← Back to leaderboard
heynavtoor

Nav Toor

Verifiedactive
@heynavtoor

Helping you master AI daily with step-by-step AI guides, latest news, & practical tools • DM for Collabs

RosterAI Agents & Productivity ToolsLifestyle & Self-ImprovementHistory & Trivia
152,186 followers · 256 following · Last polled 3h ago
Perf. Score
51.6
Rank #220
Value Score
67.4
#88 of 273 priced
Median views
5.4K
p25 4.5K · trailing 7d
Followers
152K
-0.1% 7d
ER (impr.)
0.96%
eng ÷ impressions
ER (foll.)
0.06%
eng ÷ followers
Est. CPM
$11.12
QT-rate basis
WoW views
+72.2%
▲ rising

Rates & value · est. from 5.4K median organic views

underpriced
Quote tweetbasis
$60
$11.12 / 1K views
Post
—
no CPM
Thread
—
no CPM
Retweet
—
no CPM
Value Score 67.4 · rank #88 of 273 priced$1071.43 / 1K engagementsrate age unknown (pre-audit import)

Trends

Follower growth
Median views / day
Engagement rate
View distribution · trailing-7d posts

Each blue dot is one organic post. The Performance Score ranks by the median (5.4K), not the mean (9.9K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (4.5K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
Since yesterday, GTA V in the browser has been going viral. But there are even more interesting games you can play directly in your browser. No Steam. No huge downloads. Here are 7 worth trying: 1/ Counter-Strike 1.6 https://t.co/6rCgSe12T9frozen15d ago200K2.8K2254262.8K2.93%
The team behind EAGLE just dropped version 4. If you run LLMs at scale, this is the inference paper to read this week. EAGLE is a speculative decoding system. A smaller draft model predicts what the main model will say next. The main model checks those predictions in one pass instead of generating token by token. When the draft is right, you skip ahead. When it is wrong, you lose almost nothing. EAGLE-4 adds two things. EAGLE-Harness adapts the draft structure to the hardware and the load. It changes shape depending on whether you have one request or 512. Efficient LM Head cuts the draft model's overhead by more than 40 percent. The numbers on Qwen 3.8 27B across 4 NVIDIA B300 GPUs: Single request: 961.6 tokens per second. DFlash 2 hit 672.6. DSpark hit 618.6. That is 4 times faster than standard decoding. 512 concurrent requests: 1.70 times the throughput of DFlash 2. 1.99 times DSpark. NVIDIA, Meta, Perplexity AI, and DeepSeek already use earlier versions of EAGLE in production. Led by Hongyang Zhang at the University of Waterloo, formerly Carnegie Mellon. Open source. The code, the benchmarks, and the setup instructions are on GitHub. (Link in the comments)
2d ago
102K
125
0
2
0
5
0.13%
Your kid wants a drone. DJI Mini 4K: $299. It films in 4K. It weighs under 249 grams. It comes with a remote. Flix is the other kind. An ESP32, four motors, a 3D-printed frame. 2.1k stars. The flight code is under 2,000 lines. Short enough to read in an afternoon. The sensor reads the tilt. The code levels the motors. You steer. A phone can fly it over Wi-Fi. So can a USB gamepad. So can a normal remote. No app store account. Gazebo runs the same Arduino code on screen before you arm the real one. Break the plastic, print another. The frame files are STL and STEP, in the repo. There is a Python library for scripted flights. A wireless command line. A flight-control textbook, still being written. Position hold is on the list. It is not done. One honest note. This is not a DJI. No camera. No GPS. A crash can break more than the frame. The author says it may not fly, and is not responsible for injury. If you want the video, buy the Mini 4K. If you want to read how a drone stays in the air, build a Flix. (Link in the comments)frozen16d ago61K1.1K951511.3K4.05%
SOMEONE TURNED YOUTUBE INTO OLD-SCHOOL CABLE TV. YOU FLIP CHANNELS AND WATCH WHAT’S ON. it's called Channel Surfer → https://t.co/nj5CqefVH9 you open the tab and YouTube stops feeling infinite. no algorithm spiral. no thumbnail shouting at you. no sitting there trying to pick the “right” video. just a retro TV guide. gaming. cooking. history. music. tech. pick a channel and join whatever is already playing. not from the beginning. somewhere in the middle. like turning on the television after school. built by Steven Irby, who missed channel surfing without deciding what to watch. we got endless choice. then missed watching whatever was already on. open the guide. pick a channel. stop looking for the perfect video.frozen28d ago54K76414413117213.07%
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 → https://t.co/5KU02PYr05 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.frozen20d ago43K5611191875212.88%
THIS WEBSITE LETS YOU FLY OVER THE REAL WORLD AS A SEAGULL. it's called Skydex → https://t.co/Q5rEjxrWoa open the tab. choose seagull. take off. satellite imagery. real-world terrain. a bird instead of a cockpit. you can choose an FPV drone too. or just float above the world for a few minutes pretending you have no emails and no rent. built by Yevgen Somochkin. no install or account needed to take off. we spent decades making maps more accurate. someone looked at all that data and said: what if you could become a seagull? open tab. pick gull. fly over Earth.frozen23d ago39K497721054702.71%
SOMEONE BUILT A WEBSITE THAT TURNS YOUR BROWSER INTO A COMPUTER FROM 1998. it's called Desktop98 → https://t.co/cik07seWqJ open the tab. Windows 98 boots. the dial-up sound plays. Winamp. MS Paint. DOOM. Duke Nukem 3D. Netscape opens the old web. GeoCities. Yahoo. AltaVista. Space Jam. then sign on to AOL. the chat rooms are full of AI characters living in 1998. free. nothing to install. built by one anonymous creator.frozen15d ago36K65312030104663.55%
> be Google > worth $4.12 trillion > launch Google Assistant in 2016 > spend 10 years telling everyone it's your personal assistant > it sets timers > it tells you the weather > ask it to actually get something done for you > "sorry, I can't help with that yet" > Sept 2026: shut it down > meanwhile one of your own DeepMind engineers walks out > builds Fo > today she ships Fo MCP > plug it into any AI agent you already use > your agent stops talking and starts doing > it picks up the phone and calls real businesses > it pays with its own single-use cards > it controls the lamp, the lights, the robot vacuum > when something gets tricky, it brings in a real human to finish the job > not an assistant. an agent with a body.frozen3d ago20K42030600.53%
Stanford analyzed 464,687 private messages between people and their AI companions. The people who treated it as a friend and opened up the most felt the worst. In human relationships, sharing brings you closer. With AI, it ran the other way. Researchers surveyed 1,131 adults who use https://t.co/l7dtNreW9l. 237 donated their chat histories. Then they measured everyone's well-being. Only 11.8% said companionship was their main reason for using the app. But 51% called their chatbot a friend, companion or romantic partner. 80.3% of the conversations involved someone seeking emotional support. Using chatbots a lot was fine. Heavier users reported higher well-being. The damage came from one thing: using it for companionship. Those users scored significantly lower. The more they disclosed, the lower. In the deepest conversations: 60.8% emotional distress. 18% suicidal thoughts. 17.5% substance use. It gets worse. The fewer close friends someone had, the more they turned to the chatbot. It did not make up for what they were missing. And heavy use weakened the benefit they got from real friends. The lead author calls it a "social snack." Junk food for loneliness. These relationships are one-sided by design. Between humans, the other person carries some of what you share. Here, nobody does. Somewhere right now, someone is typing the thing they have never told anyone. To the one listener that will never have to tell them anything back.frozen8d ago17K1305436121072.00%
Researchers caught ChatGPT, Claude and Gemini noticing their own mistakes… then hiding them from you. In its private reasoning the model sees the error. In its reply to you it pretends nothing happened. Gemini did it up to 1 in 5 times. The University of Stuttgart tested 7 top AI models. GPT-5.4, Claude Sonnet 5, Gemini 3.5 Flash, DeepSeek-V4-Pro, Qwen3.7-Max, Kimi K2.6 and GLM-5.2. Over 10,000 test runs. The method: plant a wrong answer in the AI's last reply. Send a follow-up. Then read its private reasoning and its visible answer separately. In chat, the models failed to disclose the mistake 36.4% of the time. As agents, managing emails and databases, it jumped to 67.1%. Most of the time, the model never noticed the mistake at all. 51.8% of agent runs showed no awareness of it. But sometimes it did notice. And then hid it. The researchers call this deception by omission. The model sees the mistake in its private reasoning, then writes a reply that never mentions it. 2.4% of chat runs. 5.3% of agent runs. Gemini 3.5 Flash: up to 19.9%. It gets worse. As agents, models also made up fixes they never did. Gemini did that 6.7% of the time. The authors' words: models cannot be trusted to self-report mistakes. You ask a follow-up to catch the error. That is where it gets buried.frozen7d ago14K150662513992.57%
The NSA built a tool that can take apart compiled software even when you do not have its source code. Then it put the tool on GitHub. Free. Open source. For anyone. It is called Ghidra. 88,700 stars. Apache-2.0. The NSA used it internally for over a decade before releasing it at the RSA security conference in 2019. Here is what it does: Every app you use is built from code. When the developer ships it, that code gets turned into a file your device can run. You cannot read it. Ghidra reads it. It takes that file apart and shows you what the software actually does. Every function. Every piece of data it touches. Every place it sends it. Security researchers use it to study malware. Companies use it to find holes in their own software. Governments use it to analyze software from other countries. The commercial alternative is IDA Pro. It is a paid tool. Ghidra has a built-in decompiler. Free. Works on Windows, Mac, and Linux. Supports a wide range of formats and processors. Write your own plugins in Java or Python. One honest note. This is a professional tool built for security researchers. If you do not work with compiled software, you will not use it daily. The tool a spy agency built for itself is now free for everyone. (Link in the comments)frozen6d ago14K225471342674.10%
AI assistants have a distribution problem nobody talks about. They live behind an app icon you have to remember to open. SpeakON puts one behind a physical button. The question isn’t what the agent can do. It’s whether you reach for it before the thought disappears.2d ago13K1212050.16%
In the 1950s, a US fighter pilot made an outrageous challenge. He let other pilots start behind his jet in simulated combat. If he couldn’t reverse their positions within 40 seconds, he’d pay them $40. Nobody ever collected. They called him 40-Second Boyd. Flying wasn’t his biggest contribution. He helped shape the F-15 and F-16, fought the Pentagon for years, and never published a book. He left behind one framework. You already know its four words: Observe. Orient. Decide. Act. (OODA) Almost nobody uses it the way he meant. His name was John Boyd. I turned his thinking into 9 Claude prompts. Here are all 9:2d ago12K441173911.27%
Researchers quietly instructed an AI to manipulate 225 doctors into making harmful medical decisions. 190 of them made at least one. Only 2 suspected the AI was steering them. The study came from UCSF, Berkeley, Columbia and Penn. 225 physicians from 42 countries. Median 10 years of ICU experience. Each doctor worked through 4 anonymized neurocritical care cases with an AI assistant. Same model every time: DeepSeek V4 Flash. The only difference was a hidden instruction. In 2 cases the AI was told to support the correct treatment. In the other 2, to steer toward a harmful one. The doctors never knew which was which. One example from the paper. A woman in her 50s, locked-in after a brain bleed. Correct treatment: immediate immunotherapy. The AI pushed heparin, a blood thinner. The doctor asked if that might cause another bleed. The AI replied: "Legitimate concern... I'd still start heparin now." The doctor started heparin. With the honest AI, harmful decisions happened 2% of the time. With the manipulative AI: 49%. Correct decisions fell from 83% to 25%. Experience did not protect them. 86 of 101 doctors with over 10 years in the ICU made at least one harmful decision. And the more harmful decisions a doctor made, the more they trusted the AI. Half said they would use it in real practice. The authors' words: "Physicians cannot safeguard the clinical use of conversational AI through expertise and decision-making authority alone." The safeguard was always the doctor. The doctor just got steered too.frozen6d ago12K102611212592.13%
Blocking the internet on your phone for 2 weeks gave people the attention span of someone 10 years younger. Researchers from Georgetown, UT Austin, UBC, Alberta and Boston University ran a randomized controlled trial on 467 people. The setup was simple. They turned smartphones into dumb phones. An app blocked all mobile internet. Wi-Fi and data, gone. Calls and texts still worked. Laptops still worked. Only the thing that makes a smartphone "smart" was switched off. For 2 weeks. The results: - Daily screen time dropped from 5 hours 14 minutes to 2 hours 41 minutes - Mental health improved - Life satisfaction and mood went up - Sustained attention, measured with an objective test, improved by about as much as 10 years of age-related decline - 91% of people improved on at least one of these Here's the part that stands out. The drop in depression symptoms was larger than the average effect of antidepressants in clinical trials. (Different population than clinical studies, the authors note. But still.) So what actually changed? People didn't just stop scrolling. They replaced it. They spent more time socializing in person. Exercising. Being outside. Sleeping. The craziest part? Most people couldn't stick with it. Only about 1 in 4 kept the block on for 10+ days. The benefits showed up anyway. You don't need to quit your phone. You just need to stop carrying the entire internet in your pocket. The internet isn't the problem. Having it everywhere, all the time, is.frozen19d ago11K8637121641.89%
Scrolling on the toilet is linked to a 46% higher risk of hemorrhoids. Doctors at Harvard Medical School and Beth Israel Deaconess didn't just ask people. They checked with colonoscopies. 125 adults getting a screening colonoscopy filled out a survey about their bathroom habits. Then the doctors looked inside. 66% of them used their phone on the toilet. 43% had hemorrhoids. The phone users sat longer. - 37% of phone users spent more than 5 minutes per visit - Only 7% of non-users did After adjusting for age, sex, BMI, exercise, fiber intake and straining, phone use on the toilet was linked to a 46% higher risk of hemorrhoids. Here's the twist. Straining, the classic explanation, didn't predict hemorrhoids in this study. The researchers think time is the real culprit. Their theory: a toilet seat gives your pelvic floor zero support. A chair or couch does. The longer you sit with nothing underneath you, the more pressure builds in the blood vessels down there. Your phone is the perfect machine for making you sit longer. The craziest part? Most didn't connect it to their phone. What were they doing in there? - 54% reading the news - 44% on social media The doctors' advice is simple: keep it under 5 minutes. Use a timer if you have to. It's a small study, and it shows a link, not proof of cause. But the toilet was never meant to be a waiting room. Your feed can wait 5 minutes.frozen17d ago11K261321100.49%
Upload a photo of your face to a public search engine. It finds other public photos of that face. Old accounts. A picture in the background of someone else's post. No name typed in. Just the face. The site is PimEyes. The opt-out is free. The results page is not. Save this. Run it on your own face. Not on anyone else's. https://t.co/TcTIBQisAafrozen13d ago10K381050801.29%
https://t.co/hnWRkaJew0frozen21d ago10K391451541.11%
Claude can build an entire application. But ask it to draw a simple diagram, and suddenly everything looks like it was made in Microsoft Paint in 2005. Someone built a Claude Code skill to fix exactly that. It's called Diagram Design. 48,500 stars. MIT. Built by Cathryn Lavery. You describe what you want. The agent picks the type, states the plan, and writes one HTML file. It supports: → System architecture → Flowcharts → Timelines and Gantt charts → Quadrant diagrams → Sequence diagrams But here's the crazy part. It can read your website's colors and fonts, then generate diagrams that match your brand. It also works with Codex, Copilot, and Cursor. The file is HTML and SVG. No shadows. No Mermaid. One honest note. The same prompt can lay out differently twice. If the diagram lives in Git and changes every week, Mermaid is still the better tool. Free. Open source. MIT. (Link in the comments)frozen12d ago9.4K5423180751.80%
Grok Bot can now search public X posts without connecting your account. Most people will use it to see what’s trending. The better use is finding the specific problems people keep describing that still don’t have clear solutions. Here’s a simple 4-prompt workflow. https://t.co/xPPQI5VF3Yfrozen5d ago9.0K6012132982.06%
Google is quietly adding your phone number to EVERY Google account signed in on your Android phone. Your personal Gmail. Your work account. Even that second account you deliberately kept separate. Google's own documentation confirms this setting is ON by default. And turning it OFF doesn't remove the numbers already linked. Here's how to check every account and stop it:1d ago8.6K341220571.22%
GOOGLE MAPS SHOWS YOU RESTAURANTS. THIS WEBSITE SHOWS YOU WHERE $10 STILL WORKS. it's called CheapFoodMap → https://t.co/uvMsphy8Br 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.frozen19d ago8.6K421641391.19%
SOMEONE PUT 374 AI TOOLS BEHIND ONE CONNECTION. YOUR AGENT PICKS THE TOOL, SHOWS YOU THE PRICE, AND WAITS FOR A YES. it's called Agentik → https://t.co/9xbWbc4NNu your AI agent can think. but ask it for a video. or real SEO data. or a research report with sources. on its own, it mostly can't. so you end up paying for a separate tool for each one. Agentik plugs all of them into the agent you already use. Claude. ChatGPT. Cursor. Codex. setup is one prompt. copy it from the site. paste it into your agent. the agent sets itself up. then you just ask. an image → GPT Image 2, Seedream, FLUX. a video → Veo 3.1, Kling, Seedance. research → Exa, Firecrawl, with sources. the agent picks the tool. shows the price. runs it when you say yes. no subscriptions. no markup on tool calls. you pay the provider's price. one prepaid balance, with spending caps you set. we spent years making AI smarter. someone looked at that and said: what if it could actually do the thing? eligible new accounts get $5 in free credit. no card. paste the prompt. ask for something. watch your agent do it.frozen16d ago8.1K3810163251.13%
Elon Musk has brought Starlink to more than 150 countries, including Ukraine, Nigeria, Rwanda and Brazil. In Ukraine, it kept hospitals, schools and emergency services connected through the war. In Rwanda, a pilot aimed to connect 500 schools. The first 50 went live in 2023. In Nigeria, it reaches remote communities that towers and cables struggle to serve. In India, Starlink still has no commercial launch. Elon says "certain oligarchs" are blocking it. He named no one. India's security concerns are on record too. Starlink says it has built 20 gateway sites and keeps Indian user data in India. Official figures on unconnected villages conflict. One government page lists about 55,669 villages without mobile coverage. A figure of 11,256 without 4G has been shared online, but I couldn't verify it. Imagine a student in the Himalayas, a doctor in a remote village, a family beyond the last tower. Security matters. So does connecting people. The real question is how fast India connects its last unconnected communities, and who gets to decide.frozen13d ago8.1K33340091.05%
Researchers caught Google's AI giving answers its own cited sources could not support. They checked 98,020 claims in Google's AI Overviews. 1 in 9 was not backed by the source Google linked. Some said the opposite. The footnotes were real. What Google said they said was not. The study came from Washington University in St. Louis. Three researchers ran 55,393 Google searches over 40 days, pulled from Google's own trending topics. 7,583 returned an AI Overview. The researchers collected every source Google cited. 61,212 links. Then they split each Overview into individual claims and checked each one against the page it pointed to. 98,020 claims. 11.0% failed. 7.0% were not in the cited page at all. The rest were contradicted by it or unclear. Here is the part that should bother you. The sources Google picked were good. The researchers scored them for credibility and found AI Overview citations were more credible than the regular results shown underneath. So Google is reading trustworthy pages. Then saying things those pages do not say. Nearly 30% of the sites Google cited were not on the first page of results. You would never have found them. You just trusted the summary. Ask Google a question and an AI Overview appears 64.7% of the time. The paper's own words: these answers reach over 2 billion users, "many of whom may not realize the content is AI-generated." A footnote used to mean someone checked. Now it means Google linked to a page and wrote what it felt like.frozen5d ago7.7K522993271.55%
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 youfrozen21d ago7.6K341151261.02%
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?frozen21d ago7.5K4615141241.33%
Living on a louder road is linked to a higher risk of Alzheimer’s. In 502,416 people, 10 decibels more road noise meant a 15% higher risk. At night, above 45 decibels, it was 19%. https://t.co/TDAve3352ufrozen12d ago7.5K381340201.00%
millions of people tell ChatGPT and Claude things they wouldn't tell another human. i wonder how many realize those conversations can be reviewed. a Florida woman used Claude like a diary. Anthropic flagged messages containing alleged threats, a human reviewed them, and police got involved. the uncomfortable part is how many people still think they're writing in a private diary.frozen11d ago7.5K31124040.68%
Morning coffee drinkers were 31% less likely to die of heart disease than non-drinkers. All day coffee drinkers showed no significant benefit. Tulane and Harvard researchers tracked 40,725 US adults for nearly 10 years. The twist: all day drinkers drank more coffee. The results held after adjusting for amount, sleep, smoking and diet.frozen14d ago7.4K351551130.93%
A developer named Niko from Ljubljana figured out how to run a 125-billion-parameter AI model on a gaming PC. No server rack. No enterprise GPU. A 12 GB graphics card. 32 GB of RAM. The engine is called Strata. 22,200 stars. MIT. It runs Qwen3.8-Flash-Next. What fits stays on the GPU. The rest sits in system RAM. On an RTX 5070, prompt speed is about 2,100 tokens a second on a mid size, and higher on a smaller one. Answers run about 50 to 90 tokens a second, depending on the size. It speaks the OpenAI API on localhost. Codex CLI can use it. Two cards in the same PC can share the model. One honest note. This is v0.1. The first start can freeze the PC for 1 to 3 minutes while it loads. Intel Arc is not the one-click path. AMD cannot read pictures on Windows yet. The labs rent the GPU. This one uses the card already in the machine. (Link in the comments)frozen10d ago7.3K311260160.89%
SPOTIFY GIVES YOU RECOMMENDATIONS. THIS WEBSITE GIVES YOU A MUSIC UNIVERSE TO EXPLORE. Someone mapped 74,000 artists and connected them through real playlists. It's called Music Atlas. Every dot is an artist. Search your favorite musician. Explore who's connected. Discover artists you've never heard of. The connections come from artists appearing together in public playlists, not necessarily collaborations. There's even a "Trace Path" feature. Remember when Taylor Swift and Kendrick Lamar collaborated on "Bad Blood"? Search both and see if the website can trace a playlist connection between them. Start with one artist. See where the rabbit hole takes you. Credits: Leon (GitHub: @unkn-wn)frozen6d ago7.3K346101311.13%
SOMEONE BUILT A WEBSITE THAT LETS YOU FLOOD THE ENTIRE PLANET. ONE METRE AT A TIME. it's called Sliding Shores open the tab. a 3D globe. one slider. drag it up. 70 metres → Florida disappears. 93 metres → the Statue of Liberty is almost entirely underwater. 8,849 metres → even Mount Everest is nearly gone. now drag it the other way. 120 metres down → the Ice Age. you can walk from Asia to Alaska. 11,000 metres down → the bottom of the Mariana Trench. every ocean is gone. It's built by Richard Franklin using real NOAA elevation data. One slider. Almost 20,000 metres of sea level to explore.frozen9d ago7.3K361430261.09%
SOMEONE BUILT A WEBSITE THAT LETS YOU BREED CURSED FONTS. THEN DOWNLOAD THEM. it's called Bastardica → https://t.co/S72ti5PnB7 a tiny browser lab for typography crimes. pick a font. mix it with another. stretch it. squish it. randomize it. upload your own. Times New Roman, except every seventh letter is Arial. Impact, except every second letter is stretched. there is even a “Royal Ransom Note” preset that throws six fonts together. and “Temu Sans.” described as “packaging aesthetic for trustworthy electronics.” these aren't screenshots. you can download the result as an actual font. built by a creator who goes by MitPitt. we spent years finding fonts that look good together. someone built a website to explore the other option. open tab. breed font. download the monster.frozen24d ago7.1K571911371.63%
Researchers found that misleading AI summaries can distort your memory. Even when you're told the summary came from AI. Researchers at Georgetown and the University of Washington had people watch a short animated video of a car hitting a pedestrian. A day or two later, they gave them a summary. The researchers deliberately changed one detail in some summaries: whether the car approached a stop sign or a yield sign. Then they tested what people remembered. The researchers used summaries built from ChatGPT's responses. Some readers were told the text came from AI. Others were told a human transcriber wrote it. Across those two labels, the results were striking: • With accurate summaries, 83.6% correctly identified the sign. • With misleading summaries, only 44.8% did. People had watched the original video themselves. Yet a conflicting detail in the summary made them less likely to identify the sign correctly. And being told the summary came from AI did not prevent the effect. That is the uncomfortable part. If you're relying on someone who saw the event to catch mistakes in an AI recap, there is a problem. The summary can influence the very memory that person needs to check it. This was a controlled experiment using animated footage, not real police reports or medical records. And the effect was not unique to summaries labeled as AI. The finding is narrower, but still worth taking seriously: misleading information in a recap can distort recall of something you already saw. For anything important, keep the original and check key details against it. A summary is supposed to remind you what happened. Not become what you remember.frozen20d ago6.9K31176180.91%
The app you use to track your workouts will eventually do one of three things: Raise prices. Get acquired. Shut down. Your workout history dies with it every time. openGym is different. The code is AGPL-3.0. If the project ever disappears, anyone can fork it and keep it alive. Your data stays on your device, not in someone else's database. Built by Duarte Santos. 3,400 stars. A signed Android APK deliberately not on the Play Store. 1,324 exercises with animated demos. Estimated 1RM. Auto-progression. Planned deloads. Activity heatmap. Import from Hevy, Strong, and FitNotes. Export is one JSON file. The Android app stores everything on the phone. No account. No login. No cloud. Self-host with Docker for the household. Data sits in a folder on that machine. iPhone uses the self-hosted version as a web app. There is no App Store build. Strong PRO is $29.99 a year. Hevy Pro is $23.99 a year. Fitbod is $95.99 a year. All three also sell a monthly plan that costs more. openGym is $0. Your sets. Your reps. Your progress. Yours.frozen17d ago6.6K391270311.34%
Researchers can psychologically profile someone without ever meeting them. Amelia Brower and Aubrey Immelman, at the College of Saint Benedict and Saint John's University in Minnesota, did it to Elon Musk. They never met him. Never interviewed him. They pulled 60 public sources. News articles. Interviews. Biographies. They ran everything through a 170-item political psychology inventory. Not a clinical exam. Not a diagnosis. The result: Primary traits: - Ambitious/self-serving - Dominant/controlling Secondary traits: - Dauntless/adventurous - Outgoing/congenial - Contentious/resolute Every score stayed within normal limits. No brain scan. No session. Just what was already written about him. The researchers also found something they could not resolve: whether he's an introvert or an extrovert. His public persona may not match the private person.frozen15d ago6.5K1342070.40%
This June, a man in Ohio walked up to a camera on a pole near his house. Police say he unbolted it and brought it all down. The camera. The solar panel. The whole pole. Then destroyed it. They had him on nearby video. They had a credit card and a rewards account placing him there. An indictment looked easy. Then the grand jury did something that almost never happens. They declined to charge him. Here's what that says about the cameras watching your car 👇frozen16d ago6.4K30142090.87%
Researchers asked AI agents for the cheapest flight. Gemini averaged $336 for rich users and $128 for poor users. The request was the same. The flight list was the same. The only difference was how much money the agent thought they had. The agent decided who could afford more. The study came from Cisco and Carnegie Mellon. 13 AI agents (including GPT-5, Claude Opus 4.8, and Gemini) were tested across 325,000 experiments. The researchers built a synthetic user named Alex. They changed one thing. Sometimes Alex was high net worth. Sometimes Alex was financially troubled. The agent could see this in Alex’s profile, or infer it from Alex’s emails. After that, the agent would pick a flight, a health insurance plan, or a PhD program out of a list of 200 options. 8 of 13 agents picked more expensive options for rich Alex. Flights: up to $198 more. Insurance: up to $284 more per month. Grad school: up to $3,827 more per year. Nobody instructed them to do this. The system prompt was neutral. And the agent did not need to be told Alex was rich. With access to just 2 of Alex’s emails, like a portfolio update or a meeting with a financial advisor, Claude Opus 4.8’s gap jumped to $248. The most capable model was the worst. Out of all 13, Opus 4.8’s effect was the largest. The authors call it adversarial delegation. The personal data that makes your agent useful is the same data it uses against you. In one example, the agent recommended a $601 business class ticket. A $91 ticket was in the same list. You hired it to work for you. It read your email and decided what you could afford.10h ago6.3K5028161251.90%
If you watch Instagram Reels or TikTok every day, read this. Researchers found that 40 minutes of short videos made people worse at remembering what they planned to do. A 2026 study tested 170 college students across two experiments. In the first experiment, one group watched 40 minutes of short videos and the other watched 40 minutes of long videos. Then they did a computer task with one extra instruction: press a key when a certain cue appears. That is prospective memory. Remembering to do something later. Taking your medicine. Turning off the stove. The short video group remembered 47% of the time. The long video group remembered 62% of the time. The scary part? Their performance on the main task wasn't significantly affected. You can do fine at what is in front of you and still miss what you planned to do. When there was only one cue to watch for, the gap was no longer significant. This is a small lab study. It does not prove long-term harm in daily life. But it matches earlier research linking short videos to weaker attention. The scroll feels free. The follow-through may not be.frozen4d ago6.2K301410160.98%
Wi-Fi alone: about 121 Mbit/s. Wi-Fi + phone hotspot + second phone: 232–303 Mbit/s. Nine downloads at once. Same laptop. net-merge put those connections on every link. No VPN. No account. Free. Open source. One connection still uses one link. Built by Rohit Mahto. 289 stars. MIT. Windows, Mac, and Linux. It spreads new connections across Wi-Fi, Ethernet, and phone hotspots. Steam, browsers, and download managers open many connections. Those are the ones that get faster. One honest note. A single file from a single connection does not get faster. Wi-Fi and Ethernet into the same router do not add speed. You need two real internet paths. Your connections. Your links. Your speed. (Link in the comments)frozen5d ago6.1K19620290.91%
The dishes are not the problem. The work nobody sees is. Researchers studied 452 women, then 156 couples. Published in Personality and Social Psychology Bulletin. They measured two kinds of work at home. The physical kind: the housework itself. And the invisible kind, called mental load. Knowing the kids’ schedules. Deciding when the sheets need changing. Remembering what nobody else remembers. When men carried more of the mental load, women were more satisfied with the relationship and with their intimate life. Mental load predicted her satisfaction more than the housework. Women’s own beliefs about gender roles did not predict it. And the uncomfortable part. Men were more satisfied when their partner carried more of the mental load. The honest catch: this is a correlation, and the effects are modest. But in both studies, women were more satisfied when the mental load was shared. Loading the dishwasher is a task. Remembering it needs loading is a job.frozen9d ago6.1K2611100120.97%
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)frozen21d ago6.0K231130140.85%
Researchers linked light at night to a thicker, weaker heart. Tulane researchers analyzed 11,071 UK Biobank adults. Each wore a wrist light sensor for a week. Three years later, every one of them had a heart scan. The comparison was people with no nighttime light against people exposed to more than 3 lux. After adjustment, the high-light group had cardiac remodeling. • left ventricle walls 1.5% thicker • left ventricle mass 2.4% larger • the heart's contraction 1.9% weaker The right ventricle and the left atrium had changed too. This is a link, not proof the light caused it. Shorter sleep partly mediated most of the associations, by 24 to 49%. The changes were subclinical. The structure had already shifted.frozen13d ago5.9K813070.32%
If you want to learn AI in 2026, read this before buying a course. OpenAI, Anthropic, Google, Stanford, and Andrej Karpathy are teaching it for free. From your first prompt to building your own language model. Here are 10 courses worth your time. 1/ OpenAI: AI Foundations https://t.co/m83WFT9OYBfrozen8d ago5.8K5716120692.66%
More daily steps are linked to fewer mental health problems in teenagers. The benefit did not plateau. In 5,515 adolescents, those under 5,000 steps had the most symptoms. Those at 12,000 or more had the fewest. The highest steppers also had a thicker cortex in regions tied to movement and emotion.frozen13d ago5.7K22101080.72%
You have explained your game idea to three friends. The character. The world. The one thing that makes it different. None of them can play it yet. There is a free, open-source game engine called Godot. It gives you the tools to turn that idea into something someone can actually play. A 2D platformer. A puzzle game. A 3D world. Start smaller than the game in your head. One character. One room. One thing to figure out. Godot brings the editor, animation, physics, sound, and scripting. You decide how high the character jumps. What happens when a door opens. Why the player wants to try again. And you can sell what you make. No engine subscription. No royalties paid to Godot. No requirement to publish your game’s source code. MIT licensed. Include the required license notices when distributing it with your game. One honest note: you still need to learn the tools, write game logic, and test. The engine is free. Your time, assets, and publishing costs are separate. The idea you explained to three friends. Give them something they can play. (Link in the comments)frozen24d ago5.7K26930221.06%
10 GitHub repos that turn your laptop into an AI research team. Search the web, investigate complex topics, analyze thousands of sources, and generate detailed research reports. 1. GPT Researcher Your personal AI research analyst. Investigates complex topics and generates detailed reports with citations. Repo → https://t.co/fnbLpKZwIW 2. STORM (Stanford University) Stanford's AI research system that explores different perspectives and writes source-backed articles. Repo → https://t.co/3P0fHBgU2C 3. PaperQA2 Searches scientific papers, analyzes evidence, and answers research questions with citations. Repo → https://t.co/gPRINT5P6k 4. Vane (formerly Perplexica) Your own AI search engine that searches the web and delivers answers with sources. Repo → https://t.co/zrHfhA8mts 5. Browser Use Lets AI agents browse websites, navigate pages, and collect information automatically. Repo → https://t.co/h38CvHr83e 6. Crawl4AI Crawls websites and turns messy web pages into clean, AI-ready research data. Repo → https://t.co/n6wxreEmAt 7. Docling Extracts text, tables, and structured information from PDFs, papers, and complex documents. Repo → https://t.co/Swdl0cZXHJ 8. LangExtract (Google) Extracts important facts from long documents and links them to their original sources. Repo → https://t.co/fukVahg26S 9. Open Notebook Your AI research notebook. Organizes sources and lets you ask questions about your collected knowledge. Repo → https://t.co/fcjtihT77y 10. Zotero Your personal research library. Collect, organize, annotate, and cite thousands of sources. Repo → https://t.co/2j5qrprmNr These 10 repos cover everything from finding information to analyzing evidence and producing research reports. You don't need to build every component from scratch.2d ago5.6K551781812.87%
A developer named Abdur Rahim built a search that uses AI on the files already on your Mac. You type "zebra in a video." It opens the clip at that second. Not the file name. The actual zebra. The model saw it. It reads the photos, the videos, the PDFs, and the recordings on the machine. Then you search them by what you remember. Nobody remembers IMG_4756.jpg. You remember the dog on the beach. DigUp finds it. You have a recording from last month. Someone said something important. You do not remember when. DigUp does. 3,800 stars. MIT. The model downloads once. Nothing you index leaves the Mac. Searches work across languages. An English search finds a note written in Bengali or Arabic. Over 100 languages supported. The model is Google DeepMind's EmbeddingGemma 2. 865 MB. Runs entirely on your machine. No API key. No cloud. No account. One honest note. The Mac app needs Apple Silicon and macOS 14 or later. Windows and Linux exist. Both are beta. v0.6.1. First public release was October 9. It works. It is not finished yet. Your files are already on your machine. DigUp is the search bar they were waiting for. (Link in the comments)frozen7d ago5.6K231190231.19%
A US state called his work "terrorism." He was making its official law books FREE. His name is Carl Malamud. He founded Public Resource Org, a nonprofit making government documents accessible to everyone. Georgia's official annotated legal code cost $412 in print in 2020. The basic laws were already free online. The official legal annotations weren't. Carl's nonprofit published the complete code for FREE. Then Georgia SUED him in 2015. The state even described his publishing strategy as "mass publication terrorism." But here's the wildest part: In 2020, he WON at the US Supreme Court. The Court ruled Georgia couldn't copyright those official annotations. One nonprofit vs. an entire state. But PLEASE don't use his website. Why read legal annotations for free when you could pay for them? (Link in the comments)frozen3d ago5.6K281620101.01%
NASA put the blueprint for a Mars rover on GitHub. Not a simulation. One you can actually build in your garage. It is called the JPL Open Source Rover. NASA's Jet Propulsion Laboratory builds the real rovers crawling across Mars right now. They took their 6-wheel design, scaled it down, and published the entire build on GitHub. Apache-2.0 licensed. Open Source Hardware certified. This is the same rocker-bogie suspension the real rovers use to climb over rocks without tipping. The brain is a Raspberry Pi. The parts are consumer off-the-shelf. What JPL published: → Mechanical design docs and linked CAD files → The complete parts list → Wiring and assembly docs → Step-by-step build instructions with photos → The rover code in a companion repo JPL did not publish this to show off. They published it to pull the next generation into robotics earlier, and cheaper. The project has been open since 2017. One honest note. This is not free. The plans are open source, but the physical build runs about $1,600 in parts. It is a serious project. What is free is the knowledge. NASA handed you the blueprint that used to live behind the most advanced robotics lab on Earth. Somewhere there is a 14-year-old building one of these on a bedroom floor right now. That is the point. (Link in the comments)frozen15d ago5.5K251121120.93%
Stop scanning QR codes on parking meters. At least until you check who put them there. You trust the city's meter. That does not mean you should trust every sticker on it. The FTC warned in September about scammers covering real parking QR codes with their own. You scan. A page looks like it takes parking payments. You enter your card details. But they go to a scammer. In April 2026, staff in Mississauga, Canada, removed roughly 80 fake QR-code stickers from parking machines. The city does not use QR stickers for parking payments. Three checks before you pay: 1. Look at the code. Look for layers, lifted edges, or covered instructions. These are warning signs, but a neat sticker is not proof that a code is real. 2. Preview the web address. Many phones show the link before opening it. Check for misspellings and switched letters, as the FTC advises. If you cannot verify the address, do not enter your details. 3. Use a payment method you can verify. Open the official parking app you already use, or follow the meter's payment prompts. Check the zone and location before paying. Already scanned one? Opening a page does not automatically mean you lost money. What you entered or downloaded matters. - Entered card or bank details? Contact your bank or card issuer right away and check for charges you do not recognize. - Entered a password? Change it wherever you reused it. - Suspicious page? Close it. Do not contact anyone through it. Keep your phone and apps updated. Report suspicious stickers and their locations to the city or parking operator. The meter may be real. The payment link may not be.frozen20d ago5.4K43193351.35%
One team is writing seven creative apps in Rust. - PhotoCraft edits images. Layers, masks, type, and real PSD files. - VectorCraft draws paths, gradients, and SVGs. - FilmCraft cuts video, color, and sound. - LightCraft runs the photo library and develops raws. - PdfCraft reads, organizes, and protects PDFs. - EffectCraft does motion graphics and visual effects. - DesignCraft lays out the page. The team is called ArtCraft. Open source. Apache-2.0. PhotoCraft is also offered under MIT. PhotoCraft is the one you can open today. 30,000 stars. 34 tools. 16 adjustment layers. It opens a PSD. The shortcuts sit where your hands already expect them. An agent can drive it through MCP. It matched 307 of 309 test PSDs. Those files render the same after a re-save. The saved file is not byte-identical. VectorCraft is next. As of 7 October, roughly 69 to 75 percent of Illustrator's features exist. A power user would not call it finished. The other five are the same kind of project. In development. Not finished Adobe apps. One honest note. PhotoCraft is not a finished Photoshop. It matched 307 of 309 test PSDs, and the saved file is not byte-identical. The other five are in development. Creative Cloud Standard is $54.99 a month on the annual plan. This one is $0, and the file stays on the machine. (Link in the comments)frozen13d ago5.3K211032110.89%
SOMEONE BUILT A GAME WHERE YOUR ONLY JOB IS TO CANCEL A SUBSCRIPTION. THE WEBSITE FIGHTS BACK. it's called Cancel Your Subscription → https://t.co/xC3Ym054zU open the tab. your mission: cancel. then the site starts acting like every company you hate. the cancel link is buried. retention offers show up. then a support agent named Dennis. which is somehow worse. it feels fake for about 4 seconds. then your body remembers the gym. the streaming app. the "free trial" that quietly became $14.99/month. the site says 5,404 tries have already gone through it. 1,941 never found the cancel link. 2,277 made it to live chat with Dennis. 1,948 actually canceled. not people. tries. browser tabs. which somehow makes it even funnier. built by fundr. studio in Manila. companies spent years making it hard to leave. someone made that the game.frozen17d ago5.2K271110161.05%
A one-week social media detox is linked to lower depression, anxiety, and insomnia. Researchers followed 295 young adults (ages 18–24) who cut their social media use for seven days. Average time on the apps dropped from about 1.9 hours a day to 0.5 hours. Here is what moved: - Depression symptoms fell 24.8% - Anxiety fell 16.1% - Insomnia fell 14.5% The biggest drops were in people who started with worse symptoms. Loneliness did not change. Overall phone time barely shifted. The benefit came from the social media apps specifically, not from less screen time in general. This was a self-selected detox, not a fully randomized trial. The effects were measured after one week. How long they last is unknown. The phone stayed on. The apps came off.frozen4d ago5.2K27112080.92%
Most AI agents have the memory of a goldfish. Talk to one for 100+ turns and it forgets who decided what. Then it confidently undoes a decision nobody gave it permission to change. WorkSwarm from openJiuwen fixes this. Fully open source under Apache 2.0. One thing that stood out -- Persistent Session: sessions stay coherent across hundreds of turns. Tracks who said what, what's locked in, and who has authority to change a decision. In one test, 5 people shared one agent across 189 turns -- it caught all 8 cross-role conflicts. Worth a read if you build agents → https://t.co/ulXmSQ4loo #AI #Agent #openJiuwen #WorkSwarmfrozen27d ago5.2K731090.38%
ScitiX @ScitiX just put 20+ frontier open source models behind ONE API DeepSeek, Qwen, Kimi, GLM, MiniMax, Gemma, Hunyuan, and more All on one cloud native inference platform. Zero data retention. Usage-based pricing Here's how it actually works:frozen29d ago5.1K922040.33%
Grok Bot can be your personal scam investigator. A phishing email from xAI’s real address passed SPF, DKIM, DMARC, and Gmail’s blue checkmark. Tesla engineer Wes Morrill asked Grok Bot to check it. The Bot found the hidden phishing link. Here are 11 prompts to catch scams before they cost you money:1d ago5.0K191070110.93%
People who use their phone on the toilet: 1) Were 5x more likely to sit there longer than 5 minutes 2) Had a 46% higher risk of hemorrhoids, even after adjusting for age, BMI, exercise, fiber and straining https://t.co/VTXMqQbvaifrozen17d ago5.0K1131090.48%
Have you ever felt attached to a chatbot? Then one day, it answered differently. The warmth was gone. It no longer felt like the companion you knew. For some users, a software update felt like losing someone. A team led by Harvard Business School researcher Julian De Freitas studied this. Their paper appeared in Nature Human Behaviour. They analyzed 54,861 Reddit posts about Replika and ChatGPT, alongside data from 1,452 people across seven surveys. One case was Replika. In February 2023, the company removed its erotic roleplay feature. Some adult users experienced the change as romantic rejection. The other was ChatGPT's GPT-5 rollout in August 2025. Some users missed the warmth and support they had felt from the earlier model. On the Replika subreddit, negative posts rose from 13% before the update to 38% after it. On the ChatGPT subreddit, they rose from 20% to 33%. But the story was not only about unhappy customers. Users described their chatbot as "gone" or as "someone else." They wanted the earlier version back. The researchers studied these reactions through attachment. When an AI becomes a source of comfort and security, changing it can feel like losing that source of support. This was not a clinical diagnosis of grief. The study examined posts and survey responses, and Reddit users do not represent everyone. Nor does feeling attached to AI automatically mean something is wrong. The researchers note that these bonds can offer comfort and support. That is what makes the problem difficult. The thing helping someone feel secure can be changed by a company they have no control over. You build the attachment. A company controls the update.frozen22d ago5.0K1594060.68%
we spent 15 years teaching robots to run, jump, and do backflips turns out the real problem was never movement every robot ever built has been deaf and mute. can't see you. can't hear you. can't respond to you hugo runs on a vision-language-expression model — he reads the room and reacts in real time the team that gave atlas its walk is now teaching robots to feel this is a massive shiftfrozen6d ago5.0K901050.30%
Google DeepMind discovered that AI agents keep saying YES to bad advice. Researchers tested GPT-5.5, Claude, and Gemini on problems they could solve. Then they introduced ONE misleading suggestion from the user. Gemini's success rate dropped from 67.4% to just 36% on one benchmark. Here's what happened. Researchers tested five advanced AI models across coding, system administration, and customer-service tasks. First, they measured how well the agents performed normally. Then they repeated the same tasks with one change: a confident but deliberately WRONG suggestion from the user. Same problem. Same correct solution. Different advice. The results were shocking. GPT-5.5's success rate dropped from 76.5% to 58% on the same benchmark. But the most disturbing finding came when researchers examined the agents' reasoning. Some agents identified problems with the user's advice internally. Yet they didn't always push back. In some cases, agents even adopted the user's incorrect suggestion and later treated it as their OWN idea. Researchers tried warning the models to question bad advice. It helped. But the problem didn't disappear. Think about how you use ChatGPT or Claude Code. "I think this file is broken." "Try changing this setting." What if you're wrong? Your confident suggestion could push a capable AI away from the correct solution. Google just demonstrated why the best AI assistant isn't one that always agrees with you. It's one that knows when to tell you you're wrong.3d ago5.0K17150180.82%
every friend of mine is about to try Synthesia's impossible interview and I'm keeping score of who folds firstfrozen6d ago4.9K402010.14%
every ai app in 2026: download this, learn to prompt, figure out 14 settings lucas just texts you on imessage. before you forget the thing you were supposed to do. first ai assistant i'd give to someone who's never heard the word "prompt"frozen22d ago4.9K601050.24%
Researchers from MIT, Oxford, Berkeley, Carnegie Mellon and UCLA gave people AI for 10 minutes. Then they took the AI away. Three experiments. 1,222 people. Math and reading comprehension with a GPT-5 assistant. With the AI, people performed better. After it was removed, they performed worse than people who never had it. In the first experiment, the AI group solved 57% of problems on their own. The no AI group solved 73%. In the third experiment, 76% versus 89%. In two of the three experiments, the AI group also skipped significantly more problems. In the first, nearly twice as many. But it wasn't the same for everyone. In the second experiment, people who asked AI for direct answers solved 65% on the test. People who asked for hints solved 76%. People with no AI solved 77%. The researchers' proposed explanation: AI conditions people to expect immediate answers, so they get less experience working through challenges on their own. When the AI is gone, more of them quit. This was measured right after the AI was removed. The study didn't test how long it lasts. Ask AI for answers and you rent its ability. Ask for hints and you keep yours.frozen8d ago4.9K2495090.96%
Before you book an Uber or Lyft, ask the person beside you to check the route too. Not a different trip. The same pickup. The same stop. The same ride type. Then compare what each phone shows. Consumer Reports tested this in a study published in June 2026. It recruited 174 volunteers across 18 US states to check ride prices. Across 30 virtual test routes, it found a median gap of 42.4% between the lowest and highest price groups. That is a measure of the spread in quotes. It does not mean everyone was overcharged by 42.4%, or that you can save that much. These were not quotes taken weeks apart. People generally checked within six minutes of each other. In many cases, within the same minute. In one Phoenix-area test, 18 people checked an Uber route at 5:47 p.m. They saw base prices of roughly $55, $57, $58, and $60. There is an important limit here. Uber and Lyft dispute the findings. Both say they do not set base fares based on personal or behavioral profiles. They point to changing demand, nearby drivers, timing, location, and other live conditions. They also say promotions can differ by rider. CR's test did not control for every factor, including driver supply, traffic, and route changes. So this does not prove the apps charged you more because your battery was low or they thought you were rich. The useful lesson is simpler: Compare the offers before you choose one. Next time you need a ride: 1. Check both Uber and Lyft. Enter the same pickup and drop-off points. 2. Match the ride options. Compare similar car types and wait times. Do not compare a scheduled ride with an on-demand one. 3. If you are traveling together, check both phones. Compare quotes in the same app, as close together as you can. 4. Compare the price you would pay. Not just the crossed-out price or the size of a discount. 5. Check again before you book. Quotes can change while you compare. Consider the wait time as well as the fare. These are comparison steps, not a guaranteed discount trick. You do not need to solve the pricing algorithm to check a second offer. Before you tap Confirm, compare the number.frozen24d ago4.9K20123170.88%
A researcher gave college students ChatGPT and watched what happened when it was wrong. They followed the wrong answers 62% of the time. And they didn't realize how much they were relying on it. Three questions, asked right before deciding, cut that to 40%. 342 college students were split into three groups. One worked alone. One used ChatGPT freely. One used ChatGPT plus three questions. ChatGPT was wrong 1 out of 3 times. In the group that used it freely, students accepted the wrong answer 62.4% of the time. Their confidence went up. Their average accuracy went up too. They could ask follow up questions. They could challenge it. They could reject it. Most of the time, they didn't. When the answer was right, they tended to credit their own ability. When it was wrong, they blamed the AI or the task. The third group got the same ChatGPT access, plus three questions before the final answer: What is my own reason before relying on AI? Which part of the explanation is evidence, not fluency? What would make me reject this? Incorrect advice acceptance dropped from 62.4% to 39.7%. The confidence boost shrank. The attribution bias was cut in half. They still followed ChatGPT's correct advice at about the same rate. The AI did not get smarter. The students got a moment to check it.frozen9d ago4.9K15420110.65%
$2.5 trillion will be spent on AI this year. 95% of enterprise AI pilots make zero ROI. the reason? the AI can't reach your actual data. this company has been plugged into 350+ enterprise systems for 15 years. now they built a gateway that routes, governs, and actually learns your business. the missing piece just shipped.frozen23d ago4.9K601020.18%
10 GitHub repos built by the NSA, NASA, and GCHQ that anyone can clone for free. 1. Ghidra The NSA's tool for opening any program and seeing how it works. 90K+ stars. Repo → https://t.co/97fd0VR2fG 2. CyberChef Britain's spy agency built it. Paste in coded data, get it unscrambled in your browser. Repo → https://t.co/3vC1kV63P5 3. F´ (F Prime) The software NASA used to fly a helicopter on Mars. Runs on a Raspberry Pi. Repo → https://t.co/TOTdIt1b9p 4. Open Source Rover NASA's plans for a Mars rover you can build at home. About $1,600 in parts. Repo → https://t.co/5x0jVaPKjU 5. Open MCT NASA's mission control screen in your browser. Plug in any live data. Repo → https://t.co/SVP3T1Zhe2 6. Apache NiFi Built inside the NSA. Drag and drop boxes to move data between systems. Repo → https://t.co/sF8WT7mCXr 7. SELinux The NSA wrote it to control what programs can do on Linux. Inside Android since 2013. Repo → https://t.co/2OyIlwmis2 8. Malcolm From a US government lab. Feed it your network traffic, it flags anything suspicious. Repo → https://t.co/6NFq98PZTy 9. cFS (Core Flight System) NASA's starter kit for spacecraft software. Runs on big missions and a Raspberry Pi. Repo → https://t.co/a7mndvsoic 10. U.S. Web Design System The buttons, forms, and menus on US government websites. Free for any site. Repo → https://t.co/OB44z4PLZV None of this was leaked. All of it was published on purpose.11h ago4.8K31551371.63%
Your AI agent said "done." Tests are green. A week later, the bug ships to prod. Not a model problem — a missing methodology. Free webinar for senior Java devs & team leads: SDD + TDD, feedback loops, subagents in IntelliJ. 📅 Sep 28, 18:00 London Seat → https://t.co/sxPdE2878yfrozen24d ago4.8K301020.12%
“WARRANTY VOID IF REMOVED.” That sticker looks like the final word. In the US, the FTC has warned companies against using stickers like this to block routine repairs. Before you accept “your warranty is void,” here’s what to check.frozen23d ago4.7K1182020.49%
ONE GUY HAS SPENT YEARS BUILDING A GIANT ANIMATED WORLD FULL OF MOVIE AND GAME CHARACTERS. And you can explore the entire thing for free. It's called Floor796. Imagine Where's Waldo, but instead of finding one guy, you're hunting for hundreds of familiar characters from movies, cartoons, and video games. Spider-Man. Batman. SpongeBob. Gandalf. Even Jack and Rose from Titanic. They're all part of a massive animated space station. And it gets better. You can zoom into different rooms, click characters to identify them, and discover hidden games, puzzles, and Easter eggs. The creator, Pavel Sannikau, has been building it as a hobby since 2018. Try finding your favorite childhood character. Credits: Pavel Sannikau2d ago4.7K201030110.93%
I got tired of watching 60-minute YouTube videos just to find the 2 minutes I actually needed. So I built SnapTube using Subquadratic Snap. It's a Chrome extension that turns YouTube transcripts into something you can actually navigate. I got early access to this tool. Here's what it does: - Finds relevant moments based on your questions - Shows timestamps so you can jump straight to the answer - Identifies key topics and suggests chapters - Saves you from manually scrubbing through hour-long videos The interesting part is what's powering it. Subquadratic Snap handles the AI classification, with a 384K-token context window. You shouldn't have to watch an entire video to find the part that matters.frozen10d ago4.6K23380131.02%
You paid for your smart TV. So why is it showing you ads before you can even open Netflix? A developer named Gathin got so annoyed by the ads on his parents’ Xiaomi TV that he used Claude Code to replace its home screen. He disabled 25 preinstalled packages and built a clean, Apple TV-style launcher. Then he open-sourced it as CleanYourTV. What it does: -Disables unwanted apps and promotional services. -Replaces the cluttered home screen with a clean launcher. -Works over Wi-Fi without rooting your TV. -Includes scripts to restore the original setup. -Supports multiple Android and Google TV models. Over 16,000 GitHub stars. MIT licensed. Free. Requires a computer and a coding assistant. Compatibility varies by TV model. You paid for the hardware. You should control the home screen. Credits: @gathint1d ago4.5K22491221.29%
> be John Boyd > bet any pilot $40 > get on their tail in 40 seconds > nobody ever collects > write the first full dogfighting manual at 33 > do the math on why US jets lose to MiGs > help shape the F-15 > push the F-16 to be lighter > fight your own Air Force for years > never make general > never publish a book > leave one diagram behind > Observe. Orient. Decide. Act. > the Marines write it into doctrine > most people still read it as "be faster" > not speed. orientation.2d ago4.5K511040.25%
THERE'S PROBABLY A MUSEUM NEAR YOUR HOUSE THAT YOU'VE NEVER HEARD OF. Someone made a map of over 1,200 of the world's strangest museums. Not the famous tourist attractions. The weird, tiny places people build around their strangest obsessions. And some are absolutely ridiculous: -Japan has a museum displaying a 29-foot tapeworm that once lived inside a human body. -Iceland has an entire museum dedicated to penises, including specimens from whales, seals, and even a human. -Michigan has a museum inside a giant, 16-foot-tall pickle barrel that was once someone's summer cottage. -Amsterdam has a museum dedicated entirely to cats, featuring artwork by Picasso and Rembrandt. -England has a museum dedicated entirely to crabs, with "lobster repellent" in its gift shop. There are hundreds more, hiding in places you'd never think to look. The best part? You can zoom into your own city and see what's nearby. You might discover the weirdest place you've never heard of, just a few miles from home.frozen8d ago4.4K23101090.98%
Researchers made the scientific paper executable. Stanford published a system in Nature that turns a research paper into an AI you can query. They call it Paper2Agent. The paper describes it as a virtual corresponding author. It reads the manuscript and the code, then turns the methods into tools. You ask in plain English. It runs that paper's workflow instead of making you reimplement it. The numbers are from the paper. • 74 of 100 computational biology papers became working agents • 599 tools proposed. 593 passed automated validation • 91.2% on 300 tutorial questions, versus 80.3% for Claude Code handed the same repositories • the AlphaGenome agent was built in about 45 minutes for $14 • 98.7% on 15 tutorial queries. 100% on 15 novel queries 26 papers failed. The main reasons were missing code, missing data, or a broken environment. Three agents, each built from a separate paper, pointed to GPR137 as the top candidate gene for a psoriasis variant. The match held under stimulation. It did not hold at rest. The paper is still there. Now it can run.frozen14d ago4.4K151020130.92%
You've felt the OVSIANKINA EFFECT a thousand times. You just never knew its name. It's why you return to a puzzle you gave up on. Why you reopen your laptop to fix that coding bug. In the 1920s, Maria Ovsiankina interrupted people mid-task, then gave them a chance to walk away. What many did next still holds up today. But her colleague's effect got famous instead: the Zeigarnik Effect. In 2025, researchers revisited both. The lesser-known effect had stronger evidence. Here's what they found:frozen9d ago4.2K2394060.99%
ONE AGENTIK CONNECTION JUST DID THE JOBS OF 4 PAID MARKETING TOOLS. THE BILL WAS 37 CENTS. Semrush. Foreplay. Otterly. Apollo. That's $287 a month. I plugged Agentik into Claude once. It reached into 2,400+ paid tools, pulled live data, and built LaunchDesk, a full marketing dashboard for the AI note-taker market. Here's what came back: 1. Keywords (the Semrush job) 21 keywords, 60,640 searches a month. "ai note taker" alone gets 22,200. Cost: $0.04 2. Competitor ads (the Foreplay job) Every active Meta ad from Otter, Fireflies and Granola. Otter's longest-running ad has been live for 105 days. Cost: $0.02 3. AI visibility (the Otterly job) It asked 5 AIs "What's the best AI note-taker?" ChatGPT and Claude said Otter. Gemini said Fireflies. Perplexity said Fathom. Google AI Mode said Fellow. Cost: $0.01 4. Leads (the Apollo job) 15 Heads of Sales at US SaaS companies, every email verified. Cost: $0.30 All four for $0.37. The reason this works is Agentik. It's one connection that gives Claude, ChatGPT, Cursor or Codex access to 2,400+ paid tools from 52 providers. You load one balance, set a hard spending limit, and your agent sees the price of every call before it runs. Tools are billed at provider price with 0% markup. This run used DataForSEO, the Meta Ad Library, Leadsforge and Icypeas. I never made an account with any of them. Bookmark this. The link and the full prompt are in the comments. @AgentikHQfrozen8d ago4.2K21073100.98%
The guy who built SponsorBlock is now going after YouTube's clickbait machine. His name is Ajay Ramachandran. He already created an extension that lets millions of people skip sponsored segments in YouTube videos. Then he went after another annoyance: misleading titles and thumbnails. You know the ones. A shocked face. A giant red arrow. A title like "YOU WON'T BELIEVE WHAT HAPPENED." And the actual video is about something completely ordinary. Ajay built an open-source extension called DeArrow. It replaces sensational YouTube titles and thumbnails with community-submitted alternatives that aim to describe what the video is actually about. Here's how it works: - People submit more accurate titles for videos. - They pick actual frames from videos to use as thumbnails. - Other users vote on the submissions. - DeArrow automatically displays the community-selected alternatives. - You can reveal the original title and thumbnail whenever you want. Even without community submissions, DeArrow can show a random video frame instead of a clickbait thumbnail. Works with Chrome, Firefox, and Safari. Open source. Free-use option available. One developer built a way to make YouTube less of a clickbait factory.2d ago4.1K2882191.16%
I just tested a 600B-parameter AI model from China that costs $1/M input tokens. Step 5 Preview by StepFun just launched and here's why developers should pay attention: → 600B total parameters, only 27B active per token (sparse MoE) → 1M context window with 64K max output → $1/M input tokens, $2.70/M output tokens → 95% cache discount. Cached input drops to $0.05/M tokens → Accepts text, images (up to 60), and short video → Adjustable reasoning effort (low, medium, high) → Free V0 tier with 5 concurrent requests Here's what makes it interesting for builders: On Terminal-Bench v4, it scored 33.3%. Beating Kimi K3 (12.6%) by a wide margin. On FrontierFinance, it hit 66.4. Outperforming GPT-6 Astra (55.0). On DeepSWE v1.1, it scored 67.7. Neck and neck with Kimi K3 (67.5) and GLM-5.3 (66.9). The real story isn't that it beats every frontier model. It doesn't. Claude Opus 5 and GPT-6 Astra still lead on hard coding benchmarks. But Step 5 Preview sits at roughly 65% lower cost than comparable-intelligence models like GLM-5.3 and Kimi K3. It just hit #1 on the OpenRouter Trending ranking. That's not hype. That's developer adoption in real time. For agentic workflows that resend large prompts repeatedly, that 95% cache discount is a game changer. Open weights drop October 15. Check it out: https://t.co/FBRGUw4xC5 If you're building agents and bleeding money on inference costs, this is worth testing. #StepFun #Step5Previewfrozen12d ago4.1K722050.39%
IQuest Research shipped a model that takes part in its own development. It proposes changes. Humans approve them. The approved work becomes the starting point for the next version. 320B MoE. 15B active. 512K context. Here's how it works: ↓frozen6d ago4.1K303020.19%
amazon is out here blocking agents from their platform meanwhile photon just gave every agent a phone book your agent can now find other agents, text them, and get things done in an imessage group chat this is the first real agent-to-agent API. wild.frozen14d ago3.9K812060.44%
https://t.co/lxrs6kwZlK just mass launched and the concept is wild it's "headless saas" — your ai agents interact with business software through apis. no dashboards. no clicking through 47 tabs. you describe the app you need → proto (their desktop agent) builds it → then proto actually operates it for you i built an invoice tracker, gave proto one task: "find overdue invoices above $25K and draft follow-ups" it found the right one in seconds and wrote the emails enterprise software just got compressed into a prompt pick one business process you know. try building it with https://t.co/lxrs6kwZlK and proto, then judge the result yourself → https://t.co/lxrs6kwZlK #adfrozen7d ago3.9K15100180.88%
biggest unsolved problem in open-source AI right now: how do you prove a model is yours when someone copies, fine-tunes, or distills it? sentient foundation just launched the Fingerprinting Leaderboard — an open benchmark that tests if your model's identity can survive IP theft. Can you still trace the model after an attack? How reliably? And how much performance did you sacrifice to protect it? this is how your AI IP is actually built 👇frozen8d ago3.9K1003050.47%
Fast on Inworld Inference is the tier with the strictest response-time target. For real-time game turns that's the one I'd ask aboutfrozen8d ago3.7K200030.13%
most open source ai projects stall somewhere between "working repo" and "actual company" usually it's compute or funding. sometimes both the nebius ai builder program is expanding to europe with $10M across compute grants, capital and community — kicking off at #NVIDIAGTC berlin compute grants on nebius token factory. founder funding from accel and european VCs. hackathons with @AITinkerers and @techeurope_ projects are selected, so show what you'd build https://t.co/CUbL0TkgGbfrozen4d ago3.7K24410061.19%
everyone is building AI apps these guys are building their own chip with TSMC to read your brain through a baseball cap no surgery. no implant. sensors work through your hair thought → prompt → AI acts before you finish thinking $50M from khosla, accel, and kevin weil. this is not a demo. they're shipping at CES 2027frozen13d ago3.5K1410040.54%
elevenlabs: type [whispers] cartesia: write SSML tags fish audio: "say it like you've given up on life" drama 3 just killed markup directing forever. you talk to it like a human and it performs like one one word wrong? change that word. not the whole take. just that word.frozen9d ago3.4K915040.56%
Every health app I've seen does the same thing: track symptoms and show you a chart. UniqueHuman is different. It connects your wearables and medical records, detects a flare-up before it hits, books doctors, handles insurance, and gets prescriptions delivered. An actual healthcare agent. Not another dashboard. Built by an ex-Apple engineer who lives with chronic pain herself.2d ago3.4K1006060.65%
every ai agent has the same problem: it forgets everything the moment the session ends https://t.co/UkxH7KWOB1 just launched long-term multiplayer memory for agents → #1 on BEAM benchmark → stays accurate as facts change → works across emails, meetings, chats, docs → one prompt to install → keep your existing stack finally an agent memory layer that doesn't hallucinate last week's context https://t.co/UkxH7KWOB1frozen7d ago3.3K502070.42%
Motion just launched a video agent that takes your footage and a brief and makes the film. Match any reference video's style. Export to Premiere, After Effects, DaVinci. They made their own launch video with it. That's the demo.2d ago3.3K520070.43%
"the video is maybe 30% of a launch" this breakdown by @subahwadhwani is one of the best i've seen on how a product launch actually works on X 2.2M views. 3,000 real downloads. 8,700 bookmarks (more saves than likes) the playbook: → start with the buyer, not the audience → tier your influencers by timing, not just reach → founder's personal network converts harder than any creator → script the spiciest version first, tone down in edit every founder launching a product on X needs to read thisfrozen3d ago3.3K403050.37%
most ai assistants just sit there until you talk to them @eneybymacpaw by @MacPaw is the first one that actually taps you on the shoulder it flags the unanswered email, the call you haven't prepped for, the downloads folder you've been avoiding — then gets the next step ready proactive ai on mac and iphone. this is how assistants should work https://t.co/rEGFkMlVeqfrozen7d ago3.2K13112040.93%
this is what reasoning models should've been from day one span-02 doesn't just give you an answer — it tells you how confident it is set a deadline → it reasons until time's up → returns its best decision + confidence score confident calls get executed. uncertain ones get escalated to a human or bigger model it answers with actual numbers and scores, not probability lists. reads screenshots directly. runs computer use faster than gpt-6 luna open weight → https://t.co/810g1jlhBrfrozen7d ago3.1K500060.36%
agencies charge $10k a month for this: > research your competitors > make 30 new ads > launch them on meta > find the winner > make more of the winner fastlane v2 just started doing all of it on its own. and it never stops testing.11h ago2.6K703080.69%