Vasu (@VaseGod) header
Vasu (@VaseGod) profile picture

TBH, i don't like Vases that much🏺🐘🏺 🤹‍♂️ lAMATEUR JUGGLERl🪑🙋🏾‍♂️I ML/AI👾 l https://nitter.cf/t.co/3J9IWSNgch l Axiom l LoopForge l Eureka Engine

NEW YORK
Joined July 2015
Photo by Vasu (@VaseGod)Photo by Vasu (@VaseGod)Photo by Vasu (@VaseGod)Photo by Vasu (@VaseGod)Photo by Vasu (@VaseGod)Photo by Vasu (@VaseGod)Photo by Vasu (@VaseGod)Photo by Vasu (@VaseGod)Photo by Vasu (@VaseGod)Photo by Vasu (@VaseGod)
Microsoft AI just unveiled seven specialized models at Ignite 2026 alongside MAI Licensed Weights, letting enterprises fine tune frontier weights on private data and host them inside their own Azure tenants. 1. MAI Cyber 2 Flash beats GPT 6 Astra on the CyberGym end to end security benchmark at less than 2 percent of the cost per task, while MAI Image 2.7 takes the number one spot on Artificial Analysis image editing at one fifth the price of GPT Image 2.6 Sunburst. 2. Decision 1.1 routes prompts across models 50x faster and 79x cheaper than GPT 6.1 Sol, while MAI Realtime 1 brings live speech to speech down to 1 cent per audio minute. 3. Under the hood, Microsoft confirmed its second generation Maia 200 AI accelerators are now running production frontier models at 1.7x higher tokens per watt. Enterprise AI is splitting into two clear winners: ultra cheap specialized models running 24/7 on custom silicon, and licensed weights that never leak proprietary company IP. microsoft.ai/news/microsoft-…
Today, we’re releasing MAI-Code-2 and MAI-Cyber-2-Flash! I’m incredibly proud of what the team has accomplished in just 3 months. Thank you everyone. We’ll be sharing more third-party evaluations soon to demonstrate our models’ capabilities. We’re also hiring!
11
Google DeepMind just launched Genie 4 inside Project Genie for AI Ultra subscribers, turning generated 3D worlds into live interactive environments you can edit mid session with text prompts. 1. While exploring a playable world in your browser, you can type natural language commands on the fly to shift the weather, spawn characters, or alter physics alongside real time adaptive sound effects and dialogue. 2. A new Project Genie agent orchestrates multimodal models to build custom worlds from conversation and acts as an in world referee, while Genie 3 gets upgraded to 3 minute sessions seeded from 360 degree panoramas. 3. Everything generated carries SynthID watermarking and can be published to a new community gallery where anyone can step inside and remix other creators' worlds. World models are rapidly moving from passive 6 second video clips to persistent, interactive sandboxes where agents and robots can train before touching the real world. blog.google/innovation-and-a…
Genie 4 is now available in Project Genie. It brings native audio, real-time editing, and a new agent to help you create interactive, virtual worlds.
1
46
Two critical unauthenticated remote code execution flaws (CVE 2026 123633 and CVE 2026 123007, both rated 9.2 Critical) were just disclosed in NGINX, the web server powering over 225 million websites. 1. The worst bug is an HTTP/2 flow control integer underflow that has been hiding in NGINX since version 1.9.5 in 2015, allowing attackers to run arbitrary code even on servers serving plain static files over HTTPS. 2. Wiz discovered both flaws using its AI code scanning harness and confirmed full remote code execution with ASLR bypass using Gemini 4 Carbon and ChatGPT 5.6 Sol, while four other research teams independently found the exact same decade old bugs. 3. Public proof of concept exploit code is already on GitHub, so patch immediately to NGINX 1.30.7 stable or 1.31.9 mainline, or set http2 off in your server blocks right now. When five separate security teams independently uncover a decade old memory bug in NGINX the same month AI code auditing goes mainstream, complexity is no longer a defense. wiz.io/blog/nginx-http2-rce-…
49
The first large national study tracking millions of US students before and after school cellphone bans just found that locked phone pouches produce measurable test score gains. 1. Published today by researchers at Harvard's Center for Education Policy Research, NWEA, and Dartmouth, the study compared student achievement growth before and after strict bell to bell phone restrictions. 2. Schools requiring students to lock phones in pouches like Yondr all day saw modest but statistically significant gains in math and reading equivalent to roughly one to two extra weeks of learning per year. 3. Researchers estimate in school smartphone distraction accounts for up to half of the decade long slide in eighth grade test scores since 2013, meaning recovery will build gradually over several school years. Removing smartphones from classrooms is not an instant magic wand, but locking devices away all day measurably restores lost instructional time. nytimes.com/2026/11/17/upsho…
12
Exactly 350 years ago this month, Danish astronomer Ole Rømer proved light has a finite speed by using one of Jupiter's moons as a cosmic clock. 1. While tracking Io's eclipses at the Paris Observatory, Rømer noticed Io emerged behind schedule when Earth moved away from Jupiter and ahead of schedule when Earth moved toward it. 2. He stunned the French Academy by predicting Io would emerge from Jupiter's shadow 10 minutes late on November 9, 1676, proving light takes time to cross space rather than arriving instantly. 3. His measurement put light's travel time across Earth's orbit at 22 minutes (roughly 225,000 km per second, within 25 percent of modern physics), proving that looking into deep space means looking backward in time. Before 1676, humanity assumed the universe happened instantaneously, and one astronomer timing a moon shadow revealed the cosmic speed limit. sciencenorway.no/culture-his…
10
Apple's Developer System Status page just flipped back to all green after an internal update knocked App Store Connect, iCloud, and Apple Pay offline earlier today. 1. Starting around 12:45 PM ET, developers and users were locked out of App Store Connect, iCloud sign in, Apple Music, Apple Pay, and Apple School Manager for hours. 2. Apple confirmed the disruption was triggered by an internal update rather than a cyberattack, marking the third major tech or banking outage in five days after Microsoft on Friday and JPMorgan Chase yesterday. 3. All 45 Apple Developer services including TestFlight, CloudKit, APNS, and Xcode Cloud now show as operating normally on the status dashboard. When a single internal update can halt app distribution, cloud sync, and school classrooms at once, centralized cloud dependencies remain the biggest single point of failure in software. developer.apple.com/system-s…
1
61
Microsoft AI just unveiled MAI Code 2, matching 96 percent of Claude Opus 5.5's coding performance at just 5 percent of the cost per task. 1. Across DeepSWE, Terminal Bench, SWE Atlas, and FrontierCode, MAI Code 2 averages a 63.2 percent pass rate at $0.31 per task compared to $6.31 for Claude Opus 5.5 and $0.76 for GPT 6.1 Sol. 2. In blind side by side UI evaluations across 258 human reviewers, developers preferred MAI Code 2's frontend apps and dashboards over GPT 6 Sol 67 percent of the time. 3. Its smaller sibling MAI Code 2 Flash beats both GPT 6 Luna and Claude Haiku 5.5 at $0.05 per task and is rolling out across GitHub Copilot, including Copilot Free. Frontier labs spent two years racing for raw benchmark supremacy, and now the real enterprise war is delivering 96 percent of frontier coding quality for pennies per pull request. microsoft.ai/news/introducin…
Introducing MAI Code 2 and Code 2 Flash. Code 2 brings frontier price-performance for everyday enterprise coding tasks. Code 2 Flash brings affordable and high-quality coding for routine work at scale. Try both today in GitHub Copilot.
1
55
OpenAI just launched GPT 6.2 Astra and Sol, cutting frontier agent costs by more than two thirds while hitting a major AI self improvement milestone. 1. On Agents Last Exam, GPT 6.2 Astra beats GPT 6 Astra by 5.2 points at less than one third the cost and beats Claude Fable 5.2 by 5.5 points at just 5 percent of the cost. 2. On Harvey's legal benchmark, Astra more than doubles Claude Fable 5.2's pass rate at one sixth the cost, while API prices drop 20 percent to $8 input and $40 output ($0.40 cached). 3. GPT 6.2 Astra is OpenAI's first public model to hit the High threshold for AI Self Improvement, prompting OpenAI to build an automated kill switch system called Control Room. When frontier models double legal and scientific benchmark accuracy while slashing per task token costs by 70 to 95 percent, enterprise automation economics change overnight. openai.com/index/introducing…
Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…
61
MIT just published the full 2026 syllabus and open archive for its legendary How To Make Almost Anything hardware engineering course. 1. Run by MIT's Center for Bits and Atoms, the site gives anyone free access to weekly lab notes, CAD files, PCB design guides, and firmware tutorials. 2. The 2026 edition adds dedicated recitations on using AI agents and Edge AI alongside classic hands on weeks covering machining, molding, and mesh networks. 3. Even better, the archive links every single student build and project portfolio going all the way back to 1998 across 29 years of hardware hacking. Bookmark this site if you want a complete self study roadmap for turning raw silicon, sensors, and code into working physical machines. pub.cba.mit.edu/863.26/site/
22
A new Groundwork Collaborative audit of over 350 Amazon Black Friday and Cyber Monday deals found nearly 1 in 3 offered zero real savings. 1. Almost a third of the featured holiday deals sold for the exact same price or cheaper during the 30 days leading up to the sale event. 2. Nearly 1 in 10 advertised deals were actually priced higher during the sale than in the weeks prior, with a median markup of 19 percent above pre sale prices. 3. One pair of UGG boots listed on Cyber Monday sale for $179.98 had sold for $98.40 four weeks earlier, an 83 percent price hike disguised as a discount. Amazon quietly deleted the archived deal pages after researchers reached out, so always check a 90 day price history tracker before buying on Friday. groundworkcollaborative.org/…
33
A software bug in JPMorgan Chase's on premise hardware systems triggered cascading online banking and payment delays across major US banks today. 1. Customers at Chase, Bank of America, Wells Fargo, Truist, and US Bank all reported login errors and stalled transfers starting around noon ET. 2. Chase confirmed late this afternoon that an internal software bug slowed transactions whenever Chase was a counterparty, causing spillover delays on Zelle transfers at partner banks even though Early Warning Services' core network stayed online. 3. Coming just three days after Friday's global Microsoft cloud outage and the nationwide FAA ground stop, one bank's on premise bug rippled across the US payment rails in minutes. Even when core payment networks stay up, tightly coupled settlement between mega banks turns a single institution's software bug into an industry wide outage. apnews.com/article/bank-outa…
110
Alibaba just launched Qwen4 Max, a 2.4 trillion parameter multimodal model that cuts per task agent costs by 62 percent. 1. Built with Gated DeltaNet and Sparse Attention across a 1 million token context window, it activates 71B parameters per token and trims Terminal Bench 4.0 token usage by 49 percent. 2. Average Work360 task cost dropped from $4.43 down to $1.67 while beating Qwen3.8 Max across coding, legal due diligence, and 3D modeling. 3. In live automotive manufacturing tests, it cut CNC workpiece programming time from 475 minutes down to 24 minutes. Cutting token burn in half while scaling past two trillion parameters is how frontier models become cheap enough for 24/7 factory and office agents. qwen.ai/blog?id=qwen4-max
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! 🚀 - Blog: qwen.ai/blog?id=qwen4-max - API: qwencloud.com/models/qwen4-m… - Token Plan: qwencloud.com/pricing/token-… More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
2
79
A developer wired a $3 Arduino Nano into his 1994 Jeep Cherokee and used Claude and Codex to reprogram the engine computer and pass a California smog test. 1. Built before OBD2 existed, the 32 year old Jeep had no modern diagnostic port, so Claude reverse engineered Chrysler's legacy SCI protocol over three wires. 2. Dumping the ECU firmware revealed that the factory California fuel table starved the engine at 25 mph, causing high emissions during the exact smog test load. 3. Because the 1993 hardware could not be reflashed normally, Codex injected live code straight onto the car's EEPROM to fix the fuel mix. Frontier coding agents are no longer trapped inside web apps; give them a $3 microcontroller and they can patch 1990s automotive hardware. youtube.com/watch?v=WPj3aQ-x…
39
One of the most fascinating open source repositories on GitHub is the archived codebase for IRS Direct File, the free federal tax filing system shut down before the 2026 tax season. 1. Built by a joint team from the IRS, USDS, and 18F, the repo translates the US Internal Revenue Code into plain English interview questions that compile directly into Modernized eFile XML returns. 2. Buried inside the codebase is Fact Graph, a declarative XML knowledge graph written in Scala and transpiled to JavaScript via Scala.js that reasons over incomplete taxpayer data in both the browser and backend. 3. Although the repository is now archived with 4,700 stars and a warning not to run it in production, its tax logic test suites and Fact Graph rules engine remain public under a federal open source mandate. Before lobbyists and budget cuts killed it, government engineers quietly built and open sourced a clean declarative rules engine for the US tax code. github.com/IRS-Public/direct…
27
Two brothers built an interactive browser museum letting you compare all 13 retro releases of The Secret of Monkey Island pixel for pixel with authentic WebGL2 CRT monitor shaders. 1. Every edition across DOS VGA, EGA, CGA, Amiga, Atari ST, Mac, FM Towns, Sega CD, and monochrome Hercules was captured across 12 identical game moments at 1600x1200 with exact 4:3 aspect correction. 2. Instead of fake scanline overlays, they ported DOSBox Staging's real CRT shaders to WebGL2 so 31 kHz VGA renders 400 fine lines while 15 kHz Amiga and TV monitors render authentic phosphor gaps. 3. You can drag any two editions side by side, zoom in 4x to inspect pixel dithering, and tweak phosphor curvature and vignette settings live in the browser. It is one of the finest interactive digital preservation projects ever built for classic PC gaming history. monkeyisland.pages.dev/
1
2
52
Smart glasses, AI pins, and Apple Watch transcription tools are quietly breaking the oldest rule of personal privacy: your settings no longer protect you from the stranger sitting next to you. 1. In a sharp new warning, the Electronic Frontier Foundation points out that Meta and Google smart glasses, Amazon Bee pendants, and Apple Watch Audio Intelligence are engineered to record and summarize conversations continuously in public. 2. Tech companies brag about encryption and on device controls for the buyer, while offering zero technical opt out for friends, coworkers, or strangers caught in the microphone and camera range. 3. Even Google's senior product director admitted to CNN that pinging nearby phones when someone is recording leaves bystanders feeling weirdly disempowered because there is nothing you can practically do about it. Pulling out a camcorder or voice recorder in a restaurant used to carry an obvious social signal, whereas ambient AI wearables are designed to hide surveillance inside everyday objects. eff.org/deeplinks/2026/11/te…
72
An embedded hardware engineer priced out a Raspberry Pi 4 field rig at $347 and realized a $50 used Android phone already packs better hardware inside one pocket sized slab. 1. Building a Pi 4 setup with a 64MP camera, GPS, 64GB storage, real time clock, UPS battery backup, 5 inch touchscreen, and 4G cellular modem adds up to $347 before batteries or an enclosure. 2. A secondhand Redmi 13C 5G or Huawei Mate 30 5G sells for $45 to $75 and comes built in with a 50MP camera,5G modem, GPS, battery backup, touchscreen, and USB C host support to drive custom microcontrollers. 3. For remote kiosks and field sensors, built in cellular data lets you SSH in and push updates over 5G without begging a venue for WiFi access or open firewall ports. Smartphone economies of scale amortized billions in sensor, battery, and radio R&D into a $50 used device that makes modular dev board stacks look overpriced. so.long.thanks.fish/why-pi-w…
1
1
1
73
Security researchers audited 159,041 web apps built on AI coding platform Lovable and found thousands silently leaking user databases to anyone without a login. 1. Out of 34,545 Lovable apps wired to Supabase, 24,664 exposed database records on initial page load, and follow up read only API queries uncovered personal email records across 7,403 websites. 2. Exposed tables included over 1,000 passport and identity document records, 80,000 user profiles and 200,000 private chat messages from a single forum, 280,000 customer emails, contracts, invoices, and private photo albums. 3. While Lovable markets that zero coding knowledge is needed to launch an app, its terms place 100 percent of database access control liability on non technical users who have no idea how Supabase Row Level Security works. AI lets anyone ship a full stack web app in ten minutes, which means millions of production databases are now deployed by people who have never heard of access control rules. crissyfield.de/research/lova…
21
SpaceX just caught the Starship upper stage in mid air with the launch tower chopsticks for the first time in spaceflight history. 1. After launching from Starbase on Friday evening and deploying 26 Starlink V3 satellites across 14 orbits over 22.5 hours, Ship 42 survived orbital reentry and hovered directly into the steel catch arms at Pad 2 at 4:20 PM ET today. 2. While Super Heavy Booster 22 diverted to an offshore Gulf splashdown during ascent due to a low fuel alarm,catching an orbital upper stage returning from space at hypersonic velocity solves the hardest half of full rocket reusability. 3. Three previous flights proved SpaceX could catch the booster, and today's catch proves the entire two stage super heavy lift vehicle can return straight to the launch mount. Full and rapid orbital reusability just moved from a slide deck promise to an operational reality in South Texas.
Congratulations to an incredible team at SpaceX!!!
44
Engineers upgrading the Large Hadron Collider just built next generation superconducting magnets exceeding 11 Teslas using brittle niobium tin coils. 1. The existing Large Hadron Collider bends proton beams using niobium titanium magnets capped around 8.4 Teslas, which is already more than 100,000 times stronger than Earth's magnetic field. 2. Pushing past 11 Teslas required switching to niobium tin superconductor coils, a notoriously fragile alloy that shatters like ceramic if wound improperly under extreme cryogenic stress. 3. Stronger dipole and focusing magnets squeeze particle beams tighter at the collision points, dramatically multiplying luminosity and rare particle discoveries without digging a larger tunnel. Before building a bigger particle accelerator, materials science lets physicists extract radically higher collision density out of the exact same 27 kilometer ring. spectrum.ieee.org/large-hadr…
18