🤯 8B model > GPT-5 on math ? 🌠Introducing Parallel Coordinated Reasoning (PaCoRe)——open source deep think. 📚Decouple reasoning from context limits → multi-million-token TTC. 🤩With this new paradigm, even an 8B model can push the effective per-problem TTC context to multi-million tokens, surpassing proprietary systems like GPT-5 Thinking. 🔢 Math (HMMT25): 94.5%，beating GPT-5's 93.2%, powered by ~2M-token TTC 🤖 Coding (LiveCodeBench2408-2505) : 78.2% , competitive with frontier models。 Model checkpoints, training data, and inference pipline are now MIT-licensed and fully open-source. Fork it, test it, build with us！

+2 more

Revealing the Coder Color of the Year… https://t.co/C503ZccKuP

24/7/365. Rain or shine. CEO midnight meetings. All hands on deck building the world’s best data centers with maniacal speed and efficiency. https://t.co/W3TuzXFUKF

When Whites are a minority in a country they become targets and are systematically forced into a sub-class where they live in fear and lose everything (Rhodesia/South Africa) where the end goal is total White genocide. When non Whites are a minority they are granted millions upon millions in social programs and welfare and worshipped by liberal White people to the point they can do no wrong. There's a discernible difference in outcomes.

In S3E12 of @GoingnativeTv, Drew discovers how Winnipeg is embracing its Indigenous roots and revitalizing public spaces and institutions. He even gets his hands dirty. Visit https://t.co/9aYARUiZZ5 for more information. https://t.co/cAp112uISe

+5 more

You can now train LLMs 3× faster with no accuracy loss, via our new RoPE and MLP kernels. Our Triton kernels plus smart auto packing delivers ~3× faster training & 30% less VRAM vs optimized FA3 setups. Train Qwen3-4B 3x faster on just 3.9GB VRAM. Blog: https://t.co/kL6JM6skH1

Real-time audio transcription. Entirely on-device. This hands-on tutorial shows you how to build it from scratch. No cloud dependencies, no API calls, complete privacy, using LFM2-Audio-1.5B by @liquidai Enjoy ↓ https://t.co/7Kn7f6Alv3 https://t.co/ax3MlZ8p0X

MAX just demonstrated something important: you can break AI vendor lock-in without compromising performance. New case study with @tensorwave shows what we achieved on AMD MI355 GPUs in production: 🧵 https://t.co/PDG1PZtBaf

These gains are not theoretical. They come from real usage across multiple workloads, backed by benchmarks that we published together with @tensorwave Watch the video: https://t.co/B3eYyLzNPP

By combining MAX software optimizations + AMD MI355's higher GPU memory + @TensorWave's liquid-cooled infrastructure we achieved ~2× greater inference throughput on MI355 clusters and 40-60% reduction in inference costs. Full methodology and benchmarks: https://t.co/XegRDVrvvm

@corbin_braun You should try https://t.co/wCUGHijxku

TOMORROW: Inside Helion: Live Q&A with Jason Ansel, @oguz_ulgen, @weifengpy, and Jongsok Choi from @Meta’s PyTorch Compiler and Helion teams. Hear how the Python-embedded DSL compiles to Triton, accelerates kernels, and what’s ahead for Helion. 🔗 https://t.co/aVnYgsdwis #PyTorch #AIInfrastructure #OpenSourceAI

Finally, AI designs you can edit. Moda - create brand assets on a fully editable canvas. This is not another ChatGPT or nano banana wrapper. We actually taught an agent layout, typography, and color. We gave it taste. Use it for slides, social, infographics, e-books … https://t.co/0AgNYpHvfb

Introducing WindowSeat - our new method for removing reflections from photos taken through windows, on planes, in malls, offices, and other glass-filled environments. Try it with your own photos in this demo: https://t.co/PMOvQP79cK

I just released an open source AI Agentic RAG system that can auto-schedule Google Calendar events and send Gmail replies when the policies allow, inspired by customers of @auto_manus. The agent checks your vector database, reasons over the policies, and decides when to call the right tools based on user requests. In this tutorial I'll walk through: • a simple system design for an AI Agentic RAG • a working demo that auto-schedules Google Calendar and sends Gmail with policy checks • a full-stack code walkthrough and deployment using Supabase, FastAPI, and Google Cloud Today's Github: https://t.co/M7nuINGYOk ☕️ Buy me a coffee: https://t.co/QHLDvxrcJU You can build anything. You can learn anything. 💪

Today we’re launching Vybe to the world and announcing our $10M Seed round to make vibe-coding actually work inside companies. This is why, how and our vision: Over the last few decades, every fast-growing company has quietly built the same mess behind the scenes: internal ops glued together with rigid SaaS, fragile spreadsheets or custom-coded tools nobody wants to maintain. Meanwhile, eng teams are stretched thin. Internal tools never make it to the top of the backlog. Vibe-coding is changing the game but it’s mostly been good for prototypes, landing pages, and side projects disconnected to production data. Our belief is simple: in the next few years, most internal software will be vibe-coded by teams working with AI, engineers and business teams together. Vybe is built for that collaboration: 1/ Business teams own the surface area:  Business teams (Ops, CX, PMs etc.)  can build and iterate on apps themselves: flows, UI, fields, and logic; without waiting weeks for eng to pick up another “internal tools” ticket. 2/ Engineers own the foundation:  Integrate production data (Postgres, Salesforce, Jira, and 3,000 integrations), define SQL definitions once, set up SSO auth, access control, and keep everything in Git to help when needed (from their favorite IDE!) 3/ Secure by design:  Our security and permissioning layer is not vibe-coded and can’t be modified by AI. Everyone can sleep at night. 4/ Team-ready out of the box:  SSO, Auth, environments, deployments, and review flows are built in. Over the last few months, we’ve been in closed waitlist mode and have hand-onboarded teams to pressure-test Vybe on real production workflows: - A YC Founder runs his entire CS operation on Vybe and saves ~2 days per week. - Another company ingested millions of rows from their warehouse to build BI-like internal views that would break typical AI builders. - One team fully replaced Metabase/Looker by plugging Redshift into Vybe and just… prompting their way to MAU, DAU, funnels… Remix apps from world-class operators To make it even easier to get started, we’re launching templates co-created with operators who’ve already solved these problems at scale: - @collinmathilde (CEO @ Front) – how she runs 1:1s - @lennysan (yeah, that Lenny!) - how to manage up, do perf reviews and write PRDs - @sushmars (CTO @ 23andMe) - her  7Cs Framework for Build vs. Buy Decisions - and many more from the best Tech leaders Backed by people who’ve lived this pain We’ve raised $10M in Seed funding, led by @firstround with participation from @ycombinator and an incredible group of operators and founders, including: The CEO Datadog, CEO Grammarly, CEO Reforge, CTO Intercom, Head of Product at OpenAI, Head of Product Anthropic, and 50 more incredible operators who believed in our vision! Huge thank you to our early customers, team, and investors for believing in us this early. 🙏 We’re now in GA: no more waitlist! https://t.co/jqEYOWYGSQ

This is one of the coolest VR experiences I’ve seen in a while, it lets you play mini golf from the comfort of your own living room. https://t.co/19Y5oKDTR6

https://t.co/rAThImaYKo

https://t.co/rAThImaYKo

Perfection. https://t.co/uZFg3DhOBn

Perfection. https://t.co/uZFg3DhOBn

Who we let into our country matters. A 15 year old girl was raped by two Afghan asylum seekers in Warwickshire England. The lawyer blamed the attack on “cultural differences” and said his client “was not used to a society where women are free and deemed equal to men”. The United States should be unapologetically strict with who we let into our country.

AI-native doesn’t mean “now with AI.” It means systems built to interpret, respond, and evolve without relying on one-size-fits-all models. 🤖⚙️ It’s a shift in how we think about content infrastructure and semantic frameworks that can collaborate with intelligence. And it’s already changing how organizations structure workflows, personalize experiences, and maintain control over data. Read the full post for a closer look at what AI-native really means, and why openness matters more than ever. 🔗 https://t.co/wmiXNbEbDA

The agent experience in @code just got a major upgrade. With today's release, you'll see: ✨ Agent sessions integrated into the chat view for a unified experience ✨ Support for isolated background agents powered by Git worktrees ✨ Seamless delegation across agents 🧵 https://t.co/KTkcvFxGRG

Agent sessions are now integrated directly into the Chat view. See session status, progress, and file changes at a glance. Archive sessions to stay organized. Select any session to view the full conversation history. https://t.co/Exchy0KN3c

Background agents work in their own isolated workspaces, so you can run multiple background agents without conflicts. Git worktree support isolates each agent's changes in a separate folder. Review and merge changes back when done - or apply directly to your workspace. https://t.co/i79uYycIlM

Hand off tasks seamlessly across agents. Start brainstorming locally, then continue with a background or cloud agent using "Continue in". Your chat context transfers automatically and the original session archives after handoff. https://t.co/4EUXsqEXZy

Seamless agent collaboration in VS Code, tailored to the way you work best. Read the full release notes: https://t.co/A9p63tiyY2 Happy coding! https://t.co/gavGtTcM26

+1 more

🚨X SUSPENDS PRO-RUSSIAN HACKTIVIST ACCOUNT, COOPERATES WITH FBI ON CYBERCRIME INVESTIGATION X confirmed the suspension of the account for what it called “clear evidence of criminal conduct.” In its statement, the platform said it is working directly with the FBI’s Los Angeles field office in an ongoing investigation into DDoS attacks and other hacking activity targeting U.S. infrastructure. The banned account is well known in cybersecurity circles for ties to the Russian-linked “Killnet” collective, a hacktivist network accused of launching DDoS-for-hire operations against government and financial systems. The FBI said the account had been used to coordinate and promote attacks on U.S. and allied networks. This marks one of X’s most public collaborations with law enforcement since Elon’s acquisition in 2022 and reflects a shift toward active enforcement under its newer zero-tolerance policy for criminal activity. Source: @Safety

Funny how people proudly share their IQ test results only when the score looks impressive. When it doesn’t… suddenly it “never happened.” 😂 Meanwhile, the testing company still has all the data in their database. 🤣🎉 #IQTest #DataPrivacy #DarkHumor https://t.co/rFxj16CeAm

🔑 Your house key is basically API data now. We still think of “physical” keys as safe because they’re not online. But in 2025, that line between cyber and physical security is paper-thin. With cheap, off-the-shelf tools you can now: 1️⃣ Capture the profile of a key with a small hacking gadget (Flipper-style devices) 2️⃣ Decode the bitting pattern with simple software 3️⃣ Print a working copy on a basic 3D printer in minutes At that point your key isn’t a piece of metal anymore — it’s data that can be: 📥 read 💾 stored 📤 shared 🖨️ reproduced on demand And it’s not just keys: 🔐 badges & RFID cards 🔐 intercoms & pin pads 🔐 mobile access apps All of them are merging into one attack surface where “cyber” and “physical” are the same problem with different UI. ⚠️ Important note (because ethics > clout): This kind of research should ONLY be done: – in legal, controlled environments – on systems/locations you own or are explicitly allowed to test – with the goal of hardening security, not breaking it The point isn’t “haha look what we can break”, it’s: our access-control model is stuck in the 90s while the tools are in 2025. 💭 Honest question: Are we massively underestimating physical access control in modern security architecture? #CyberSecurity #PhysicalSecurity #RedTeam #FlipperZero https://t.co/FFWblGvx5h

Just found the wildest bug in Traefik (3B+ downloads): TLS verification "on" actually turned it OFF 🔥 CVE-2025-66491 The vuln was 3 chars. NGINX "on" collided with Go semantics => inverted security intent 🙃 It evaded scanners & passed all tests until @Aisle_Inc's discovery! https://t.co/TzrwhrD0LE