chatgpt is now indexing the web they're indexing the web like google researchers found its name in ChatGPT's traffic called Labrador from May 21 to July 21, every search result in ChatGPT carried a label for where it came from three of the labels were scraping services that pull Google results. the fourth was Labrador, which is OpenAI's own index a few things stood out to me it barely overlaps with Bing. Researchers ran the same searches on both, and only about 1.5% of Labrador's results were in Bing's top 20 if you've been treating your Bing rankings as your ChatGPT rankings, they're not the same thing. free users mostly get Labrador in the default Instant mode, ChatGPT didn't open a single page in 93% of answers it answered from a stored snippet of about 200 characters Paid Thinking mode is different about 75% of its results came from scraped Google, and it actually reads the pages your snippet comes from your H1 and the text right around it ChatGPT ignores your meta description, and the snippet doesn't change until the page gets re-indexed updating a page doesn't update ChatGPT. it keeps a cached copy of your page. pages nobody asks about can stay stale for months. it also doesn't run JavaScript, so anything that loads with JS isn't there OpenAI runs three separate bots. GPTBot collects training data. OAI-SearchBot decides whether you show up in ChatGPT search. ChatGPT-User fetches a page live during a chat. I keep seeing sites that blocked all three to stay out of training and then wonder why ChatGPT never cites them so what I'd do: allow OAI-SearchBot in robots.txt. make sure your H1 still makes sense if someone reads only that line answer the question in your first paragraph put the date and author in visible text near the top render your content on the server ChatGPT still mixes in Google and Bing results. But OpenAI is already testing "prefer our own index" in shopping, so I'd expect more answers to come from Labrador over time Research credit: Peec AI, Search Engine Land, RESONEO, a lot of X posts

Educational content is about to get really good. Scroll to travel through 12 landmarks of Rome. Click Explore to orbit each building and zoom into the details. Explore Rome: I directed the design around a few specific things: White as the foundation. White marble and pale stone, with terracotta roofs, aged brick, granite, dark bronze, soft greenery, and turquoise water. I wanted a predominantly white experience that still reflected the buildings’ actual materials. Recognizable architecture. I kept asking for greater accuracy in the façades, columns, entrances, inscriptions, obelisks, fountains, and statues—checking for overlapping elements and details in the wrong places. A proper 3D sculpture. The simplified horse and rider wasn’t enough. We developed a full model in Blender, reviewed it from multiple angles, and repeatedly refined the face, clothing, proportions, and detail before placing it in the Vittoriano scene. The feeling of flying. The camera needed to look toward the next building, travel above roofs rather than through them, and let you stop to explore. A minimal interface. Just a small Explore control and elegant serif landmark names floating above the buildings, appearing as you arrive and fading as you leave. A carefully directed opening. Large, luxurious serif letters spelling ROME. We tried drawing them, but the motion didn’t feel smooth enough. Instead, each letter appears in pearly white, like the buildings, before the whole word darkens together and fades into the scene. Detail that stays visible. I wanted the finer architectural features preserved throughout the journey, including removing a sweeping shadow boundary that was washing them out. Built with React, @threejs, custom shaders, and the Blender sculpture. The architecture, lighting, water, and camera movement render live in your browser. Landmarks load as you approach and release memory as you move on. Built with @OpenAIDevs Astra, shaped through references, screenshots, and a lot of hands-on visual direction.

How does Sphere Studios deliver breathtaking visuals at an incredible scale? Hundreds of Lenovo ThinkSystem servers, powered by @AMD EPYC Server CPUs and @NVIDIA accelerated computing, help Sphere Studios rapidly render the ultra-high-resolution content behind their awe-inspiring experiences. See how Lenovo technology is helping bring the future of live entertainment to life at Sphere:

Three Kami releases this week. Sequence, class, and ER diagrams now render in Kami’s own palette, clean enough to print. Installs and packaging steadier too. Good content deserves good paper.

After Google acquired my AR startup, I started thinking about the next interface, this time for AI. Text and voice work fine, just like DOS did before the Mac. But nothing’s more intuitive than face to face. After two years of research, today we’re launching Atmee v2: a real-time avatar model that makes live video agents feel truly life-like. • Any character from 1 image + 10s of audio • Contextual expressions, gestures, even interacting with the environment • Live video calls, voice only, or video generation At its core, Atmee v2 is a video foundation model, it can render anything, but it’s built to excel at characters, which is what video models like Seedance 2.5 and Wan 3.0 are mostly used for. In comparison: Atmee v2 • Up to 50x faster • 10-100x cheaper Build tutors, sales coaches, support agents, companions, game characters, even whole living worlds. The interface after text and voice starts today.

It's interesting how when we talk about AI agents, we talk about the human being in the loop. But the human is actually the gap in the loop. An ideal agentic loop runs closed; to involve a person, it has to halt, freeze state, render it legible, pass it over, and sit there waiting before it starts again on a changed input. Despite this latency, I do think human-in-the-loop is a necessary evil for AI agents. Even trained on our own context and history, our internal agents land wrong roughly 2/10 times. And that's fine when the agent is drafting something, but not when it's set to do a task. So ideally, every human-involved checkpoint in the agentic process should buy you accuracy if it's costing you time.

I've been enjoying Victoria Whitworth's new work, The Book of Kells: Unlocking the Enigma. I've actually never seen the Book of Kells in person, somewhat to my embarrassment. I've been doing some reading about the origins of Christianity this year, however, and I figured I should know something about the most famous Irish manuscript. (Perhaps the most famous manuscript, full stop.) Reading the book, I was struck by how much the contents have suffered over the past ~1200 years (enduring everything from water damage to reckless malfeasance in attempted nineteenth century restoration), and I wondered whether AI could help give a sense for how the work might originally have appeared. I downloaded the Internet Archive's PDF and asked my friendly neighborhood agent to use gpt-image-2 to render each page the way it imagines it might have originally appeared. Remarkably, this all worked with a single prompt, with the agent spinning up 48 workers, since each page took a minute or two. (I'm sure that someone wiser than me could prompt the model better, ensuring somewhat more historical accuracy in color restoration and so forth. There is no gold leaf in the Book of Kells!) This part of the project went from conception to completion before I'd finished my morning coffee. I then wanted some easy way to view the results online, so I asked Stripe Projects ( to host the result on Vercel. That also worked in basically a single prompt: I also figured that people might want an easy way to download the full PDF of updated images, but it's a large (~200MB) file, so I decided that I should charge $0.10 to cover bandwidth costs using @MPP. I asked my agent to set this up, and it basically worked smoothly, though I had to tell it what MPP is (I guess it's not yet in the pretrain) and also manually set up the Cloudflare account that actually hosts the PDF and configure the API key. (Vercel seemingly has a 100MB limit.) The purchases now show up in my Stripe account alongside all other activity. The site now has a ready-made agent prompt for anyone who wants to download the whole thing. I'm guessing that we'll see a lot more UIs like this in the future. I remain pretty intrigued by the intersection of agents, micropayments, and stablecoins. I don't know much about managing crypto wallets from the CLI, but now AI can do that for me, while Stripe seamlessly handles turning it all back into fiat. So what is the moral of the story? • Whitworth's book is very good, and you should buy it. • The Internet Archive continues to be wonderful and a civilizational treasure. • While there are rough edges, setting up third-party services via the CLI now basically works. I'm pretty sure I wouldn't have bothered with any of this if I couldn't have outsourced almost all of the work to AI. • The image models have gotten very good. • There will probably continue to be all kinds of interesting applications of AI to history. (The Vesuvius Challenge of course being a shining pioneer.) • These days, I often find myself building single-use sites for things I'm learning or for books I'm reading. I think this is a cool new category of software.

I’ve set the agenda for the Hospitalogy AI Retreat, 11/2-11/3 in Phoenix...an opportunity for health system senior leaders to navigate healthcare’s hardest problems with their peers. The agenda: • Reorganize or Explain: Healthcare’s AI Fork in the Road. Dr. Jonathan Slotkin of Geisinger on what’s real, what’s vaporware, and the governance frameworks Geisinger runs on. • Who Has to Say Yes: AI Governance and Physician Alignment. I’m moderating Dr. Neil Gheewala (US Heart and Vascular), Michael Batt (Hall Render) and Dr. Nadia Smati (Faculty at Houston Methodist and Partner at Ainsley Health) on the gap between the C-suite that bought the tool and the clinicians who will (probably) use it. • Hardwiring Trust and Social Capital in a Time of Turmoil. Dr. Tom Lee of Press Ganey on why trust is the rate-determining step in AI adoption. • Hospitalogy Hot Seat. No slides. Attendee participation. This is going to be a lot of fun. Content runs two hours a day. The rest is roundtables, 1:1s, workshops and probably some golf. Travel and hotel are covered for approved attendees. Apply here: @GeisingerHealth @HallRender @MethodistHosp @PressGaney @Hospitalogy_HC

A useful but little-known Claude API feature 🌊 Waiting for the entire tool call argument to finish before doing anything? You can start earlier. Claude's Fine-Grained Tool Streaming delivers tool call arguments as they're generated, token by token. It enables real-time UI updates and early processing of large arguments. 📌 Title: Fine-Grained Tool Streaming 🔗 URL: 🧩 Overview Standard tool calls wait for the full argument JSON before execution. But when arguments are large (long code blocks, documents, etc.), waiting for completion wastes time. Fine-Grained Tool Streaming sends partial argument data as streaming events (input_json_delta) while the model is still generating. You can render progress in the UI or start preprocessing in parallel. 🛠 How to use it Call the Messages API in streaming mode. Tool call arguments arrive as input_json_delta events incrementally. Display them in your UI as they arrive, or begin preprocessing once key parts are confirmed. The actual tool execution still happens after the full JSON is complete, but users see progress in real time. 🏗 Building it into production ・Coding UIs: show code being generated in real time in the editor. Users watch the code being written instead of staring at a spinner. ・Document generation: when producing long text as a tool argument, render a live preview as content streams in. ・Data transformation pipelines: start validation or schema checks on partial argument data as soon as key fields are confirmed. ・Progress UIs: visualize streaming progress of tool arguments so users can see what the agent is about to do. 💡 Use cases ⌨️ Real-time code generation preview 📝 Incremental document rendering ✅ Early argument validation 📊 Tool call progress visualization ⚠️ Watch out Streaming data is incomplete JSON, so don't try to parse it as valid JSON mid-stream. Use partial data for preview and preprocessing only; execute the actual tool only after the full JSON arrives. The event handling code is somewhat complex, so leverage your SDK's built-in helpers. ✨ Showing nothing during wait times is a big UX miss. Stream tool arguments to build interfaces that feel responsive and alive. #Claude# #LLM#