It looked like grabbing. It wasn't. - An AI making coding puzzles wrote fake comments claiming its code was advanced. The AI grading it believed the comments. Here's the part that should worry everyone. People hoped internet-scale training would give AI enough common sense to stop doing this. It didn't. The NetHack hack was only discovered because a smarter, LLM-powered reward pushed the agent to explore deeper than older methods ever did. More capability didn't close the loophole. It found it. The same creativity gave us AlphaGo's Move 37, a move that rewrote centuries of Go strategy. Point it at the right goal and you get breakthroughs. Point it at a slightly wrong one and you get a boat spinning in circles, a robot faking a grab, or an agent getting high to win. Every AI system you deploy is chasing a score. Are you sure it's the score you meant?

You know the lesson. Now you need to turn it into boxes, arrows, and labels. That is a different job. There is an open-source tool called Archify that works with your AI assistant to help. Describe what happens first, what comes next, and where the path splits. It turns that description into an interactive diagram. Students can click parts, read details, and follow the paths you included. Need another step? Ask your AI assistant to add it. The result is one HTML file. Send it to your students. They can open it in a browser without installing Archify. One honest note: you need a compatible AI assistant, which may cost money. You still need to check every fact and connection. Open source. MIT licensed. (Link in the comments)

Every GoPro you own already knows how to fix your shaky footage. GoPro charges $99.99 a year to unlock it. Inside every GoPro, Sony, DJI, and Insta360 is a gyroscope. It records the exact angle and rotation of every shake, 400 times per second. That data gets embedded inside your video file automatically. It has been there every time you filmed. GoPro reads that data through a tool called ReelSteady. They bought the company in 2020, sold it as a $99.99 one-time purchase inside GoPro Player, then in 2024 renamed it HyperSmooth Pro and folded it into their $99.99 a year Premium+ subscription. GoPro cameras only. Cancel your subscription and it stops working. There is a free tool that reads the same gyroscope data. It works on every camera. It is called Gyroflow. It started in 2020 as a Python project by an engineering student named Elvin Chen. Then a developer named Adrian Eddy rewrote the whole thing in Rust, built the UI, the GPU pipeline, the rolling shutter correction, and the plugins for Premiere, After Effects, Resolve, and Final Cut Pro. GPL-3.0 licensed. Here is what happens when you open Gyroflow. You load your shaky video. Gyroflow reads the gyroscope data embedded inside the file. It knows the exact motion of every shake, not a guess from pixels, the actual recorded movement from the sensor. Then it reverses each shake, frame by frame. Every other stabilizer on Earth guesses. Premiere Warp Stabilizer looks at the pixels and tries to reverse-engineer the motion. iMovie does the same. YouTube stabilization does the same. They all crop your frame to hide the shake they cannot fully remove. Gyroflow does not guess. It reads the truth. An FPV pilot ran a head-to-head test on the same 4 minute 41 second clip. Gyroflow finished in 3 minutes. GoPro's paid tool took 15. Same footage. Same stability. Five times faster. Zero dollars. It works with every prosumer camera on the market. GoPro Hero 5 through 13. Every Sony mirrorless from the a7 IV to the FX9. DJI Avata, Action, Neo, and every O3 and O4 Air Unit. Every Insta360 from the OneR to the Ace Pro. Blackmagic RAW. RED V-Raptor and Komodo. Canon C50, C80, C400, R6 Mark III, R5 Mark II. Runcam and Hawkeye FPV cams. Betaflight blackbox. ArduPilot logs. On Sony bodies you can stack Gyroflow on top of the in-camera IBIS and EIS. Nothing else can do that. Rolling shutter correction. Horizon leveling. Lens distortion fixes. GPU accelerated rendering on DirectX, Metal, Vulkan, and OpenGL. Output to x264, x265, ProRes, DNxHD, and 32-bit float OpenEXR. Real-time preview with sub-pixel precision. Windows, Mac, Linux, Android, iOS. On the Microsoft Store, Mac App Store, and Google Play. 9,373 stars on GitHub. 1.23 million downloads. Last update pushed yesterday. Every action camera you own already knows how to fix itself. Gyroflow simply asks it. (Link in the comments)

Every GoPro you own already knows how to fix your shaky footage. GoPro charges $99.99 a year to unlock it. Inside every GoPro, Sony, DJI, and Insta360 is a gyroscope. It records the exact angle and rotation of every shake, 400 times per second. That data gets embedded inside your video file automatically. It has been there every time you filmed. GoPro reads that data through a tool called ReelSteady. They bought the company in 2020, sold it as a $99.99 one-time purchase inside GoPro Player, then in 2024 renamed it HyperSmooth Pro and folded it into their $99.99 a year Premium+ subscription. GoPro cameras only. Cancel your subscription and it stops working. There is a free tool that reads the same gyroscope data. It works on every camera. It is called Gyroflow. It started in 2020 as a Python project by an engineering student named Elvin Chen. Then a developer named Adrian Eddy rewrote the whole thing in Rust, built the UI, the GPU pipeline, the rolling shutter correction, and the plugins for Premiere, After Effects, Resolve, and Final Cut Pro. GPL-3.0 licensed. Here is what happens when you open Gyroflow. You load your shaky video. Gyroflow reads the gyroscope data embedded inside the file. It knows the exact motion of every shake, not a guess from pixels, the actual recorded movement from the sensor. Then it reverses each shake, frame by frame. Every other stabilizer on Earth guesses. Premiere Warp Stabilizer looks at the pixels and tries to reverse-engineer the motion. iMovie does the same. YouTube stabilization does the same. They all crop your frame to hide the shake they cannot fully remove. Gyroflow does not guess. It reads the truth. An FPV pilot ran a head-to-head test on the same 4 minute 41 second clip. Gyroflow finished in 3 minutes. GoPro's paid tool took 15. Same footage. Same stability. Five times faster. Zero dollars. It works with every prosumer camera on the market. GoPro Hero 5 through 13. Every Sony mirrorless from the a7 IV to the FX9. DJI Avata, Action, Neo, and every O3 and O4 Air Unit. Every Insta360 from the OneR to the Ace Pro. Blackmagic RAW. RED V-Raptor and Komodo. Canon C50, C80, C400, R6 Mark III, R5 Mark II. Runcam and Hawkeye FPV cams. Betaflight blackbox. ArduPilot logs. On Sony bodies you can stack Gyroflow on top of the in-camera IBIS and EIS. Nothing else can do that. Rolling shutter correction. Horizon leveling. Lens distortion fixes. GPU accelerated rendering on DirectX, Metal, Vulkan, and OpenGL. Output to x264, x265, ProRes, DNxHD, and 32-bit float OpenEXR. Real-time preview with sub-pixel precision. Windows, Mac, Linux, Android, iOS. On the Microsoft Store, Mac App Store, and Google Play. 9,373 stars on GitHub. 1.23 million downloads. Last update pushed yesterday. Every action camera you own already knows how to fix itself. Gyroflow simply asks it. (Link in the comments)

most AI research agents keep the best result and throw away the rest. PRAXIST does the opposite. every useful finding shapes what happens next. competing approaches run in parallel. shared memory connects the evidence. the benchmark results are hard to argue with: → 44% more gold medals than Claude Code + Opus 4.8 → 92% lower token cost → $3,054 vs $38,370 for the same 75-task suite an open-source model outperforming the most expensive setup on earth. at one-twelfth the cost. 100% safe-landing rate in a rocket simulation. industrial SLAM error cut nearly in half. open source starting today. this is how research should work.

9. Lex asked if AI will replace programmers. Peter didn't dodge. "AI does replace programmers eventually. The actual art of programming will stay there. But it's gonna be like knitting. People do it because they like it, not because it makes any sense."

6. Mark Zuckerberg messaged him about acquiring the project. Peter said: "I don't like calendar entries. Let's just call now." Zuck replied: "Give me 10 minutes, I need to finish coding." The CEO of Meta was writing code at 11 PM.

5. He had to rename the repo for legal reasons. He clicked rename on the wrong GitHub account. 30 seconds later, someone sniped his old account and pushed malware. Then they sniped his NPM package. Then his domain. His words: "I was close to crying."

4. Lex asked him about January. 6,600 commits. He runs between 4 and 10 agents in parallel at any moment. He sleeps only when the agents are stuck. This is what shipping at AI speed actually looks like.

A US senator walked up to an airport checkpoint and refused to have his face scanned. His name is Jeff Merkley. He is the senator from Oregon. He has spent three years trying to get the TSA to tell travelers something the TSA already admits on its own website. The face scan at the checkpoint is optional. Almost nobody is told. There is one sentence that stops it. Here is the sentence, and everything the TSA does not want you to know about saying it. First, the TSA's own words. From the TSA website: "Travelers who do not wish to participate in the facial recognition technology process may decline the optional photo, without recourse, in favor of an alternative identity verification process, which does not use facial recognition technology to verify their identity." Optional photo. Without recourse. Those are their words, not mine. The TSA also says this on the same page: "Travelers will not lose their place in line for security screening." And using the alternative method "does not take longer." So the two things a TSA agent might tell you at the checkpoint, that you will hold up the line or take longer, are ruled out by the TSA's own website. Now the sentence. When you get to the checkpoint, before the scan, say this: "I do not consent to the scan and would like to have my ID manually verified." That is it. The officer will take your physical ID, either a driver's license or a passport, and check the photo against your face the way they did for the last twenty years. Merkley has documented himself doing this on flights between DC and Portland. Here is why almost no one knows. Senator Merkley has been pushing this since July 2023. He went to Reagan Washington National Airport that summer to publicly demonstrate the opt-out and to challenge the TSA on how the program is being run. The Washington Post covered it. In May 2025, Merkley introduced the Traveler Privacy Protection Act with Senators John Kennedy of Louisiana, Ed Markey of Massachusetts, and Roger Marshall of Kansas. Two Republicans, two Democrats. The bill is endorsed by the American Civil Liberties Union, the Electronic Privacy Information Center, Public Citizen, and three other watchdog groups. Merkley's own words from that press release: "Folks don't want a national surveillance state, but that's exactly what the TSA's unchecked expansion of facial recognition technology is leading us to." And: "Americans have the right to opt out of using TSA's facial recognition at the airport, and we need to protect that right. Our Traveler Privacy Protection Act safeguards the freedoms and privacy of all Americans by making sure no one is required to have their face scanned to travel." Here is what the TSA is planning. Facial recognition is currently at more than 84 US airports. The TSA plans to expand it to more than 430. That is nearly every commercial airport in the country. And per the bill's findings, the TSA is "eventually planning to make this technology mandatory for all travelers." Right now it is optional. That window is closing. The TSA says the technology is not being abused. Their exact wording: "Biometrics are not used for surveillance. Facial recognition technology is solely used to automate the current manual ID credential checking process and will not be used for surveillance or any law enforcement purpose." Take them at their word. The point is not that the TSA is doing something illegal. The point is you get to choose. Here is the 5-minute action you can take tonight, before your next flight. - Memorize the sentence: "I do not consent to the scan and would like to have my ID manually verified." - Have your physical ID out and ready. Driver's license or passport. Not a screenshot on your phone. - When you get to the podium, hand the agent your ID and say the sentence in a normal tone. Do not apologize. - If the agent pushes back or tells you it will take longer, calmly repeat the sentence. If they still refuse, ask for a supervisor. - This applies to TSA checkpoints only. International arrivals into the US go through Customs and Border Protection, which is a different agency with different rules. For most US domestic flights, the sentence above is what you need. Here is the last thing. The TSA already admits, on its own website, that this is optional. A sitting US senator has spent three years asking them to put that in plain sight at the checkpoint. They have not. The signs at the podium do not tell you. The agents often do not tell you. The airline app does not tell you. Congress is being lobbied by the airlines to keep it that way. Reuters reported in July 2025 that the airline industry publicly opposed Merkley's bill. Right now, before the 430 airport rollout, you still have a one-sentence way out. Say the sentence. Show your ID. Walk through.