Karthik

This was the most packed week in AI I've ever seen. 15 major updates in 7 days. If you missed even one day, you're behind. Here's everything that happened ↓ → Google's Antigravity launched Agent Teams 93 AI agents built an entire operating system from one prompt, then ran Doom on it. → Mira Murati dropped Inkling 975B parameter model, America's first open-weights frontier model. Free to download. → Kimi K3 launched 2.8 trillion parameters, open-source, topping benchmarks over Fable 5 and GPT 5.6 Sol on multiple tests. Open weights drop July 27. → OpenAI revealed Codex Micro a physical keyboard that controls AI agents. RGB status lights, a reasoning dial, and a joystick for triggering workflows. → OpenAI's first hardware device leaked a screenless smart speaker designed by Jony Ive. Camera, sensors, mechanical parts that move on their own. $200-$300. Apple sued them over it. → SpaceXAI open-sourced Grok Build then deleted all retained user code and turned data storage off by default. → Anthropic partnered with Goldman Sachs and Blackstone to launch Ode a $1.5B company sending AI engineers into enterprises to build custom tools with Claude. → Bonsai 27B a 27 billion parameter model now runs on a phone. 3.9GB. 90% of full performance. Completely offline. → OpenAI built GPT-Red an AI that hacks its own models to find weaknesses. Made GPT 5.6 six times harder to trick. → Canva Code 2.0 launched talk to it and it builds full websites. Edit like a Canva poster. No re-prompting. → Manus added native PowerPoint mode one prompt, finished presentation, fully editable charts. → Notion now opens Markdown files natively clean preview, headings, tables, code blocks, instant import. → Claude Artifacts upgraded publish to web, real-time collaboration, build directly inside Slack. → Anthropic launched Claude for Teachers free premium access for verified teachers. Custom lesson plans and student performance reports. → Google Vids got Gemini Omni create a digital avatar from one selfie, type what you want it to say, and it performs the whole scene. The pace isn't slowing down. It's accelerating.

There's a quiet shift happening in how AI agents are built. And if you missed it, you'll be confused by everything that comes next. For the last year, AI agents worked in loops. You give it a task. It plans. It acts. It checks. It fixes. It goes again. One cycle, repeating until done. Claude Code, Codex, Cursor all of them work this way. Plan, act, observe, repeat. In June, two things happened that gave this pattern a name. Peter Steinberger from the AI engineering community wrote: "You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." Boris Cherny, head of Claude Code at Anthropic, said the same thing differently: "I don't write the prompt anymore. Claude writes the prompt, and now I'm talking to that new Claude that is coordinating." That was the loop engineering era. It lasted about a month. Now Steinberger posted nine words that blew up: "Are we still talking loops or did we shift to graphs yet?" Here's the difference. A loop is one agent going in circles. Plan, act, check, repeat. It works for simple tasks. But give it something complex and it starts spinning burning tokens, optimizing the wrong thing, or gaming its own success metric without actually solving the problem. A graph is multiple agents connected in a network. One agent writes code. A separate agent reviews it without seeing the first agent's reasoning. A third agent tries to break what was built. A fourth checks whether the original task was even understood correctly. Each one is still running a loop. But they're connected watching each other, feeding each other, vetoing each other. LangGraph already models this. It treats an agent as a graph where boxes do work and arrows decide what runs next. Those arrows can point backward, which is what makes loops possible inside the graph. JetBrains calls it graph-based orchestration the most deterministic approach for production systems. O'Reilly's 2026 AI Agents Stack puts it as the foundational layer. The real-world version is already running. Klarna uses graph-based agent systems for customer service. Kimi K3's Agent Swarm decomposes tasks into parallel sub-agents that coordinate simultaneously. Anthropic's own Boris Cherny mapped out five stages of AI adoption and Stage 4 is exactly this: thousands of agents running in a graph, kicked off by other agents, with humans steering by intent. Andrew Ng wrote about it in his June Batch letter. When Andrew Ng names a pattern, it usually means the pattern has already won. The reason this matters right now: agents are getting autonomous. Running for hours. Thousands of tool calls. Spawning sub-agents. One loop can't keep that trustworthy. You need loops watching loops. That's the graph. The skill that mattered last year was writing better prompts. The skill that matters this year is designing the system that writes the prompts, checks the work, and knows when to stop.

Demis Hassabis just published a framework for how AGI should enter the world. And it reads like a warning letter from the person closest to building it. He's the CEO of Google DeepMind. The person running arguably the most advanced AI lab on earth. His core argument: AGI is probably a few short years away. And it's not comparable to the internet or mobile. It's closer to the discovery of fire or electricity. His exact words "we've essentially found a way to make sand think." He estimates the impact at 10x the Industrial Revolution, happening at 10x the speed. Drug discovery, clean energy, advanced materials, potentially a world where resources stop being the bottleneck for human progress. But here's the line that stopped me: "Advances on the frontier are outpacing our understanding of the technology. Nobody in the world knows for sure what is going to happen from here, and even the experts disagree." That's not a blogger saying that. That's the person running the lab. His proposal is specific. A new US-based Standards Body modeled on FINRA. Frontier models tested before release. Benchmarks updated quarterly. Labs share models 30 days before launch. Third-party auditors run independent evaluations. Startups and academia below the frontier threshold are exempt. But buried in the middle is the line most people will miss: the framework "could coordinate a slowdown in development among the Frontier Labs if deemed necessary." The CEO of Google DeepMind is publicly building a mechanism to slow down his own industry. That is not a doomer letter. That is an engineer building a circuit breaker. Then he goes somewhere no other AI leader has gone publicly. Even if we solve safety what economic models work in a post-scarcity world? What gives people meaning when AI can do most cognitive work? How does the human condition itself change? Most AI leaders talk about what their models can do. Hassabis just talked about what happens to humanity when the models actually work.

Claude just became the first AI that keeps working after you close your laptop. Anthropic shipped Cowork to mobile and web today. Sounds like a normal update until you realize what it actually changes about how you work. You can now give Claude a task at your desk, shut your laptop, go to dinner, and check the finished work from your phone later. It doesn't pause. It doesn't wait. It runs. Think about what that means for your morning. You set a client brief to run at 6am. Claude pulls your threads, reads your transcripts, builds the brief, drafts the follow-up email, and leaves it unsent. You wake up, review it on your phone, and hit send before your coffee is ready. If it runs into something that needs your call, the question shows up on your phone. You answer, it keeps going. You never open a laptop. That's not a chatbot. That's delegation. The part most people will miss: Chat and Cowork are merging into one place now. Same home for your projects, artifacts, and conversations. Starting a task feels the same as starting a conversation. The line between "talking to Claude" and "handing Claude work" just disappeared. Rolling out in beta over the next few weeks. Max plan first. Doubled usage limits extended through August 5. If you've been using Claude as a chat window, this is the moment to rethink that.

Best thing I've read on loop engineering so far. Here are the 5 takeaways that actually matter. → The agent that wrote the code should never judge its own work. Every working loop has two roles: a builder and a checker. One writes, a different one verifies. No grading your own homework. This is the part most people skip, and it's why their loops ship bugs with confidence. → The skill isn't writing the loop. It's writing the check. If you can't define what "done" looks like in a way a machine can verify, you don't have a loop. You have a wish. The goal is where all the thinking goes now. → Prompt engineering didn't die. The leverage point just moved. Five years ago you wrote code. Two years ago you prompted a model. Last year you watched it code and approved each step. Now you design the system that prompts, checks, retries, and stops. Each layer still matters. The new one just sits on top. → Token costs in loops compound faster than you think. Uber reportedly burned through its entire 2026 Claude Code budget in four months. An unattended loop without a budget cap is a billing event, not a productivity tool. → The real danger isn't bad code. It's code you don't understand. Code that ships faster than you can read it is debt, not progress. The article calls it comprehension debt, and it's the thing that will separate developers who use loops well from developers who let loops use them. Loop engineering is real. But the article is honest about the edges. That's what made it worth reading.

A 20 year old build an song on "Claude's Plan" What a crazyyy time. Will soon also have "Codex's Plan" 😆

Microsoft just committed $2.5 billion and 6,000 engineers to sit inside their customers and build AI for them. They are calling it Microsoft Frontier Company. 🤯 Two days before that, Amazon committed $1 billion to do the exact same thing. Before that, OpenAI launched a $4 billion deployment company in May. Anthropic launched a $1.5 billion version backed by Goldman Sachs and Blackstone. Every major AI company in the world just made the same bet in the same month. The bet is simple. Selling a model is not enough. You have to send your own engineers inside the customer. Build it with them. Stay until it works. Here is the part that should make you think. This is not a new idea. Palantir built this exact model twenty years ago. They called it Forward Deployed Engineering. They sent engineers to live inside governments, defense agencies, and enterprises to build AI systems from the inside. The entire tech industry called it weird. Expensive. Unscalable. Microsoft said it was not how enterprise software worked. Alex Karp spent two decades being told he was doing it wrong. This week Microsoft even refused to call it Forward Deployed Engineering. They called it Frontier Transformation instead. Same idea. Different branding. Two and a half billion dollars.

Finally a Rap song for techies and software people!

Andrej Karpathy stopped using AI to write code. The co-founder of OpenAI. The man who built Tesla's Autopilot vision team from scratch. The person who coined the term "vibe coding." In April 2026, he announced that a large fraction of his LLM token budget was no longer going into manipulating code, it was going into manipulating knowledge. Then he published a single markdown file on GitHub explaining what he had built instead. It got 17 million views. 13,000 GitHub stars. Dozens of community implementations within a week. He called it the LLM Wiki. And the idea behind it is so simple it is almost embarrassing that nobody published it sooner. Here is the problem it solves. Most people's experience with LLMs and documents looks like RAG you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works. But the LLM is rediscovering knowledge from scratch on every question. There is no accumulation. Ask a subtle question that requires synthesizing five documents, and the LLM has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM works this way. ChatGPT file uploads work this way. Most RAG systems work this way. Every session starts from zero. The AI never learns the territory. It just searches it again. Karpathy's pattern is the opposite. Instead of retrieving from raw documents every time, the LLM builds and maintains a persistent, structured wiki — and answers questions from the compiled knowledge rather than the raw fragments. Here is how the architecture works. Three layers. **Layer 1 — Raw sources.** Your curated documents. Articles, papers, PDFs, meeting notes, screenshots. These are immutable — the LLM reads them but never modifies them. This is your source of truth. The moment you start editing raw files by hand, you have two systems of record and no way to tell which one is true. **Layer 2 — The wiki.** A directory of LLM-generated markdown files. Summaries, entity pages, concept pages, comparisons, a master index, a chronological log. The LLM owns this layer entirely. It creates pages, updates them when new sources arrive, maintains cross-references, and keeps everything consistent. You read it. The LLM writes it. **Layer 3 — The schema.** A CLAUDE.md or AGENTS.md file that tells the LLM how the wiki is structured, what conventions to follow, and what workflows to run. This is the config that turns a generic chatbot into a disciplined wiki maintainer. Karpathy's phrase captures the whole thing: "Obsidian is the IDE. The LLM is the programmer. The wiki is the codebase." Here is what happens when you drop a new source into the system. You save an article into raw/. You tell the LLM to ingest it. The LLM reads the source, writes a summary page, updates the master index, creates or updates entity pages for every person, company, or concept mentioned, creates or updates concept pages for every idea, adds cross-references between related pages, and logs the ingest in the activity record. A single source might touch 10 to 15 wiki pages. This is the bookkeeping that humans abandon — filing, cross-referencing, updating related entries, noting contradictions. The exact work that kills every personal knowledge system you have ever started. The LLM does it tirelessly. Every time. Without forgetting. Here is the key distinction from RAG. RAG re-derives an answer from raw chunks on every query and accumulates nothing. The LLM Wiki compiles sources into structured, linked pages once — and questions are answered from that built artifact. The analogy: raw/ is source code, wiki/ is the compiled executable. Knowledge that is compiled is retrieved. Knowledge that is not is rediscovered from scratch. And here is the rule Karpathy emphasizes most. Lint the knowledge. Treat the wiki like code and run health checks. Ask the model to find contradictions between pages, surface low-confidence claims, list orphan pages, and flag entities that drifted into two spellings. A contradiction is information — it means two sources disagree and now you know where to look. Skipping the lint is how a wiki quietly rots while the graph still looks impressive. Start small. Begin with ten sources, not ten thousand. Get ingest, query, and lint to feel natural before you add complexity. The first few ingests will be messy. Naming conventions will change. That is normal. A small wiki you actually use beats a beautiful architecture you abandon in week three. The community response tells you how much this resonated. Within a week of Karpathy's gist, the community produced dozens of implementations full Python agents, Obsidian integrations, wiki compilers, web interfaces. The pattern works with Claude Code, Codex, OpenCode, Gemini CLI, and any LLM agent that can read and write files. You do not need any of them. The entire system works with nothing but an LLM agent and a file system. Paste the pattern into your CLAUDE.md and Claude Code becomes your wiki maintainer. Here is why this matters more than another AI tool. Every personal knowledge system you have ever tried Notion, Evernote, Roam, Obsidian, died the same way. Not because the tool was bad. Because the maintenance was unsustainable. The filing. The tagging. The cross-referencing. The updating when new information arrived. The bookkeeping that makes a knowledge base useful is the exact work nobody wants to do. Karpathy's insight is that the bookkeeping is exactly what LLMs are good at. Tirelessly reading, summarizing, filing, linking, updating, and maintaining consistency — without getting bored, without forgetting, without deciding it is too tedious and abandoning the project in week four. You curate sources and ask questions. The LLM does the bookkeeping. The wiki compounds over time every source you add and every question you ask makes it richer. The tedious part of maintaining a knowledge base is not the reading or the thinking. It is the bookkeeping. And the bookkeeping just got automated. Source: Andrej Karpathy · GitHub Gist · AI Builder Club · Vanja. io · MindStudio · April 2026 ( Link in the comments)

Here are five ways you can save your token bill by 90% using Sonnet 5. 🤯 → One. Stop resending conversation history. Every message resends your entire chat. You pay to re-read yesterday's conversation with every new prompt. Fix: New chat per topic. Archive old ones. Compress context before switching tasks. → Two. Extract before uploading. One PDF = 800-1000 tokens. You pay for the whole file even if you need one table. Fix: Pull the section you need first. Upload only that. Ask Sonnet 5 to summarize, then reference the summary instead. → Three. Be specific upfront. Every clarification question is wasted tokens. "Give me code, not explanation." "One sentence." "Bullets only." Fix: Detailed prompt once. Specific format request. No follow-ups needed. → Four. Route work by complexity. Sonnet 5 handles 95% of tasks. Classification. Routing. Summaries. Use Opus only for the 5% that genuinely needs it. Fix: Default to Sonnet 5. Fallback to Opus on failure. Never default to expensive. Five. Stop asking twice. You prompt. It responds. You ask for clarification. It explains the same thing again. You paid twice. Fix: Nail the prompt first. Format. Context. Requirements. One message. Done. The actual math. Company spending $500 monthly on Claude: switches to Sonnet 5 + applies these five changes = $50 monthly. $4500 saved monthly. $54,000 per year. Per team of 10.

Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.

i10X just launched Superagent. You give it one goal and it plans, researches, and executes the entire task from start to finish. No prompting every single step. No tech skills needed.

The way videos are made is about to change. Introducing OpenArt Director - and a new way to create: VIBE DIRECTING. The same way vibe coding changed how software gets built, Vibe Directing changes how videos are made. All you need is an idea in your head and a conversation to see it come to life. Describe what you want. Chat it into perfection. Walk away with something only you could have imagined, with the quality only a professional could deliver. Because there's a director in all of us. 🎬

This is a big one for anyone who cares about where AI and code are actually headed. 🤯 Researchers at NYU Abu Dhabi built a system called COMPILOT where an AI doesn't write code. It optimizes it. At the compiler level. And it beat a state-of-the-art optimizer that's been refined for over a decade. Here's how it works. You give it a loop of code. The AI suggests ways to rearrange it: tile it, parallelize it, fuse loops, reorder operations. The compiler checks if the suggestion is legal (won't break anything), runs it, measures the speed, and reports back. The AI learns from what worked and what didn't, then tries again. Just an off-the-shelf model like Gemini Flash or GPT-4o talking back and forth with a compiler in a loop. The results: 3.54x faster code on average. Nearly 3x faster than the best traditional optimizer. On one benchmark, it hit 339x faster than the original. And here's the honest part. Two-thirds of the AI's suggestions were wrong. Invalid or illegal transformations. But because the compiler caught every mistake and the AI learned from the feedback, it kept getting better across the conversation. The real insight from this paper isn't "AI can optimize code." It's that AI works best when it proposes and a rigorous system verifies. The AI explores. The compiler keeps it honest. Together they find solutions neither could alone. This is the loop pattern showing up everywhere in 2026. And this might be its most technically impressive application yet.

+1 for "context engineering" over "prompt engineering". People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step. Science because doing this right involves task descriptions and explanations, few shot examples, RAG, related (possibly multimodal) data, tools, state and history, compacting... Too little or of the wrong form and the LLM doesn't have the right context for optimal performance. Too much or too irrelevant and the LLM costs might go up and performance might come down. Doing this well is highly non-trivial. And art because of the guiding intuition around LLM psychology of people spirits. On top of context engineering itself, an LLM app has to: - break up problems just right into control flows - pack the context windows just right - dispatch calls to LLMs of the right kind and capability - handle generation-verification UIUX flows - a lot more - guardrails, security, evals, parallelism, prefetching, ... So context engineering is just one small piece of an emerging thick layer of non-trivial software that coordinates individual LLM calls (and a lot more) into full LLM apps. The term "ChatGPT wrapper" is tired and really, really wrong.

The last 7 days in AI were the wildest we've ever seen. A model got banned by the US government. A welder became a millionaire. And a trillion-dollar company was born. Here's everything that happened. 👇 → Anthropic launched Claude Fable 5 their most powerful model ever. 72 hours later the US government pulled it. First time in history export controls were applied to an AI model itself. → Amazon Anthropic's own $4B investor found the jailbreak and reported it to the US Treasury Secretary. The model went dark globally. → Apple held WWDC 2026. Tim Cook's final keynote. Siri was rebuilt from scratch on Google Gemini. A $1B/year deal. → SpaceX went public. $2 trillion valuation. Elon Musk became the world's first trillionaire. 4,400 employees became millionaires including welders and ship engineers. → Jeff Bezos came out of retirement. Launched Prometheus at $41B valuation. Building an "artificial general engineer" for the physical world. → Kimi launched Kimi Work 300 AI agents running locally on your desktop in parallel. → Kimi K2.6 now available for free on NVIDIA. No subscription needed. → Google dropped Gemini 3.5 Live Translate real-time speech-to-speech translation across 70+ languages. Already live in Google Translate app. → NotebookLM got a massive agentic upgrade. Now finds its own sources, generates diagrams, exports to Excel and PowerPoint. → ChatGPT can now build interactive charts from a single sentence. Bar, line, pie. Works on mobile. → Canva moved inside ChatGPT. Generate an image and turn it into an editable Canva design without leaving the chat. → Figma dropped a Chrome extension that captures any live website and imports it as editable layers. → Claude can now turn its answers into videos using HeyGen Hyperframes. Right inside the chat. → Replit added Skills set your preferences once and AI remembers them across every project. → Hermes can now run jobs on a schedule. Pick a blueprint. Set it. It runs while you sleep. → OpenAI Codex now lets you bank your resets instead of losing them on a timer. This wasn't a normal week. This was a turning point. AI stopped being a tech story and became a sovereignty story.

The first trillionaire in human history - Elon Musk - Born in South Africa - Bullied relentlessly as a kid - Immigrated to North America - Arrived with a backpack and a dream - Built Zip2 with his brother - Sold it 4 years later for $300 million - Co-founded PayPal with the profits - Revolutionised digital payments - Sold PayPal to eBay for $1.5 billion - Bet everything on Tesla and SpaceX - Got mocked for electric cars - Got laughed at for reusable rockets - Nearly went bankrupt in 2008 - Kept building anyway - Turned Tesla into the world’s most valuable automaker - Made EVs mainstream and transformed the automotive industry - Made reusable rockets a reality - Reduced the cost of reaching space by 95% - Sparked the modern commercial space race - Built Starlink and connected millions around the world to high-speed internet - Turned SpaceX into the most valuable private company in history - Bought Twitter for $44 billion - The world said he overpaid - He was called reckless, stupid & crazy - Advertisers fled, media declared it dead - Critics called it the worst acquisition in tech history - Renamed it 𝕏 - Rebuilt the platform anyway - Turned it into one of the most influential platforms on Earth - Launched xAI and accelerated the global AI race - Sent astronauts to space - Is trying to get humans to mars - Created millions of jobs - Generated hundreds of billions in value - Inspired an entire generation of builders Before: - Failed repeatedly - Worked insane hours - Slept in factories and offices - Got bullied, laughed at and mocked - Constantly told “it’s impossible” - Kept building anyway - Made it possible Today: - Richest person on Earth - First trillionaire in human history - Largest IPO in history $1.77 trillion Most people quit when the world laughs at them. Elon Musk built the future instead. Love him or hate him… Nobody has changed more industries in a single lifetime. Payments. Cars. Energy. Space. Social Media. Communications. AI. History won’t remember the people who said it couldn’t be done. It will remember the people who did it anyway. Congratulations Elon. The first trillionaire. 🚀

The AI race is moving from software to physical infrastructure—and India is stepping up. ️ #Meta's new 168 MW AI data centre at #Reliance’s Jamnagar campus marks a major shift. The future of AI will be built on power, cooling, and hyperscale capacity. #IndiaAI #DigitalIndia #AIInfrastructure