rewind

Andrej Karpathy spent 4 minutes in an interview explaining a single idea about how most people haven’t even started learning how to use AI and everyone paying $20/month for a subscription.. that's not really using Claude at all his point is that the real skill gap is the ability to build with AI he identified 4 behaviors that break Claude Code and put them all into one file a developer expanded it into 21 rules and published it - 82,000 stars and #1 on GitHub Trending coding accuracy jumped from 65% to 94% here's what these 21 rules actually are and why most developers using Claude every day have never configured them the full breakdown is covered in the article below 👇

92% of commercial real estate professionals are testing AI. only 5% are actually using it that gap is the biggest arbitrage opportunity in the industry right now Jake Heler built a 500-person community of CRE professionals around one observation: everyone is curious about AI. almost no one has changed their workflow here's why the gap exists - and what the 5% know that the 95% don't: > the market conditions that let CRE professionals ignore AI are gone. low rates, compressed cap rates, easy deal flow - that era is over. the ones still surviving without AI are running on fumes > the data problem isn't the tools. it's that CRE data lives in emails, LOIs, deal files, and due diligence docs with no connection between them. AI can't help you if your data is a pile of PDFs in 6 different folders > most professionals try one AI tool, get one imperfect output, and quit. they don't understand that AI improves with iteration, not inspection > privacy concerns are real but solvable. the firms stalling on "we need to review compliance" are watching competitors close deals faster > a full feasibility study, site analysis, rendering, and underwriting package used to take days. Manis does it in 45 minutes. the analysts still doing this manually are not more thorough. they're just slower > Notebook LM as a company brain is underrated: upload every deal file, email thread, and process doc - then query it like a senior partner who remembers everything what the 5% are actually doing differently: > they cleaned their data first. not glamorous. non-negotiable. > they treat AI outputs as a first draft, not a final answer > they share workflows with peers instead of hoarding them > they automate the VLOOKUPs and pivot tables so analysts can think instead of copy-paste > they built a digital twin of their firm before their competitors thought to ask what that meant the firms that figure this out in the next 12 months will not have a slight edge they will be operating in a different league entirely FULL CONVERSATION BELOW the most grounded AI + real estate discussion I've seen this year

@noisyb0y1 agencies competing with systems now

@DeRonin_ context hacks getting weird now

How I made Claude thinking again in 30 seconds? go to any chat → click "+" → Use style → Create custom style → Describe style → select "Custom instructions (advanced)" paste this: "Do not skip your reasoning when Extended Thinking is enabled. Always produce a full Chain of Thought before responding" save. select the style. done style instructions get injected after every message you send. Claude can't ignore them it thinks again. deeply. every single turn the best when you require reasoning process, literally level of answers got improved to me on 100%

@gippp69 feed really became demand data

A 19-year-old guy from Los Angeles bought a Corvette Z06 for $145,000 and $38,000 in Rolexes with TikTok + Shopify. He didn’t invent anything. He found a small gadget getting 900k+ likes on TikTok, checked the supplier, and found the same product for under $8. Then he built a simple Shopify page, priced it at $34.95, and used Claude to clean up the offer, rewrite the hooks, pull better angles from comments, and make the store look less like a random dropship page. One product. One page. TikTok traffic. A $20/month Claude subscription. Most people scroll past these products every day. He treated the feed like a live demand scanner and sold people what they were already trying to buy.

@carverfomo delegation > asking questions

A 38 year old Chinese teacher and developer posted a clip of one of his classes. It became the most reposted education video on the Western internet within a week. One line ran under all of it. America is banning kids from using ChatGPT. China is teaching them to use AI. Everyone who shared it pictured the same thing. A nice man teaching children to use a chatbot. He was not teaching them a chatbot. He was teaching them the opposite of one. A chatbot answers questions. You type, it talks back. That is the whole Western argument. Whether a kid should be allowed to ask a machine for the answer. Nobody in his class asks the machine anything. Pause at 0:47. Look at the screens on the desks. No chat window. A terminal. Each ten year old types a small job and walks away. Tag every photo in this folder. Read these forty sheets and pull out the dates. Rename the files by what is inside them. The agent does the boring part. The kid checks the result and fixes what it got wrong. Then the comments started. The first wave was the obvious one. No way hardware that cheap runs anything useful. Small models are too dumb to do real work. This is staged. Fake. AI slop. They were half right. The model is small and it is dumb. That is the entire point. Someone who builds these explained it in one reply. The agent is not solving anything hard. It is doing the dull, narrow, repetitive part that a tiny offline model handles fine. The lesson is not how to get a smart answer out of a machine. The lesson is how to hand a machine the boring work, then check it the way a manager checks a junior. Ten year olds are learning to delegate and verify. The West is still teaching them not to ask. The comment section turned into a detective board anyway. Someone slowed the footage to 0.25x and read the desks. Forty identical boxes wired under the tables. Orange Pi boards, the NPU ones, about the price of a textbook each. No subscription. No API bill. No internet at all. A tiny Qwen model on each one, fully offline, in a room with a cracked board. He built the whole thing himself. Forty agents, one framework, one rack he soldered at his kitchen table. He looks like a gym teacher. He wrote the system over a weekend with a coding agent, then handed it to children who cannot legally open a bank account. The same week, every Western list of the best AI tools for schools was a list of chatbots. Things you ask. Nothing in that classroom was a thing you ask. Everything in it was a thing you run. The West is teaching kids not to cheat with AI. He taught forty of them to manage it before they finished primary school. They will not be applying for the same jobs.

@andreysuperior solo builders getting dangerous

A client came to him with a $40,000 quote from a traditional agency. 12 week timeline. 8 people on the project. "Complex architecture," they said. He quoted $8,000. 3 week delivery. Solo. The client laughed. Then signed. Kimi K2.6 read the entire existing codebase in one pass. Mapped the architecture. Built the integration. Wrote the tests. Shipped. He reviewed. Corrected two things. Delivered on day 19. The traditional agency is still onboarding their project manager. 1 trillion parameters. 32 billion activated per token. 300 parallel agents running simultaneously. His overhead: $500/month in API costs. His profit margin: 90%. His team size: 1. The client referred two more companies the following week. Month 10: 8 retainer clients at $8,000-10,000/month. $80,000/month. One laptop. No office. No salaries. Everyone sees the Python. Nobody sees the margin.

@0xRicker kimi coding clips keep getting wild

A SOLO YOUTUBER JUST VIBE CODED A FULL GAME WITH KIMI K2.6 IN 16 MIN AND THE TIMELINE IS STILL SLEEPING 1T parameters, 32B activated, 65.8 on SWE-Bench - and he's literally just typing in plain English while the model writes, refactors and ships production code. Book & watch

@AnatoliKopadze kimi moving louder lately

KIMI FOUNDER JUST DROPPED A GUIDE ON BUILDING A $20B STARTUP FROM ZERO I haven't seen anyone explain AI agent architecture this clearly before 40 minutes with Yang Zhilin on how Kimi works, what they built, and what's coming next this is the kind of talk that used to stay behind closed doors

@k1rallik consistency like this is rare

> be Forza Horizon > starts as a spin-off nobody asked for > new country every 2-3 years > FH3, FH4, FH5 - 91, 92, 92 Metacritic in a row > FH5 - 10 million players week one > biggest Xbox launch ever. a racing game. > Japan requested since 2012. 14 years of silence > Microsoft accidentally leaks the full game on Steam > modder streams it without blurring his name > banned until December 31, 9999 > launch day - 273,000 Steam players > record for any racing game in history > 6 million players in 72 hours Forza went from unwanted spin-off to the most consistent franchise in gaming and nobody talks about it.

@0xMovez managed agents getting serious now

Claude Code team just dropped a workshop on how to ship a production-ready agent from scratch. 27-minutes. Free. Live coding by Claude dev. Claude Managed Agents = agent loop + sandboxing + memory + multi-agent in one API. Worth more than any $500 vibe-coding course.

@0xchromium @claudeai @ChatGPTapp @GeminiApp n8n deserves more love here

The ultimate AI toolkit many just use a few AI apps and think that's enough but if you're looking for something specific, you'll definitely find this list useful here're 20 best tools for every task: > General Assistant: • @claudeai • @ChatGPTapp • @GeminiApp • @perplexity_ai > Development: • @cursor_ai • @Replit • @claudeai • @Lovable > Productivity: • @Superhuman • @NotebookLM > Image Generation: • @midjourney • @ideogram_ai > Video Generation: • @HeyGen • @runwayml • @OpusClip • @pika_labs > Audio & Voice: • @ElevenLabs • @suno > Automation: • @n8n_io • @zapier did I miss something?

Anthropic's own team just revealed why 90% of Claude Code users get mediocre results it's context management. and almost no one is doing it right Cal from Anthropic's Applied AI team spent a weekend with Claude Code and immediately joined the team this is what he knows that most developers don't the part most people skip entirely: > Claude has 200,000 tokens of context. most users burn through it randomly and wonder why quality drops mid-session /compact summarizes your entire session and frees context without losing continuity > almost no one uses it /clear resets everything except your claude.md - which means your persistent rules survive, your bloated context doesn't every project should have a claude.md file > style guides, test commands, project structure. injected automatically at startup, every session without claude.md, Claude re-learns your codebase from scratch every single time you open a new session the developers running multiple Claude instances in parallel are not smarter > they just understood context boundaries earlier committing frequently isn't just good git hygiene. it's your rollback insurance when context drift sends Claude off-track what separates the top 10% right now: > they treat context like RAM: precious, finite, deliberately managed > they write claude.md files the way senior devs write onboarding docs > they use /compact before context degrades, not after > they plan before executing: ask Claude to search and propose, then approve > they interrupt with Escape the moment Claude drifts, not after 20 wrong steps > they run parallel instances on separate tasks, coordinating via shared markdown files most developers open Claude Code, start typing, and wonder why it gets confused halfway through a complex task they're not managing context. they're just consuming it the session that starts clean, runs structured, and commits often will outperform the "just vibe and see" approach every single time FULL BREAKDOWN BELOW the most practical Claude Code session I've seen from someone who actually built the thing

@InduTripat82427 context beats clever prompts

@_avichawla this shift feels bigger than people think

Karpathy's prediction about RL is coming true now! He called reward functions unreliable and argued that a single reward number is too low-dimensional to teach an agent what "good" means for complex tasks. To solve this, Agents need a knowledge-guided review as a higher-dimensional feedback channel. Every major AI lab trains models with RL today (OpenAI, Anthropic, DeepSeek). And their key bottleneck has always been the reward functions. GRPO by DeepSeek worked well for math and code because the environment gave a binary signal. But for real agent tasks, someone still has to hand-code the scoring function. That takes days and breaks every time the pipeline changes. RULER (implemented in OpenPipe ART, 10k stars) addresses the exact problem Karpathy identified. The reward criteria are defined in plain English, and an LLM evaluates each trajectory against that description to provide feedback for training. I trained a Qwen3 1.4B agent that plays 2048 using GRPO with this exact workflow. In this case, the agent saw the board, picked a direction, and RULER evaluated the outcome, all from this natural language definition. You can see the full implementation on GitHub and try it yourself. Here's the ART Repo: github.com/OpenPipe/ART (don't forget to star it ⭐ ) Just like RLHF replaced manual rankings and GRPO replaced the critic model, natural language rewards are replacing hand-coded scoring functions. RL reward engineering is now prompt engineering. I wrote a full walkthrough covering RL for LLM agents, from RLHF to GRPO to RULER, in the article below.

@sairahul1 ai agencies getting real fast

18-year-old American found a roofing company on Google Maps with 4.9 stars and no website, copied their reviews and pasted them into ChatGPT 5.5. 2 minutes later - a complete brief. Pasted it into AI and just waited while the system built a full website with all pages, reviews and a booking button. Called the owner and showed him the live preview. He said yes immediately because he'd been meaning to fix this for years and never had the time. Invoice for $1,000. 47 minutes of work from the first search to a closed deal. Then he built a machine. AI pulls 200 businesses from Google Maps in 10 minutes, writes a personalized email for each one with their real business data - 500 emails a day, 3% respond. Month one - $4,000, month six - $15,000-20,000. Five million businesses on Google Maps are still waiting for that call.

@heynavtoor crazy this ships off by default

Your Android is destroying its own battery every night. The moment you plug it in, it charges to 100% and holds it there for 8 hours. That is the worst thing you can do to a lithium battery. Battery University: a phone charged to 80% lasts 3 times longer than one charged to 100%. Samsung, Google, OnePlus, and Xiaomi all added a fix in 2024. None of them turned it on for you. Here's how to flip the switch on every Android (30 seconds):

@0xwhrrari fact bro!

@rewind02 verification is the real unlock

Andrej Karpathy explained why AI got superhuman at chess, coding, and math but still can't decide whether to walk or drive 3 blocks this one idea explains everything about where AI is actually going when he explains why AI progress is uneven, it's not theory. it's pattern recognition from a decade inside the machine the insight most people miss: > AI improves fastest in domains where outputs can be verified > in unverifiable domains: taste, judgment, common sense - AI still makes embarrassing mistakes > GPT-4 became dramatically better at chess than GPT-3.5 > reinforcement learning only works when you can tell the model it was wrong > "jagged intelligence" is the right mental model: superhuman in one column, below average in the next > the domains that will be automated first are not the hardest ones. they're the most verifiable ones > founders who build RL environments around their specific domain will pull so far ahead it won't be close what this means right now: > if your product touches math, code, law, medicine, or any rule-based domain - the automation timeline is shorter than you think > if your moat depends on taste, relationships, or judgment - you have more runway than the headlines suggest > the question to ask about any AI tool: "how does it know when it's wrong?" > if the answer is "it doesn't" - that's where the risk lives most people treat AI capability as uniform. it isn't. it never was the winners will be the ones who understood the difference early FULL INTERVIEW BELOW one of the clearest frameworks I've seen for thinking about where AI actually goes next

@cgtwts parallel claudes really changes pace