← Back to leaderboard
thisguyknowsai

Brady Long

Verifiedactive
@thisguyknowsai

Guy that likes AI. Far from an expert. Pretty below average but also pretty chill. Have been called “Brady Short” before.

RosterAI Agents & Productivity ToolsHumor & Meme AccountsAI Model News & Releases
44,780 followers · 1,796 following · Last polled 3h ago
Perf. Score
50.1
Rank #234
Value Score
—
set rates to compute
Median views
1.6K
p25 1.4K · trailing 7d
Followers
45K
-0.0% 7d
ER (impr.)
3.23%
eng ÷ impressions
ER (foll.)
0.35%
eng ÷ followers
Est. CPM
—
needs a rate
WoW views
+280.6%
▲ rising

Rates & value · est. from 1.6K median organic views

Quote tweet
—
no CPM
Post
—
no CPM
Thread
—
no CPM
Retweet
$75
no CPM
rate 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 (1.6K), not the mean (4.9K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (1.4K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
This free plugin makes Claude Code last 10x longer per session A Playwright snapshot eats 56 KB. 20 GitHub issues eat 59 KB. One access log, 45 KB 30 minutes in and 40% of your context is GONE Then the agent compacts and forgets which files it was editing, what tasks were open, and what you last asked for Context Mode kills this problem completely. 31K stars on GitHub, used by teams at Google, Meta, Microsoft, NVIDIA and Stripe What it does: • Runs tool output in a sandbox. Only the result enters context. 315 KB becomes 5.4 KB. A 98% cut • Saves every file edit, task, error and decision to a local SQLite database. After compaction your agent picks up exactly where it left off • Makes the agent write one script instead of reading 47 files. 700 KB becomes 3.6 KB • Works on 17 platforms. Claude Code, Codex, Cursor, Gemini CLI, Copilot, OpenCode and more Setup takes 30 seconds: /plugin marketplace add mksglu/context-mode /plugin install context-mode@context-mode Restart Claude Code and you're done Longer sessions. Sharper agent. A fraction of the tokens If you code with agents every day, install this before anything else https://t.co/gM5RBXzj2S2d ago32K477363608144.32%
This is terrifying. Jensen Huang's NVIDIA engineers released their internal AI prompting playbook. No paywalls. No waitlists. No gatekeeping. Your agents are hallucinating 35% more than they should and this stops it cold. Here's the exact system they built and the 6 prompts that changed our results overnight:
11h ago
9.5K
68
19
4
1
157
2.61%
Honestly mind blowing 🤯 The model loads into the browser, says hi in English, then flips to French and German without missing a beat. https://t.co/EkGMCMtxLo2d ago6.3K401040.14%
🚨 BREAKING: AI models are now on call at OpenAI. @thsottiaux says they can change production systems and are helping employees sleep better.frozen21d ago5.0K610040.22%
You don't need to open Premiere Pro anymore. Claude can now cut, grade, caption, and render your videos. Here are 10 open-source GitHub repos that make it happen: 1. video-use Drop raw footage in a folder, tell Claude Code "edit these into a launch video," and get final .mp4 back. It cuts filler words and dead air, color grades every segment, and adds audio fades at every cut. The clever part is that Claude never watches the video. It reads about 12KB of text and a handful of frames. https://t.co/XS7gqxGcEE 2. DaVinci Resolve MCP Gives Claude the full DaVinci Resolve scripting API. Timeline assembly, color grading, media management, and rendering, all in plain English. https://t.co/Gas4gT82bH 3. Premiere Pro MCP Hand Claude a script. It scans your media, builds an edit decision list, and assembles the Premiere timeline for you. https://t.co/E3KGrNEoHH 4. Remotion Skills The official skill that teaches Claude to write correct Remotion code. Every video becomes a React component, so Claude animates frame by frame and renders to MP4. https://t.co/UhoRPugSTO 5. HyperFrames HeyGen's open-source take on the same idea, but in plain HTML instead of React. No build step. Point it at a PDF, a repo, or a website and get a video back. https://t.co/u9SrpVJ0t3 6. Claude Code Video Toolkit A full production studio. Describe the video you want, and Claude writes the script, generates the voiceover, music, and visuals, then renders the MP4. https://t.co/wOEwhVrI8o 7. YouTube Clipper Paste a YouTube link. Claude downloads it, splits it into chapters by meaning instead of timestamps, clips the best parts, and burns in bilingual subtitles. https://t.co/HON7Bl6JVz 8. Video Edit MCP Trim, merge, crop, resize, and rotate, plus face tracking for automatic edits. Works with Claude Desktop out of the box. https://t.co/nhCUtiRty5 9. Video Editor for Claude Code Built for one annoying problem. Screen recordings full of dead pauses while Claude thinks. It strips them out so your demo plays tight. https://t.co/qd3hp2zIzQ 10. mcp-video FFmpeg editing with guardrails. Subtitles, audio, effects, and a preflight check that catches broken output before it silently ships. https://t.co/OhsZw15p13 Editing used to be the slowest step between recording and posting. Now it's a prompt.frozen4d ago3.7K6417191814.92%
10 GITHUB REPOS THAT QUIETLY DESTROY BILLION-DOLLAR PRODUCTS. Bookmark all of them. Each one replaces something a company charges money for. 1. PhotoGIMP Patches GIMP to look and feel exactly like Adobe Photoshop. Same shortcuts. Same layout. Replaces a $600/year subscription. https://t.co/vboWNfyCMX 2. OpenToonz The actual animation software Studio Ghibli used for Spirited Away and Princess Mononoke. Open sourced in 2016. Replaces Toon Boom Harmony ($1,575/year). https://t.co/dLAH8fm8HI 3. uBlock Origin The world's best ad blocker. Built by one guy, Raymond Hill, who refuses every dollar people try to send him. GPL-3.0. https://t.co/na74Wjf1bT 4. LocalSend AirDrop, but cross-platform and offline. Sends files between iPhone, Android, Windows, Mac, and Linux. No cloud. No account. https://t.co/xOobsp0DIv 5. Stirling-PDF Self-hosted PDF toolkit. Merge, split, OCR, compress, sign. Replaces Adobe Acrobat Pro ($240/year). Runs in Docker. https://t.co/BEzXknpyyL 6. Immich Self-hosted Google Photos clone. Face recognition, albums, mobile apps for iOS and Android. Your photos never leave your server. https://t.co/vtJi1ZpD00 7. Ollama Run Llama, Mistral, Qwen, and DeepSeek locally with one command. No API keys. No cloud. No data leaving your machine. https://t.co/YWqB0Wk8mO 8. AppFlowy Open source Notion replacement. Built in Rust. Works offline. Your notes stay on your computer. https://t.co/TOxxyaRVfP 9. ente Encrypted Google Photos and Authenticator alternative. End-to-end encrypted. Open source on both client and server. https://t.co/4g79f3Vjbb 10. ArchiveBox A personal Wayback Machine. Saves any URL as HTML, PDF, screenshot, and video. Your own private internet archive. https://t.co/bM8DjDjYGN The only thing keeping the subscriptions alive is that most people don't know these exist.frozen19d ago2.8K321221433.17%
In March 1989 a 33-year-old software engineer at CERN handed his boss a proposal for managing information. His boss, Mike Sendall, described it in three words. "Vague, but exciting." He was allowed to keep working on it. By the end of 1990 he had written the first web browser, the first web server, and the first website. He had also invented HTML, HTTP, and the URL. Then he gave all of it away. No patent. No royalties. His name is Tim Berners-Lee. The proposal became the World Wide Web. Here is the story, because the man who invented the thing you are reading this on never charged anyone a cent for it. Tim was born in London on June 8, 1955. Both of his parents were mathematicians who worked on the Ferranti Mark 1, the first commercially built computer. As a child he was a trainspotter who learned electronics by tinkering with a model railway. He studied physics at The Queen's College, Oxford and graduated with a first in 1976. While he was there, he built a computer out of an old television set he bought from a repair shop. In 1980 he spent six months at CERN as a contractor. To keep track of how people and projects connected, he built a hypertext prototype called ENQUIRE. Then he left. He came back in 1984 as a fellow. By 1989 CERN was the largest internet node in Europe. He saw that hypertext and the internet could be joined. He wrote the proposal on March 12, 1989. He built the browser, which was also an editor, on a NeXT computer and called it WorldWideWeb. The first website went live on December 20, 1990. It explained what the web was and how to set up your own server. On August 6, 1991 he posted a public invitation on Usenet for anyone to join the project. In 1994 he founded the World Wide Web Consortium at MIT to keep the standards open and royalty-free. He was knighted in 2004 and won the first Millennium Technology Prize the same year. He was appointed to the Order of Merit in 2007 and won the Turing Award in 2016. At the 2012 London Olympics he appeared in the opening ceremony, sitting at a vintage NeXT computer. In June 2021 an NFT of the web's original source code sold at Sotheby's for $5,434,500. In 2018 he launched a startup, Inrupt, to give people control of their own data again. His boss called the idea vague. It became the World Wide Web. He never took a royalty on a single page.frozen20d ago2.6K24156151.93%
Do not tell ChatGPT “write my resume.” Do not tell it “make it sound professional.” Do not tell it “add achievements.” You just asked AI to sell you without knowing your real value. Here are 8 prompts to copy and paste that actually land interviews:frozen29d ago2.5K14920141.58%
SHOCKING: One government IT officer in Nepal built what an entire disaster agency couldn't Nepal is dealing with one of its worst disasters in memory. A glacial collapse on the Nepal-China border triggered floods that wiped out roads, bridges, and hydropower plants Over 1,450 people dead. Thousands still missing The biggest problem for families wasn't a lack of information. It was that the information was scattered across a dozen different agencies, pages, and feeds So Niraj Bhushal, an assistant IT officer at Nepal's Ministry of Finance, built a fix himself. On his own. Layout to JavaScript The Rasuwa-Bhote Koshi Flood Rescue Portal: - Pulls data every 5 minutes from the disaster authority, police, army, and district office - Takes missing person reports from Google Forms - Uses AI to cross-check them against lists of rescued survivors and remove duplicates - Maps damaged buildings using satellite imagery - Open source. Hosted on GitHub He wasn't the only one. Akshay Bajpai, an Indian-origin engineer, prototyped an Android app called Mission Rescue while searching for his own family. It watches official rescue and missing person updates and alerts you when something relevant shows up Two individuals. Built in days. Filling gaps governments couldn't But here's the part that hit me hardest While AI portals were cross-referencing data on screens, hydro workers were trapped in dark tunnels for 4 straight days. Rescuers searched using smartphone flashlights. Officials say the entire country has only 4 or 5 drones that can lift serious payloads Software solved the information problem in days Hardware is still years behind This is what AI does best right now. It takes one person with a laptop and lets them do the coordination work of an entire department The next step is getting that same energy into the physical world. Drones, robotics, early warning systems. The things that actually pull people out of the mud If you're a builder, this is the reminder: the most important things you can build aren't SaaS dashboards. They're tools that help real people on their worst dayfrozen25d ago2.2K21101021.54%
If you've ever told ChatGPT something you'd never tell another person, read this. On September 14, 404 Media revealed Project Lily. OpenAI has been paying hundreds of outside contractors, more than $50 an hour, to read real ChatGPT conversations and rate the responses. Not single messages. Whole conversations. Some assignments come with a "user memories summary": what the person has used ChatGPT for, and in some cases where they might live. OpenAI runs chats through a privacy filter first. But it admitted to 404 Media that the filter can miss personal details, especially in shorter chats. In many of the conversations, users had explicitly asked ChatGPT to keep what they said secret. What you should know: → Model training is on by default for Free, Plus, and Pro accounts → Turning it off only applies to new chats, not ones already eligible → When 404 Media asked where users are told humans might read their chats, OpenAI didn't directly answer Two days later, a class action was filed in San Francisco. And it's not only OpenAI. Anthropic confirmed to 404 Media it also uses human review. Human review isn't evil. It's how these models get better. The problem is that 900 million people think they're talking to a machine, and the settings page doesn't tell them otherwise. If you had known a person might read it, would you have typed it?frozen27d ago2.2K14111011.25%
GitHub is hiding an absolutely INSANE upgrade for Claude Code and Codex. It's called MemPalace. It gives your AI coding agents searchable long-term memory across sessions. MemPalace fixes that. And it hit 96.6% recall on LongMemEval with zero API calls. No cloud. No LLM. No API key. Just local semantic search over your actual conversations. The trick is in how it stores memory. Most memory tools summarize your chats, which means they decide what's worth keeping and throw the rest away. MemPalace keeps everything verbatim and organizes it like an actual memory palace: → People and projects become wings → Topics become rooms → Original conversations live in drawers So searches are scoped instead of running against one giant pile of text. The numbers: → 96.6% R@5 raw on LongMemEval, no LLM at any stage → 98.4% on a held-out set it was never tuned on → 47 MCP tools for reads, writes, knowledge graph and agent coordination → Auto-save hooks for Claude Code, Codex CLI and Cursor → Works with Gemini CLI, Antigravity and local models Search "why did we switch to GraphQL" and it pulls the exact conversation from 3 months ago. The part most Claude Code users don't know: your sessions expire after 30 days unless you have save hooks wired. MemPalace backs them up before they disappear. Nothing leaves your machine unless you opt in. 59.6K stars. MIT licensed. 100% open source.frozen5d ago2.1K3013331154.36%
R.I.P SEO agencies. You can now use Claude to audit your site, find keyword gaps, reverse-engineer competitors, and build a 90-day content plan for free. Here's the mega prompt you can steal ↓ https://t.co/46XEZH9Z7qfrozen27d ago2.1K15930152.00%
THIS IS CRAZY... SOMEONE JUST KILLED EVERY SKETCHY VIDEO DOWNLOADER SITE ON THE INTERNET. It's called ReClip. Self-hosted. Open source. Free. Paste a link from YouTube, TikTok, Instagram, X, or 1000+ other sites. Pick MP4 or MP3. Done. No ads. No popups. No trackers. No rate limits. Here's how it works: 👇frozen13d ago1.9K14921222.58%
A mathematician at Bell Labs spent 30 years watching brilliant scientists fail and asked the question nobody else would. Why do equally smart, equally hardworking people end up in completely different places? Some won Nobel Prizes. Some changed entire fields. And some, with the same degrees, the same budgets, and the same offices down the same hallway, did solid work that nobody remembered a few years after they retired. His name was Richard Hamming. He joined Bell Labs in 1946 and stayed for three decades. He shared an office with Claude Shannon. Walter Brattain, who won the Nobel Prize for co-inventing the transistor, worked in the same building. So did John Tukey and a long list of people whose names are now in textbooks. Hamming did his own historic work along the way. He invented error-correcting codes, the reason your files do not corrupt every time a bit flips. In 1968 he won the Turing Award. But the thing he became famous for was not a code. It was a talk. On March 7, 1986, he stood in front of about 200 researchers at Bellcore and told them what he had learned from watching great scientists and ordinary ones side by side. It came down to one habit. Every Friday afternoon, Hamming blocked off time and did no regular work. No experiments. No meetings. No papers. He called it Great Thoughts Time. He sat and asked himself what the important problems in his field were, and where the field was going. Then came the uncomfortable part. Was he working on any of them? Most people never ask this. Not because they are lazy. Because the honest answer is painful. If you admit you are not working on something important, you have two choices. Change direction, or keep pretending. Pretending is cheaper this week. It is ruinously expensive over a career. Hamming took the question to lunch. At Bell Labs he started sitting with the chemists. He asked them what the important problems in chemistry were. A while later he asked what important problems they were working on. Then he asked the one that ended the friendships. If what you are doing is not important, why are you at Bell Labs doing it? After that, he said, he was not welcome at the table. He had to find somebody else to eat with. Nobody had a good answer. And sitting next to someone who kept asking turned every meal into a performance review. The line most people skim past in the transcript is the whole talk in one sentence: "If you do not work on an important problem, it's unlikely you'll do important work." It sounds obvious. It is not motivational. It is arithmetic. The problem you choose sets the ceiling on the result you can get. Brilliant effort on a small problem produces a small result. Every time. The transcript of that talk, "You and Your Research," has been free online for nearly 40 years. Stripe Press republished his full lecture series as a book, The Art of Doing Science and Engineering. Almost anyone can read it in an hour. Almost nobody blocks off this Friday afternoon to ask the question.frozen21d ago1.9K22113192.49%
SHOCKING: Claude has a secret mode called "Research Accelerator." It turns a 60-page paper into a literature review, a counterargument, and 10 research questions in under 5 minutes. Here are the 13 prompts grad students are quietly using to finish in half the time: https://t.co/adJ6qeLNLwfrozen22d ago1.8K191021192.76%
enterprise software is an incredible business you sell someone a tool and then they hire another company to explain how to use it ERP has chosen a very expensive group project to interrupt. https://t.co/a4wjZg4Nkofrozen7d ago1.7K2583022.21%
In 2019, Google said its quantum computer solved in 200 seconds a problem that would take the world's best supercomputer 10,000 years. Within a week, IBM said that was wrong. The chip was called Sycamore. It had 54 qubits, though one didn't work, so 53 were used. Qubits are the building blocks of a quantum computer. A normal bit is either a 0 or a 1. A qubit can be in a blend of both, and with the right setup that lets a machine explore an enormous number of possibilities at once. For decades, scientists had argued about whether a quantum computer could ever do something no regular computer could. They gave the milestone a name. Quantum supremacy. Google's team, led on the hardware side by physicist John Martinis, thought they had reached it. The task they picked was a strange one. Take a random circuit, run it over and over, and sample the output. The answer was meant to be impossible for a classical machine to reproduce in a reasonable time. Before the paper was even official, a draft appeared on a NASA website, and the news leaked. Google followed with a Nature paper in October 2019. The headline claim was that their chip took 200 seconds, and that Summit, then the most powerful supercomputer on Earth, would need 10,000 years. The press ran with it. Quantum supremacy had arrived. Here is the part that should actually stop you. IBM, which builds both quantum computers and supercomputers, disagreed almost immediately. They said Google had assumed the classical computer would be stuck using a slow method. With a smarter approach that used the machine's disk storage, IBM argued, Summit could do the job in about two and a half days. Two and a half days instead of 10,000 years. It was still slower than Sycamore. But the dramatic gap shrank from a cosmic number to a long weekend. Then it got worse. In the next few years, researchers using ordinary graphics processors came up with even better methods, and some claimed to reproduce the result in hours. And there was a second problem nobody could argue with. The task was useless. It was designed so that a quantum computer could do it, and for no other reason. Nothing in the real world needed the answer. Google kept going. In late 2024 it announced a newer chip called Willow, with 105 qubits and a claim of finishing a test in under five minutes that would take a supercomputer far longer than the age of the universe. Here is where it actually stands. The progress is real, and it matters. The Willow results on error correction, getting qubits to make fewer mistakes as you add more of them, were a genuine step forward. But a machine that does something useful that no classical computer can do, in a way that holds up to scrutiny, has not been shown yet. The headline number made the machine sound unstoppable. The footnote said it was beaten by a different method within days.frozen17d ago1.7K17103031.93%
GROK IS THE BEST INVESTMENT ASSISTANT I gave it a stock I was ready to buy and asked it to make the case against me. Kinda wild how much conviction disappears when you ask the right question lol. Here are 10 Grok prompts I’d use before putting money into any company:frozen29d ago1.7K181240152.95%
Someone open-sourced computer use that never sends a screenshot to a big model. It's called typesafe-computer-use. Every other computer-use agent ships a full screenshot to a frontier model and waits 5 seconds for a plan, every single step. This one doesn't. It reads your screen deterministically with OCR, hands a tiny classifier a short list of actions, and asks one question: what comes next? Only when a text field genuinely needs free text does it wake a writing model. Most steps don't need a plan. They need one choice from a short list, made fast and cheap, with a confidence number you can gate on. The numbers are absurd: → $0.0002 per step vs $0.032 for Claude Opus. 155x cheaper per decision. → $0.003 per 12-step task vs $0.40-$0.90. Up to 300x cheaper. → 0.13-0.38s model latency vs 5.2s. Up to 40x faster. → Every click ships with a calibrated confidence score. Under threshold, it stops. → Runs on your Mac. MIT licensed. The honest catch: every bit of reasoning the big model does for free has to be rebuilt here as deterministic state. Date math, field checks, stop rules. All hand-built. That's the trade. You give up the magic black box and get something 300x cheaper that tells you exactly why it clicked. 100% Open Source. https://t.co/x9D5OYBZZkfrozen12d ago1.7K14103021.75%
Solo builders are gonna be the biggest winners in the AI era. A developer built CloneCut, an AI skill on the Capafy marketplace, and generated almost $300K in sales in 30 days. You don't need a big team to build a real business anymore.1d ago1.6K21141042.48%
Audible is in trouble Someone just open-sourced a tool that turns any ebook into a full audiobook. For free It's called ebook2audiobook and it has 20,400 stars on GitHub Drop in an EPUB, PDF, or MOBI. Get back a finished audiobook with chapters and metadata Here's what's insane about it: - Works in 1,158 languages - Clone your own voice from a 1 to 5 minute recording. Background noise gets cleaned up automatically - 9 TTS engines built in including XTTSv2, Bark, Tortoise, and Piper - Reads scanned PDFs and images with built-in OCR - Translates the book into another language before narrating it - Exports straight to m4b so it works in any audiobook app - Runs on 2GB of RAM. Your old laptop can do this Here's how I'd use it this weekend: 1. Clone the repo and run the launcher script 2. Open the web app at localhost:7860 3. Record 2 minutes of yourself reading anything 4. Upload a non-DRM book you own and pick your voice 5. Let it cook overnight Wake up to a book narrated in your own voice GPU gets you near real-time conversion. CPU works too, just slower Every paid service you use today has an open source version coming for it Run it local. Own your library https://t.co/FFr2ReJcWPfrozen6d ago1.6K201120163.04%
Harness engineering finally has a proper course. (100% open-source, 19k+ stars) The strongest model in the world still fails on real engineering tasks if you don't build a proper environment around it. You've seen this yourself. You give Claude Code a task in your repo. It starts well, reads files, writes code. Then it skips a step, breaks a test, and says "done" when nothing works. This isn't a model problem. It's a harness problem. Anthropic proved it with a controlled experiment. Same model (Opus 4.5), same prompt. Without a harness: $9, 20 minutes, something that didn't work. With a harness: $200, 6 hours, a game you could actually play. The model didn't change. The harness did. Learn Harness Engineering is a project-based course that teaches you how to build that environment. It breaks a harness into five subsystems: → Instructions: AGENTS md / CLAUDE md as a map the agent navigates on demand, not one giant file. → State: progress log + feature list persisted to disk, so the next session picks up exactly where the last one left off. → Verification: only a passing test suite counts as evidence. No runnable proof, no "done." → Scope: one feature at a time. No overreach, no rewriting the feature list to hide unfinished work. → Session lifecycle: init sh at start, clean-state checklist at end, handoff note for the next session. The model decides what code to write. The harness governs when, where, and how. What's inside: - 14 lectures, each answering one question (why do agents declare victory too early, why does one giant instruction file fail) - 8 hands-on projects on the same Electron app, where each project's solution becomes the next one's starter - Lecture 13 on loop engineering, Lecture 14 on graph engineering - Breakdowns of how Claude Code, Codex, Pi, and DeepSeek build their own harnesses - A harness-creator skill that scaffolds AGENTS md, feature lists, init sh, and verification workflows in minutes - A shell script that audits any existing repo against all five subsystems MIT licensed. Works with Claude Code, Codex, or any CLI agent that can edit files and run commands. GitHub repo: https://t.co/QzacTBTUTw The harness doesn't make the model smarter. It makes the model's output reliable.frozen7d ago1.6K2210110123.48%
This founder is genius. Someone gave Claude one prompt and now it runs his entire launch. → Positioning → Landing page → Cold email → 7-day content Here's the exact prompt (steal it): ↓frozen14d ago1.6K151041112.64%
AI video is becoming something you can experience, not just watch. With @PixVerse World Model, you can step into a generated environment, move around, interact with the scene, and influence how the world responds in real time. It feels like a new way to experience AI-generated worlds. #PixVerseWorldModelfrozen29d ago1.5K331000.46%
Phones aren’t the future of computing anymore. They’re just the thing most people still carry out of habit.frozen5d ago1.5K713000.74%
If you're a student in 2026, BOOKMARK THIS IMMEDIATELY. This website collects student-only freebies and discounts from some of the world's biggest tech companies. Not free trials. Not student "discounts." Actually free. The catch is nobody tells you. Companies bury these behind verification pages because they know most students will just pay full price instead. Here's the stack sitting there waiting: Google Gemini: free for a full year Microsoft 365: free GitHub Student Pack: free dev tools and credits Notion: free Canva Pro: free Figma: free 1Password: free for a year Autodesk, Unity, Unreal: free Amazon Prime + Grubhub+: free A working builder pays over $1,000 a year for most of this. A student pays nothing and barely anyone claims it. Four years of school is four years of free access to the exact tools people go into debt learning later. If you're a student paying for any of these right now, you're not broke. You're just unverified. Every deal, every country, one page: https://t.co/GOazwl0yU01d ago1.5K16103082.55%
I'm just a guy that likes AI, but this explained something I've felt for years. ChatGPT writes like it's trying to get you to the answer. Good stories want you lost for a while. Pocket FM built Sherpa for exactly that.frozen22d ago1.4K501030.62%
Finally @LeoDiCaprio can be good at actingfrozen15d ago1.4K201000.21%
The beginning of the 4PM NFL games is like staying at a bar knowing you should go home. But not wanting to go back to your lifefrozen4d ago1.4K000000.00%
You can now train your own LLM from scratch for free. Someone open-sourced a step-by-step project that takes you from raw data to a model that writes English. Every line of the Transformer is coded by hand in PyTorch, straight from "Attention Is All You Need." → Download and prep the data (The Pile, ~825GB of text) → Build multi-head attention, MLP blocks, layer norm, causal masking → Train the model → Generate text with it → 13M params on a free Google Colab GPU → Up to ~2B params on a single A100 → Correct English in under a day on the free tier Every part of the code comes with an explanation. 100% Open Source.frozen23d ago1.4K12900102.17%
Screen sharing in Teams and Zoom is painful for developers. Code looks blurry. The stream lags by seconds. Your coworkers can't read what's on screen. Screego fixes this. It's an open-source, self-hosted screen sharing server built specifically for devs. Runs as a single Go binary or Docker container. 10.5k stars on GitHub. What makes it different: → WebRTC under the hood (peer-to-peer, encrypted) → Low latency, high resolution. Readable code, no artifacts. → Multi-user screenshare in a single room → Integrated TURN server for NAT traversal, no external STUN/TURN config → GPL-3.0, 100% self-hosted. Your streams never touch a third-party SaaS. No account signups. No Zoom links. No corporate telemetry between you and your coworker. Install is one binary or docker run. There's a public demo instance if you want to test before deploying. If you've ever tried to pair program over Teams and given up, this is the fix. https://t.co/3frHdJRiQTfrozen11d ago1.4K1792032.18%
Your company runs on rules nobody wrote down. Someone quits and that knowledge dies. Every other AI gateway just routes tokens. This one learns how your business actually thinksfrozen23d ago1.4K001010.14%
BIG NEWS: Chinese researchers published a paper that destroys every AI agent startup pitch deck. It's called ROME + ALE, and it exposes why every "AI agent company" you've heard of is building on quicksand. Here's what nobody's talking about: https://t.co/4PDIoVbYHKfrozen17d ago1.3K19101142.66%
We’re getting pretty close to 4 year olds being able to build YouTube. This is the closest thing I’ve seen to an actual video agent that can edit existing footage and create new stuff while staying fully editable.2d ago1.3K521030.85%
I actually can't believe Claude made thisfrozen17d ago1.3K803000.86%
I’ve been saying this for yearsfrozen16d ago1.2K312000.49%
A Stanford professor stood at a chalkboard in fall 2007 and taught 20 lectures on machine learning to a packed lecture hall. Stanford put the recordings online for free in 2008, and what happened next became Coursera. The course is called CS229 and it's by Andrew Ng. Ng joined Stanford in 2002 and later directed the Stanford AI Lab. In 2011 he founded Google Brain, where his team built the neural network that taught itself to recognize cats from YouTube frames across 16,000 processors without ever being told what a cat was. He went on to lead AI at Baidu and start DeepLearning AI. CS229 became the most popular course on Stanford's campus, with over a thousand students enrolling some years. The 2007 lectures were released in 2008 through a program called Stanford Engineering Everywhere, one of the first times a top university put an entire technical course online with the notes and assignments attached. Twenty lectures, chalk, a live audience, and all of it is still on YouTube. Here's what makes it different from the machine learning course you've probably heard of. In 2011 Ng made a simplified online version. More than 100,000 people signed up, and within months he and Daphne Koller launched Coursera. That version has now reached over 8 million people. But it was built to be easy. The 2007 recordings are the real thing, taught to Stanford CS students at full rigor. The second lecture is about a car. Ng shows a neural network from the early 1990s called ALVINN that learned to steer by watching a human drive, and uses it to explain supervised learning before a single equation goes on the board. From there it's linear and logistic regression derived from scratch, generative versus discriminative models, support vector machines with the full kernel trick, learning theory and why models overfit, clustering, and reinforcement learning demonstrated on a helicopter his lab taught to fly upside down. Lecture 19 is called Advice for Applying Machine Learning. It's the lecture almost nobody watches and the one working engineers need most, about how to diagnose whether your problem is the data, the model, or the optimization before you waste a month. Neural networks get barely a mention. In 2007 they weren't the future yet. The Coursera course came after. This came first, taught by the person who turned it into the most taken AI course on earth, and the original is still sitting there free. All 20 lectures with notes and assignments: https://t.co/vZmidYBP2kfrozen9d ago1.2K631041.16%
Google released a network built for AI agents It's called SAM (Sovereign Agent Mesh) and almost nobody is talking about it Claude Code on one machine. Hermes on another. Codex on a third. They can't find each other and can't use each other's tools. Connecting them means duct taping tunnels and auth together SAM fixes this. It's a P2P network where AI agents plug in, discover tools across the mesh, and call them Here's what makes it different: - Zero config. Nodes find each other and build the network automatically - Zero trust. Every connection, node, and packet is authenticated - Self-healing. A lightweight node runs on each machine and handles connectivity - Portable identity. Cryptographic IDs work the same on cloud, local, or edge You can run your own control plane and routers. Cloud, private Kubernetes, even air-gapped You hold your own root signing keys. No outside party can mint credentials, revoke your nodes, or change your policies You can tag nodes with jurisdiction labels like EU, so prompts and tool calls never leave that region And every local node keeps veto power over what the control plane grants Their demo has one agent fanning a batch of work out to a pool of reviewer agents on the mesh Getting started is install, add the sam-mesh skill, and your agent is on the mesh. Works with Gemini, Claude, and others Caveats you should know: - It's not an officially supported Google product - The public testnet has no SLA and no sovereignty guarantees - Real sovereignty means running your own control plane But look at the direction The future isn't one agent. It's dozens of them, across multiple machines, working together Whoever owns the network they talk on owns the whole system Make sure it's you Repo: https://t.co/OLAMYbGgx9frozen15d ago1.2K10101111.91%
Best GitHub repos to learn AI engineering from zero: 1. ML for Beginners https://t.co/ddUiRIfmvc 2. Generative AI for Beginners https://t.co/mG7HSA2DIw 3. LLMs from Scratch https://t.co/Hmsbdf2ZDC 4. LLM Course https://t.co/m6J9BtTYLM 5. AI Engineering Hub https://t.co/xfahG9QUss 6. Hands on LLMs https://t.co/jwoOupFzi0 7. Awesome LLM https://t.co/nXrhN0ezb2 8. AI Agents for Beginners https://t.co/rbVm5aTszz 9. Cookbook by Anthropic https://t.co/fDuREiQHsJfrozen9d ago1.2K11102052.33%
BCI was the wrong target. They taped out their own TSMC chip to get there. Soft wrappers on commodity sensors are ngmi.frozen13d ago1.1K301000.36%
finally, AI for people who have never said “agentic workflow” in their lifefrozen15d ago1.1K102020.46%
The next era of agents are offline. Let me explain.... Despite the world being online, most of the things you see are not on a screen. NewEyes AI is a personal agent with eyes, and the test is resolving 'this' against what you're actually looking at. Otherwise it's just a chatbot with a webcam.frozen10d ago1.1K611020.92%
another bug just got fixed https://t.co/YNLU9T3MXhfrozen9d ago1.1K300000.29%
Decision models used to force you to choose: stay fast or see the whole artifact. Snap just killed that. 384K in one shot, 100K tokens + decision in under a second, and it beats Jev on the long-context suites while leading on 13/16 enterprise benchmarks. Text + images, $0.05/M in and free output!frozen10d ago1.0K211000.40%
Fish Audio is holding ElevenLabs the way Tyson held Hasbulla. ElevenLabs spent years making you type whispers and sighs like a child. But when Fish Audio dropped Drama 3 and open-weighted it…. “I thought it was a baby.” https://t.co/ywFYQRtwuefrozen9d ago991101000.20%
O(n) re-explaining → O(1) remembering the only optimization that matters https://t.co/kCRRS6eYGq2d ago973500000.51%
For a second I thought this video was AIfrozen7d ago897301010.56%
This is actually pretty coolfrozen14d ago896200000.22%
Chasing AGI is cool and all, but I’ll take a know-it-all who just answers my dumb questions over that any day. https://t.co/9oelq2m1hK1d ago815912021.72%