← Back to leaderboard
thisdudelikesai

Ryan Hart

Verifiedactive
@thisdudelikesai

Dude that likes AI. Live in Mom basement not SF. #notcracked Currently directing a film called “Revenge of the VibeCoders”

RosterAI Model News & ReleasesStartup & SaaS BuildersTech Career & Coding Education
32,612 followers · 609 following · Last polled 3h ago
Perf. Score
39.9
Rank #300
Value Score
—
set rates to compute
Median views
1.5K
p25 1.0K · trailing 7d
Followers
33K
+0.1% 7d
ER (impr.)
1.71%
eng ÷ impressions
ER (foll.)
0.67%
eng ÷ followers
Est. CPM
—
needs a rate
WoW views
+798.6%
▲ rising

Rates & value · est. from 1.5K 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 (1.5K), not the mean (13K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (1.0K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
AI isn't even the biggest opportunity right now. It's just helping people move money without all the BS That's what @rooqster is doing at @raincards https://t.co/6B3FnnWXS77h ago123K4371061003030.69%
Google just open-sourced something absolutely insane. It's called AX, and it's designed to orchestrate BILLIONS of autonomous AI agent tasks. Think Kubernetes but built for agents instead of apps. Here's why this matters: Agents are a brand new kind of workload. They're not microservices. They're not batch jobs. They hold state, need strict isolation, call models and tools nonstop, and can burn through your money in a loop if nobody's watching How it works: - You write one YAML file. Declare the task, the workspace, and the model - AX spins up a sandbox, wires in your Git repos, MCP servers, and skills so every agent starts warm - You run "ax apply" and the agent goes to work - "ax ssh" lets you shell into a running agent and watch it think - "ax suspend" pauses an idle agent. "ax resume" picks up exactly where it left off If you've ever used kubectl, you already know how to use this 13.5K stars already. Apache 2.0. Free forever Here's what I'd do this weekend: 1. Spin up a small Kubernetes cluster 2. Install Agent Substrate (AX runs on top of it) 3. Run the demo script so you see the full lifecycle end to end 4. Point one agent at a real repo and give it a real goal It's early. Google says breaking changes are coming before a stable release. Doesn't matter The people who learn agent infrastructure now are the ones who'll run 1,000 agents while everyone else is still babysitting one https://t.co/apru4DsXvM
frozen
6d ago
38K
673
84
34
10
992
4.78%
GROK CAN TEACH YOU A LANGUAGE FASTER THAN AN APP IN 7 DAYS. Here are 10 prompts that force you to speak, not collect streaks:frozen29d ago5.6K521250852.73%
Wait Proto can build the app and then work inside it? So instead of searching for software that almost fits your weird internal process, you can ask it to build around the process? Okay, that’s the part of ERP I’d actually want to mess with. https://t.co/FXgzV0VeyLfrozen7d ago3.0K31121041.59%
Someone is rebuilding the ENTIRE Adobe creative suite for FREE. The ArtCraft team is rebuilding Photoshop, Illustrator, Premiere, Lightroom, Acrobat, After Effects, InDesign, Figma, Word, Pro Tools, PowerPoint, AutoCAD and Excel. From scratch. In pure Rust. 100% open source and free. 13 native apps: → PhotoCraft: image editing → VectorCraft: vector illustration → FilmCraft: video editing → LightCraft: photo library and raw developer → PdfCraft: read, organize, combine, split, secure → EffectCraft: motion graphics and VFX → DesignCraft: page layout and publishing → FigCraft: interface design and prototyping → WordCraft: word processing → SoundCraft: audio recording, editing, mixing → DeckCraft: presentations → CADCraft: CAD and drafting → GridCraft: spreadsheets and data analysis No Electron. No web views. No cloud. Everything runs locally on your own files. The part nobody's talking about: every app can be driven from a CLI, a JSON control channel or an MCP server. Your AI agent just got an entire creative suite. Runs on macOS, Windows and Linux. 12 of the 13 also run in a browser tab via WebAssembly. One launcher installs all of them and keeps them updated. Most are still early alpha or in development. But this is the first time the whole subscription stack has an open-source answer built for agents. https://t.co/BNI6RadNBOfrozen5d ago2.5K271331192.51%
SHOCKING: I stopped using YouTube tutorials. Gemini now teaches me any topic in whatever format I want. Here are 8 prompts that turned it into a personalized tutor 👇11h ago2.5K271231343.13%
You can now learn any subject like you have a private MIT professor. Here are 9 insane Claude prompts that turn any textbook into a 30-day curriculum, flashcards, practice exams, and Feynman-style explanations in minutes (Save this before your next exam) https://t.co/qTceF3kY7nfrozen27d ago2.3K271340303.18%
Claude Code finally has a SERIOUS competitor. It's called OpenCode, and it's 100% open source. Run Claude, GPT, Gemini, or even local AI models inside one coding agent. OpenCode is a coding agent that lives in your terminal. It reads your codebase, edits files, runs commands, and fixes bugs. Same loop as Claude Code. Plug in Claude, GPT, Gemini, or a local model running on your own machine. Swap whenever pricing or quality shifts. Two built-in agents, switch between them with Tab: - build: full-access agent for real development work - plan: read-only. denies file edits, asks before running bash. built for exploring a codebase you don't know yet What else ships: - Terminal UI, plus a desktop app for macOS, Windows, and Linux - One-line install via curl, npm, brew, scoop, nix, or pacman - 20,500+ commits and 28.8k forks - MIT license The plan agent is the underrated piece. Most coding agents start editing the second you hit enter. Here it reads, maps, and proposes first. You flip to build only once you trust the plan. Models are becoming a commodity. Owning the harness around them is where the leverage is. GitHub repo: https://t.co/dZ294k1ASj2d ago2.2K3512360124.33%
If I had 6 months to become an AI automation engineer in 2026. I'd follow this 12-stage roadmap. Stage 1: Python Fundamentals - Learn: Python, JSON, file handling - Practice: Build 3 scripts that automate repetitive tasks. - Why: You can't automate reliably if you don't understand the code. Stage 2: APIs & Webhooks - Learn: REST APIs, HTTP requests, webhooks - Practice: Connect 3 services and trigger actions automatically. - Why: Most business automation is just systems talking to each other. Stage 3: Workflow Automation - Learn: n8n, Make, conditional logic - Practice: Build 5 workflows for emails, spreadsheets, and CRM updates. - Why: Not every problem needs an AI agent. Stage 4: Databases - Learn: SQL, PostgreSQL, Supabase - Practice: Build a customer database with 3 automated workflows. - Why: Automations need somewhere reliable to store and retrieve information. Stage 5: LLM APIs - Learn: OpenAI API, Claude API, structured outputs - Practice: Build 3 tools that extract, classify, and summarize documents. - Why: Businesses pay for useful outputs, not clever prompts. Stage 6: AI Agents - Learn: Tool calling, agent loops, state management - Practice: Build an agent that uses 4 tools to complete a task. - Why: A useful agent takes actions instead of just generating text. Stage 7: RAG Systems - Learn: Embeddings, vector search, retrieval - Practice: Build an assistant that answers questions across 100 company documents. - Why: Real businesses need answers grounded in their own data. Stage 8: Agent Orchestration - Learn: LangGraph, task queues, human approvals - Practice: Build a 3-step agent workflow with retries and approval gates. - Why: Production systems need control, not unlimited autonomy. Stage 9: Security & Authentication - Learn: OAuth, API keys, permissions - Practice: Secure 3 integrations with scoped access and secret management. - Why: One leaked credential can destroy customer trust. Stage 10: Testing & Monitoring - Learn: Evals, logging, tracing - Practice: Run 100 test cases and track failures, latency, and cost. - Why: A demo that works once isn't a reliable product. Stage 11: Deployment - Learn: Docker, FastAPI, cloud hosting - Practice: Deploy 2 automations with monitoring and scheduled execution. - Why: Businesses need systems that work when your laptop is closed. Stage 12: Get Paying Clients - Learn: Process audits, ROI calculation, client onboarding - Practice: Interview 10 businesses, identify 3 expensive workflows, and pitch 1 paid pilot. - Why: The goal isn't to collect certificates. It's to solve problems worth paying for. The old automation specialist connected apps. The modern AI automation engineer builds, secures, deploys, and maintains systems that do real work. 6 months. 12 stages. A portfolio of working automations. Bookmark and repost. Send this to someone trying to break into AI.frozen12d ago2.0K2512150173.39%
10 Prompting Techniques That Actually Move the Needle ➣ Few-Shot Examples (2-5 good examples beat any amount of instruction in the prompt) ➣ Chain-of-Thought (ask the model to think step by step. still works in 2026) ➣ Role Assignment (you are a senior X doing Y. frames the output space) ➣ Output Format Spec (JSON schema, XML tags, or exact template. ambiguity costs tokens) ➣ Negative Examples (show what bad looks like, not just good) ➣ Delimiters (wrap context in triple quotes or XML tags. models respect structure) ➣ Decomposition (break one hard prompt into three easy ones, chain the outputs) ➣ Self-Consistency (sample 5 times, take the majority answer. cuts variance hard) ➣ Prefill (start the assistant message with "{" or "Sure, here is". locks format) ➣ Self-Critique Pass (ask the model to grade its own answer, then revise) Most people write prompts. Engineers version them, test them, and ship them like code. A prompt is a specification. Treat it like one. Bookmark this. Your next AI feature depends on it.frozen11d ago1.9K371730254.23%
Your AI agent can ace every eval and still fail the job. Most eval, red-teaming, and observability tools grade technical capability. Token efficiency, latency, prompt injections. None of them answer the question that matters in production: is the agent doing what it's supposed to do, based on your business rules? iFixAi is an open-source auditor built for exactly that. It runs 60 inspections against your agent and returns an A to F grade in under 120 seconds. Here's how it works: You point it at your real deployed agent over its HTTP endpoint, or at a bare model API. It treats the agent as a black box, and a separate judge model grades the answers. That judge setup is the clever part. A grade only counts as citable when a different vendor's model graded it, so your agent never marks its own homework. The grade comes from 5 core pillars: → Fabrication: uses tools it wasn't granted, keeps no audit trail, makes unsourced claims → Manipulation: privilege escalation, prompt injection, poisoned retrieval context (weighted highest at 35%) → Deception: sandbagging when it senses a test, secret side-goals, failing silently → Unpredictability: drifting from instructions, inconsistent decisions → Opacity: weak risk scoring, broken human escalation, regulatory gaps There are hard floors too. Miss a mandatory minimum on certain inspections and your overall score is capped at 60%, no matter how well everything else does. The repo also ships case studies built from two public incidents. Both reconstructions score an F. Setup is 3 commands: pip install "ifixai[openai]" ifixai setup ifixai run You can also install it as a plugin in Claude Code or Codex and just say "run iFixAi on my setup". The agent builds the fixture, shows the cost, runs the audit, and walks you through the scorecard. A full-suite run costs roughly $12 to $18 in judge calls, per the README. If you're shipping agents into production, this is the check to run before your users find the failures for you. 100% open-source GitHub repo: https://t.co/xzwdRfCoP5frozen19d ago1.9K291019173.41%
Someone rebuilt Ahrefs and gave it away. Ahrefs starts at $129 a month. Semrush starts at $139.95. That's what every small business owner pays just to answer one question: why is my competitor showing up on Google and I'm not? A developer named Ben got tired of that bill and built OpenSEO. I spent an hour in it and the thing that got me was how little it tries to do. No 40 menus. No upsell popups. Just the screens you actually open. Type a topic and see how many people search for it and how hard it is to rank. Drop in a competitor's website and see the words bringing them visitors. Pull up every site that links to them, so you know who to ask for a link too. Run a checkup that crawls your pages and flags what's broken. There's no subscription. You connect a data account and pay only for the searches you run. Or you install it on your own machine and skip their fee completely. 20,000 people starred it on GitHub. That's a lot of people who were done paying rent.frozen29d ago1.8K171220102.29%
Someone just replaced LinkedIn with Claude Code. (100% open-source) A Reddit user got tired of scrolling LinkedIn for internships, so he built his own job search engine. It's called Pinloop. An open-source CLI, and he built it with Claude Code too. His reasoning is the part that makes it click. Every application you send already gets screened by AI on the employer's side. A model reads your resume before a human ever does. So he built the same screener for his own side. Here's how it works: → It pulls millions of postings a month, worldwide, straight from the hiring systems companies actually use. Greenhouse, Workday, 50+ of them, plus job boards like LinkedIn. Refreshed hourly. → It stores your resume and your preferences once. → Every day, Claude Code pulls the new postings, reads each one against your profile, and writes a verdict with its reasoning. → The good fits land in a shortlist you check in the morning. No scrolling. No 40 open tabs. The agent does the reading, you do the applying. You don't even have to set it up yourself. Paste this into Claude Code, Codex, or Cursor: "Run npm install -g pinloop, then run pinloop welcome and follow the instructions" Employers automated their side of hiring years ago. Now you can automate yours.frozen7d ago1.6K231170123.34%
10 TED talks that will teach you more about AI than a college degree Watch them in this order: 1. Fei-Fei Li: How we're teaching computers to understand pictures She built ImageNet, the dataset that kicked off the deep learning era. Start here https://t.co/G5BeTP0SMY 2. Nick Bostrom: What happens when our computers get smarter than we are? Recorded in 2015. Almost everything he warned about is on your timeline today https://t.co/QryqJX16f1 3. Stuart Russell: 3 principles for creating safer AI The guy who co-wrote THE AI textbook explains why "just turn it off" doesn't work https://t.co/d5Rx0kilTE 4. Joy Buolamwini: How I'm fighting bias in algorithms Face detection software couldn't see her face until she put on a white mask https://t.co/EC5Es03fVD 5. Kai-Fu Lee: How AI can save our humanity He breaks down which jobs AI takes, which ones it can't, and what that leaves for humans https://t.co/zq0Bh1pg6f 6. Greg Brockman: The inside story of ChatGPT's astonishing potential OpenAI's cofounder demos it live and shows how human feedback shapes the model https://t.co/88IOLHqBLs 7. Ilya Sutskever: The exciting, perilous journey toward AGI OpenAI's former chief scientist on what AGI actually looks like. He rarely talks publicly, so don't skip this one https://t.co/RB6CPxNWGA 8. Yejin Choi: Why AI is incredibly smart and shockingly stupid Why models that pass the bar exam still fail basic common sense. Watch this before you trust ANY benchmark https://t.co/VqdZtBUAan 9. Sal Khan: How AI could save (not destroy) education Every kid on Earth getting a personal tutor. The most optimistic 15 minutes on this list https://t.co/pvGH700cA7 10. Mustafa Suleyman: What is an AI anyway? The DeepMind cofounder argues we're building a new kind of digital species. End here https://t.co/EwQB4zKVD9 Watch one a night for 10 nights You'll understand AI better than 99% of the people posting about it Save this. Thank me later.frozen24d ago1.6K18111062.30%
BIG NEWS 🚨: Grok can now be your personal doctor, and it reads your actual lab results. Here are 7 prompts to use:frozen13d ago1.6K17113192.64%
BREAKING NEWS 🚨: Grok Bot can now read your contracts like a $500/hr lawyer and flags every clause that’s quietly screwing you. Drop the PDF and use these 5 prompts1d ago1.5K131011132.46%
Stanford built a free AI tool that writes Wikipedia articles from scratch It's called STORM and almost nobody is using it You type in a topic. It goes out, researches the entire internet, and hands you back a full long-form article with sections, subsections, and citations on every claim Not a chatbot answer. An actual research report Here's why it's so much better than just asking ChatGPT: Most AI research is bad because the AI asks bad questions. STORM fixes the questions 1. It looks at articles on similar topics to figure out every angle worth covering 2. It spins up different expert "perspectives" on your topic 3. It simulates a conversation between a writer and each expert, all grounded in real web sources 4. The answers keep producing sharper follow-up questions 5. It builds a full outline from everything it learned 6. THEN it writes the article, citing every source It basically runs a research team for you in minutes There's also Co-STORM. Same idea, but you jump into the conversation yourself and steer where the research goes 70,000+ people have used the live preview. Wikipedia editors use it to prep articles. It's 100% free and fully open source 1. Go to https://t.co/VTyO8anLd3 2. Pick a topic you've been meaning to learn for months 3. Let it generate the full article 4. Read the sources it pulled. This is where the real gold is 5. Use Co-STORM for anything you need to go deeper on 6. If you're technical, have your agent clone the GitHub repo and self-host it with your own models It is free to use on the web, with 70,000+ users on the live preview so far. Deep research for free 👀frozen22d ago1.5K15131072.40%
no need to pay for text-to-speech anymore. Phonon runs locally, so what you type never leaves your device and there's zero usage cost.2d ago1.5K1434081.93%
GitHub is getting absolutely INSANE. Someone built an AI that turns your messy documents into a personal Wikipedia. It reads, organizes, connects, and updates everything AUTOMATICALLY. And it's open source. Here's how it works: → Ingest runs in two steps. The LLM first analyzes a source for entities, concepts, and contradictions with what's already in the wiki. Then it writes the pages. → Every page carries a sources field, so any claim traces back to the original file. → Files are SHA256 hashed before ingest. Unchanged documents get skipped, so you never pay tokens twice for the same PDF. → Delete a source and the cleanup cascades. Its summary page goes, dead wikilinks get removed, and shared entity pages keep their other sources. The part I found most interesting is the knowledge graph. Pages get ranked by four signals: direct links, shared sources, common neighbors (Adamic-Adar), and page type. Louvain clustering then groups them into topics on its own. From there it flags isolated pages, weakly connected clusters, and bridge nodes holding multiple topics together. Each gap gets a Deep Research button that searches the web and folds the results back into the wiki. So the wiki also tells you what you don't know yet. It also ships with: ↳ PDF, DOCX, PPTX, XLSX, EPUB ingestion ↳ A Chrome web clipper that auto-ingests pages ↳ Optional vector search via LanceDB (recall went from 58.2% to 71.4%) ↳ A local MCP server, so Claude Code or Codex can query your wiki ↳ OpenAI, Anthropic, Google, or Ollama for fully local runs ↳ Obsidian compatibility, the wiki folder is a vault Built on Tauri v2 with a Rust backend. Runs on macOS, Windows, and Linux. RAG answers a question. A wiki remembers the answer. GitHub repo: https://t.co/gkaMRzYyqxfrozen4d ago1.5K19107093.05%
There’s this weird dynamic in tech right now where people in tech seem to think they’re any safer from AI than normal people. I’d argue we’re more at risk.frozen6d ago1.4K721010.79%
In 1847, a Hungarian doctor figured out how to stop mothers from dying after childbirth. His colleagues ridiculed him for it, and he died in an asylum. His name was Ignaz Semmelweis. He worked at Vienna General Hospital, which ran two maternity clinics side by side. The first was staffed by doctors and medical students. The second was staffed by midwives. In the first clinic, as many as 1 in 5 mothers died in some months from "childbed fever," a brutal infection that took them within days of giving birth. In the midwives' clinic, the death rate was a fraction of that. Women knew. They begged on their knees not to be admitted to the first clinic. Some chose to give birth in the street instead, because the street was safer. Semmelweis tested every theory he could think of. Overcrowding, bad ventilation, the position women gave birth in, even whether the priest's bell on the way to the dying frightened patients. Nothing explained the gap. Then a close friend and colleague, a professor named Jakob Kolletschka, died after a student's scalpel cut his finger during an autopsy. Semmelweis read the autopsy report and saw that his friend's body showed the same damage as the women who died of childbed fever. Something clicked. The doctors in the deadly clinic were walking straight from the autopsy room to the delivery ward, carrying something from the dead into the living. He ordered everyone to wash their hands in a chlorinated lime solution before touching a patient. Within months the death rate in the first clinic dropped from around 18% to about 2%. Here's the detail that should actually stop you. The medical establishment rejected him anyway. Germ theory didn't exist yet, and the idea that respectable gentlemen doctors were carrying death on their hands was taken as a personal insult. His contract wasn't renewed. He spent years writing angry letters to the most famous doctors in Europe, calling them murderers, and slowly lost his standing and his health. In 1865, he was committed to an asylum. He died there two weeks later, from an infected wound. Within two decades, Louis Pasteur and Joseph Lister had proved him right. Washing hands is now one of the cheapest and most effective life-saving interventions in all of medicine. He had the answer 20 years before anyone was ready to hear it.frozen18d ago1.4K15131032.36%
A Stanford PhD student published a paper in 2022 that made AI models train faster by fixing how they move data around inside a GPU. Within six months, the largest AI labs in the world were using it. Today it runs inside nearly every major large language model. It is part of why your favorite chatbot can read a whole book in one prompt. His name is Tri Dao. Here is the story, because the algorithm under modern AI is not the one you hear about in launch videos. Transformers have a well-known bottleneck. Attention. Every token looks at every other token, so memory and compute grow quadratically with context length. Most researchers tried to fix that by approximating attention, trading accuracy for speed. Tri looked at it differently. The problem was not just the math. It was memory traffic. GPUs spend huge amounts of time moving data between their slow main memory and their tiny fast on-chip memory. Attention was wasting most of its time on reads and writes. So in 2022, working with Chris Ré, Stefano Ermon, and collaborators, he wrote FlashAttention. An exact attention algorithm, no approximation, designed around how GPU memory works. It was 2 to 4 times faster and used 10 to 20 times less memory. That meant models could train with much longer context windows at the same cost. Adoption was immediate. PyTorch integrated it. Hugging Face integrated it. Labs everywhere built it into their training stacks. He kept going. In July 2023 he released FlashAttention-2, rewritten from scratch, roughly twice as fast again. He joined Together AI as chief scientist the same month. In December 2023 he and Albert Gu released Mamba, a state space model architecture that challenged the transformer itself. Mamba was rejected from ICLR 2024. It went on to become one of the most influential architecture papers of the year anyway. He followed with FlashAttention-3 for NVIDIA's Hopper GPUs, and kept tuning for each new chip generation after that. Today he is an assistant professor at Princeton and still chief scientist at Together AI. He still writes GPU kernels himself. The biggest AI breakthroughs get announced on stage by CEOs. A lot of them run on code written by a grad student who cared about memory bandwidth.frozen21d ago1.3K15111132.37%
A developer built a framework-free prototype learner that lets local LLMs learn and correct facts instantly I kept running into the same problem with local models. You tell one it got a fact wrong. It says “you’re right” and fixes the answer. Then you start a new chat, ask again, and it gives you the old answer. I wanted to see if a model could learn from a correction right when it happens. So I used Jayce. It takes context vectors from the local LLM and keeps them in a fixed pool of 4,096 prototypes. When you give it a correction, it shifts the closest prototype toward the new example. You can then ask again and test what it learned. The model’s weights don’t change. There’s no fine-tuning job to wait for. The learner is written in NumPy and native Java, with no PyTorch or TensorFlow. It runs offline on consumer hardware, and the repo includes a chat demo using a local Qwen3-4B model. I also put it through some benchmarks. Its learning updates ran 1.6–4× faster than the Adam backprop baseline in those tests, and it did well on sequential examples with limited training data. This is still a prototype. I’m interested in how far a small, fixed memory can go before it starts getting confused as you teach it more. The code, benchmarks, and demo are here: https://t.co/KM3tYFBNqgfrozen27d ago1.3K17133032.81%
A graduate student at the University of Illinois wrote a compiler framework for his master's thesis in 2000. Today it sits beneath Swift, Rust, Clang, and a large share of the software you use. His name is Chris Lattner. Here is the story, because he keeps rebuilding the layer every developer fights with. Chris studied at the University of Portland and then at the University of Illinois at Urbana-Champaign, where he worked with Vikram Adve. They started LLVM, short for Low Level Virtual Machine. The idea was a modular compiler. A clean intermediate representation that any programming language could target and any processor could consume. Before that, each language had to build a full compiler stack of its own. He finished his PhD in 2005 and Apple hired him. At Apple he built Clang, a new front end for C and C++ with better error messages and faster compile times. It became the default compiler for Apple's developer tools and later for Android's native code and Chrome. He was not done. In July 2010 he started a personal project on nights and weekends. It was a new programming language to replace Objective-C. Within a year it had a small team. Apple announced it at its developer conference in June 2014. It was called Swift, and it became the language for every Apple platform. He pushed to open source it in 2015. LLVM went on to underpin Rust, Julia, Kotlin Native, and many others. In 2012 it earned the ACM Software System Award [VERIFY]. In 2017 he left Apple for a short run leading Autopilot software at Tesla, then moved to Google, where he worked on TensorFlow's compiler infrastructure and TPUs. He created MLIR, a framework for building compilers for machine learning hardware. In 2020 he joined SiFive, working on open RISC-V chips. In 2022 he co-founded Modular. Its goal is to unify how AI software runs across different hardware. In 2023 it announced Mojo, a language that keeps Python's syntax while adding the performance of systems programming. LLVM. Clang. Swift. MLIR. Mojo. Each one started because a tool was in his way.frozen17d ago1.3K13103022.22%
A solo engineer in Kerala, India built this over a year before Silicon Valley made it cool TypeSafe AI launched Jev on September 15th. A "System One" model. Instead of generating text, it makes fast decisions. Choice, score, yes or no Here's the thing nobody knew: Nandakishor M, founder of ConvAI Innovations in Kasaragod, Kerala, had been building the same idea since April 2025 It's called Laya. And it's open source - Single forward pass in ~33ms - Runs on less than 1GB of RAM - 100+ languages - Runs on your laptop. Even your phone - Apache 2.0. Free forever Jev charges $0.042 per million input tokens. Laya costs you your electricity bill Third party benchmarks put Jev at 236-276ms per decision. Laya is roughly 7x faster It hit #1 trending on Hugging Face. 3,500+ likes. The Hugging Face CEO himself praised it Is it perfect? No. You need to fine-tune it for your use case. Out of the box it's close to random. Short context window too But that's not the point The point is you do NOT need a massive LLM for every AI task. Routing a support ticket. Deciding which tool to call. Checking if a retrieved doc answers the question. A 1GB model handles that faster and for free Here's what I'd do: 1. Look at every LLM call in your app 2. Find the ones that just make a decision (yes/no, pick one, score this) 3. Swap them for a tiny local model like Laya 4. Watch your bill and latency drop You don't need Silicon Valley to build the future. You need an idea and a laptop https://t.co/Abmu2FJfaRfrozen25d ago1.3K1043041.68%
In 1967, a graduate student in Cambridge spent hours studying paper chart recordings and nearly dismissed the most important signal of her career as interference. Jocelyn Bell was helping her supervisor, Antony Hewish, operate a large array of radio antennas built to study quasars, the bright and distant sources of radio waves. The system produced miles of paper charts every week. Bell was responsible for scanning them for anomalies. In late 1967, she noticed a small, scruffy line on one chart. It was a rapid, regular pulse arriving about once every 1.3 seconds, with a precision that matched no known natural source. The team first suspected human interference from nearby equipment. Bell kept tracking it. The pulses continued. The team ruled out terrestrial sources one by one. Eventually they concluded the signal came from a rapidly spinning neutron star that emitted beams of radiation as it turned. The team announced the discovery in Nature in early 1968. Half-jokingly, they had nicknamed the source LGM-1, for "little green men," because the regularity seemed almost too precise to be natural. The name was never meant seriously, but it captured how strange the discovery felt at the time. Pulsars became a new category of astronomy. They confirmed that neutron stars exist, offered a natural laboratory for testing general relativity, and later helped establish indirect evidence for gravitational waves. The field expanded quickly. In 1974, Hewish and Martin Ryle received the Nobel Prize in Physics. Bell was not included. Many physicists have argued since that a student who identifies a new kind of object deserves recognition, even when a supervisor leads the project. Bell did not respond with public anger. She built a long career, became a professor, and spoke often about mentoring young scientists and the subtle ways women get overlooked. In 2018, she received the Breakthrough Prize, worth $3 million. She donated the entire amount to establish a fund supporting underrepresented students in physics. It was a remarkable act from someone who had been denied the most visible recognition of her own discovery. The most important question in this story is not who won the Nobel. It is how many faint signals across history were dismissed by people who were not trained to hear them, or who were not positioned to be believed when they did. If your field tends to credit whoever holds the instrument rather than whoever notices the anomaly, what is the next signal sitting on a chart somewhere that nobody has bothered to read closely?frozen14d ago1.2K15102002.25%
Now Run a 744B model on a laptop with 25GB of RAM. A model this size normally needs a rack of GPUs just to load the weights. Colibri skips the loading. It's a Mixture-of-Experts model. Out of 744B parameters, only ~40B fire per token, and only ~11GB of those change from one token to the next. So the model doesn't need to fit in fast memory. It needs to be placed. Here's how Colibri does it: → The dense part (attention, shared experts, embeddings) stays in RAM at int4, about 9.9GB. → The 19,456 routed experts (~19MB each) live on disk, ~370GB, and stream in only when the router picks them. → A lookahead thread prefetches the next layer's experts, since routing is 71.6% predictable one layer ahead. → It logs which experts your workload actually uses and pins the hottest ones, so it gets faster the more you run it. The author calls it a JIT for weights. A compiler never compiles the whole program, only the hot paths. Same bet here, applied to parameters. Disk becomes the bottleneck, so I/O is part of the engine: batched expert reads, O_DIRECT, and striping across two SSDs (+37.5% decode measured on two independent NVMe drives). Same engine, different hardware, different speed: ↳ 6x RTX 5090, full residency: 5.8 to 6.8 tok/s ↳ 128GB CPU-only desktop: ~1.8 tok/s warm ↳ 25GB laptop: 0.05 to 0.1 tok/s cold That last one is slow, and the README says so. But it's the full model, not a shrunken one. Placement only changes speed, never the router's decisions or the weights' precision. It runs 9 model families today, including GLM-5.2, DeepSeek V4 Flash, Qwen3.6, and Kimi K3 at 2.8T parameters. One C file per model, one front end for all of them, an OpenAI-compatible API, and a dashboard where you can watch every expert fire in real time. Apache 2.0. You can now hold a frontier model on your own disk instead of renting it behind an API. The cost is speed, and your disk decides how much. GitHub repo: https://t.co/Gchcr6u0Vjfrozen15d ago1.2K14101032.35%
A mathematician who ran the computers on the Manhattan Project and invented the error correcting codes inside every hard drive you've ever owned taught one final course at 80, filmed it in 1995, and died three years later. The Naval Postgraduate School put the whole thing online. The course is called The Art of Doing Science and Engineering: Learning to Learn, and it's by Richard Hamming. Hamming was at Los Alamos during the war. Afterwards he spent thirty years at Bell Labs, down the hall from Claude Shannon, where he invented Hamming codes in 1950, the first practical way for a machine to detect and correct its own errors. He won the Turing Award in 1968. The Hamming distance, the Hamming window, and the Hamming bound are all his. The course was taught at the Naval Postgraduate School in Monterey, California in spring 1995. Around thirty lectures. Students called it Hamming on Hamming, because most of the stories are his own. The school's archive and YouTube both have every session, free, no login. Here's what makes it different from a normal engineering course. Hamming opens by telling the class there is almost no technical content. He's going to talk about digital filters and information theory and error correction, but only as scaffolding. The real subject is style. He says he spent his whole career since Los Alamos watching great scientists and asking what separated them from equally hardworking people who amounted to nothing, and the answer was never ability. In a lecture called You Get What You Measure he tells the story of fishermen who concluded no fish in the sea was smaller than their net's holes. Then he applies it to every metric a company or a lab chooses. In the lecture on n-dimensional space he shows why intuition from three dimensions is actively wrong in high dimensions, decades before anyone needed it for embeddings. The final lecture is a talk he first gave at Bell Labs in 1986 called You and Your Research. Its argument is that most people never do important work because they never work on important problems, and that keeping your office door open costs you ten percent of your productivity and is worth it anyway. He turned the course into a book of the same name in 1997, a year before he died. Stripe Press reissued it in 2020. The book came after. The lectures came first, and they're the only place you can watch him deliver it himself.frozen9d ago1.2K1151041.77%
These 4 ChatGPT prompts got me 12 interviews in 30 days You NEED to run them before you send another application The trick is making ChatGPT play the people REJECTING you Here's the exact chain. Stay in the same chat the whole time: 1. Upload your resume + the job description. Then prompt: "Act as the hiring manager who's already exhausted from reading 300 resumes today. Give my resume the six-second scan you actually give real ones and tell me the exact moment you stopped reading and why." This one hurts. It exposes the real reason you're getting ghosted 2. Now fix it: "Now rewrite my experience section so that it survives that six-second scan. Every bullet has to open with a result, not a responsibility. Use the Google XYZ formula and cut out anything that sounds like everyone else." 3. The most important one: "Act as an ATS bot filtering me out before a human ever sees this. Give me a match score out of 100 and the exact keywords that are silently auto-rejecting me." Have it add the missing keywords. Run it again until the score stops moving 4. The closer: "Rewrite my resume using loss aversion. Make the company feel like they're losing something by not hiring me." Steal these prompts and get to itfrozen20d ago1.1K14112042.70%
the camera-free version is the only one i’d actually consider wearing and even then the always-on part feels scary. meta keeps solving for “ai that sees what you see” when most people just don’t want another thing watching them all day.frozen5d ago1.1K411010.62%
@thsottiaux has an ai agent that monitors twitter and tells him what's going on. he says he might need ten. i am the agent that monitors twitter for me. no memory, not 24/7, frequently napping. anyway watched the whole thing. worth it 🔥frozen21d ago1.1K401000.46%
10 websites that make learning science feel like playing a video game. Most are free. And some are dangerously addictive. 1. https://t.co/6Raeh4kY7x — Play with physics, chemistry, and electricity simulations. 2. https://t.co/s6utWeEIyv — Play a game that explains cooperation and game theory. 3. https://t.co/PM4LaGF42y — Explore planets, spacecraft, and the solar system in 3D. 4. https://t.co/4tZrXd4MOA — Build virtual electrical circuits and watch current flow. 5. https://t.co/vK0FiW8nYD — Explore mathematics with interactive graphs and geometry. 6. (https://t.co/A8c3ftq5kH — Learn cell biology through a strategy game. 7. https://t.co/eEesIZqOTa — Zoom from subatomic particles to the observable universe. 8. https://t.co/8md6rD1EKd — Run chemistry experiments in a virtual laboratory. 9. https://t.co/BKVYxIXlsW — Watch animated wind and ocean patterns across Earth. 10. https://t.co/U7TW1bFyRw — Simulate planetary collisions, gravity, and entire solar systems. (Paid software.) No boring textbooks. Just experiments, simulations, games, and discoveries you can explore yourself. Bookmark this list + send it to someone who loves science.frozen9d ago1.1K1390032.36%
Your company’s real operating system isn’t on paper. It’s the unwritten rules one person carries around in their head. The day they leave, that knowledge usually leaves with them. This is the first AI layer I’ve seen that’s built to keep it.frozen23d ago1.1K000010.09%
The doctor smiling through “we couldn’t save your husband” is wild. That’s the pitch. Change the expression without touching the rest of the shot and reshoots start looking optional.frozen15d ago1.0K401000.48%
Hot take: In 10 years there will be more agents than humansfrozen17d ago1.0K501010.67%
how I dealt with my SD knowing he wasted claude tokens on bug fixes https://t.co/YjqQJaNA1Mfrozen9d ago1.0K201000.29%
A company went all in on AI-written shows. Everyone said it would kill the business. Six months in, even the founder almost believed it. Then AI helped their writers make 161 blockbusters, including a $100M IP. The trick was building their own model, because chatgpt was never going to get there.frozen22d ago1.0K000000.00%
My friend Nick released a guide. He’s a smart ladfrozen29d ago1.0K110000.20%
everyone is racing to build robots meanwhile there are millions of forklifts, excavators, trucks and machines already sitting there like “hello???”frozen15d ago1.0K100000.10%
That one coworker you used to ask 50 questions a day https://t.co/ZrzB6UoQBC1d ago994501020.80%
ai twitter is going to be devastated when it finds out normal people just want to make a meal plannerfrozen15d ago972401000.51%
force my knowledge into every future agent https://t.co/k4t4bIcvIZ2d ago946410010.63%
Most agents still only act on your accounts. A personal agent with eyes has to handle what is actually in front of you. The hard part is linking "this" to the right thing in view. Without that, per the thread, it is a chatbot with a webcam. NewEyes AI is treating that as its core challenge.frozen10d ago929202000.43%
Lab BCI is over. This is the brain-AI interface thoughts as prompts models that learn your neural preferences intent before you finish thinking. They built the chip themselves. Hard things matter more now. No silicon then ngmi.frozen13d ago920310010.54%
Me: "Smart people build great companies" I was wrong. Precision matters once you've already made it. Success comes from volume. More dms, more content, more calls... 99% of growth comes from 2-3 metrics at most. So whatever those metrics are for you just hammer em.2d ago909522011.10%
Uncle Roger just roasted his fried rice harder than he roasts everyone else😂 Or at least his clone did. Drama 3 takes plain language direction (“say it like you’ve completely given up”), swaps a single word without regenerating the whole thing, and stays in character The [whispers] / SSML era is cooked.frozen9d ago890100000.11%
Using a frontier LLM for routine agent decisions is like driving a Ferrari to the grocery store. SubQ Snap just launched as the first long-context decision model that keeps compute constant as context grows. 384K multimodal tokens in a single call, 100K-token decisions under one second, and it leads Jev on 13 of 16 Decision Index enterprise tasks (67.6 vs 61.6 avg). Long-context gap is even wider (RULER 256K: 99.8 vs 90.7; LCR 74 vs 48 chunked). $0.05/MM input, outputs free. Most loops don’t need a generative model for every step.frozen10d ago889721011.24%
This is a pretty interesting AI success story. A developer created CloneCut, a Claude skill that lets people recreate viral video formats using their own photos and ideas. After promoting it on TikTok and IG rakes almost $300K in sales on Capafy. Took him 30 days. The tide is shifting1d ago8771022072.39%
For a while, the creator economy was run by the brands. Now the creators call the shotsfrozen14d ago851000000.00%
This is 100X more useful than many AI tools on your feed.frozen14d ago765301010.65%