AI is useful today. The money behind the frontier is harder to read. Our new research follows the circular deals linking AI labs, chipmakers, cloud providers, and data centers. Who is paying for the AI boom? digitalmeld.io/blog/who-is-p…

IAH ➡️ SFO for OpenAI Dev Day, let’s go!!! Follow along at where.bradgroux.com.

I’m heading to San Francisco for OpenAI DevDay this week, thanks to @OpenAI! I’m looking forward to meeting a ton of brilliant people and hearing what OpenAI has coming next. My schedule includes sessions on agent architecture and Codex, plus a couple of meetups before DevDay. And if the rumors about a wearable announcement pan out, I’ll have even more to talk about. Can’t wait! I’ll be sharing live updates here on X, and at where.bradgroux.com (where you can also see my full schedule) throughout the trip. I’m in San Francisco from September 28–30, so if you’re around and want to meet up, please reach out. I’d love to hear what you’re working on.

My trip is fully set and booked for OpenAI DevDay next week in San Francisco! I'll be there Monday morning through early evening Wednesday if anyone would like to meet up. Keep track of me while I'm on the ground at where.bradgroux.com. Thanks to @OpenAI for having me!

I’ve been complaining about GPT-6 Astra, but I've been using it wrong. I was asking Eisenhower to storm the beaches of Normandy himself. Now Astra plans, delegates to 3-5 Sol threads, and reviews the work. Here’s what changed, plus three prompts to try.

Looking to onboard your credentials to Muse? Unfortunately, there is no way to bulk import credentials from iCloud Keychain or 1Password into Muse (that's by design), but you can semi-automate the process pretty quickly. Full details in the tweet below.

I kept hitting the same dumb problem: I wanted to feed YouTube videos to my AI via Codex, Muse, OpenClaw, Buzz, and every route was annoying. Paste a link and pray the agent can fetch it. Watch it burn thousands of tokens pulling the page, choking on markup, or failing outright. Or install some bloated extension full of popups, accounts, and analytics phoning home. So I built the thing I actually wanted: Minimal Transcript. Tiny Chrome extension. Open any YouTube video, click the icon, get the transcript, TXT, Markdown, SRT, or VTT, timestamps optional. One click. That's the whole product. Why it saves tokens: your agent starts from clean text instead of spending a fortune figuring out the video itself. No fetching pages, no parsing HTML soup, no hallucinating what was said. Paste transcript → ask your question → done. On a 3-hour video, that's the difference between a few thousand tokens of clean captions and tens of thousands of wasted ones. Details I sweated because they matter: • Prefers real captions over auto-generated when both exist • Strips YouTube's hidden screen-reader junk ("0 seconds", "1 minute, 5 seconds") that pollutes naive scrapes • When YouTube gates the download endpoints, it falls back to driving YouTube's own transcript panel, works whenever YouTube's panel works • Remembers your format; copy-to-clipboard if you don't even want a file • No popups, no ads, no accounts, no third-party servers. Private by design. MIT licensed. • Free, sideload in under a minute: github.com/BradGroux/youtube… If folks like it, I'll get it up on the Google and Edge extension stores soon.

After reading @pvncher’s guide, I asked Astra in Codex to audit my workspaces. I revisted approval rules, consolidated skills, and cut skill-description text by 78%. I share the changes, with prompts and code examples you can use to audit your own setup.

We love to save our clients money with AI 😃

My final assessments are in, and I'm now officially an OpenAI Certified API Expert, ChatGPT Expert, and Codex Expert.

We're now officially @OpenAI Partners at @DigitalMeld, and I've just completed my OpenAI Consultative Solutions Practitioner certification.

Did you know that you can now have up to five Ollama Cloud models available to you in Codex? Update your Ollama app, configure your five models of choice, and restart the ChatGPT app! That's it. 20x ChatGPT Pro and an Ollama Cloud Max plan are a dynamic duo. Astra and Sol along with Kimi-K3, GLM-5.3, Gemma4, and DeepSeek-v4-Flash is a hell of a toolset, all in one amazing harness. I just had Kimi-K3 pick up on a project that Astra started, without issue - it just works. Kimi read Astra's history directly from the Codex workspace and picked up where it left off. Seriously, if you're hitting usage limits with Codex and Astra, grab @ollama cloud as a backup. Their $100 Max plans are a steal and open for signups again. ollama.com/pricing#max

Way more detail, numbers and charts in the published study - go.sstb.ai/astra-study

My GPT-5.6 workflow hasn’t carried over reliably to GPT-6 Astra. I reviewed six days of records to understand why I kept repeating instructions. Here’s what I found—and what I still can’t explain.

AI progress is accelerating, and safety has to advance with it. I joined @BradGroux and Robert Groux on @DigitalMeld’s Start Small, Think Big to discuss why we need to engineer for safety while continuing to push the boundaries of what AI can do. piped.video/shorts/Pp1AE6cNg…

There is some pretty wild information about GPT-6 Astra in my findings. Usage over six days was nearly 4 billion tokens across more than a dozen projects. The article with full details will be out in about 3-4 hours.

While I've had some issues with GPT-6 Astra following PRDs and SOPs (I'll figure it out), I have had an amazing experience with Astra and @threejs! I've been working on a synthwave-inspired city builder, with fully customizable (built with blocked layers) and sharable building and vehicle designs. I started off trying to provide it with a massive PRD, and it didn't go so well, so I've started with a smaller prompt and have been doing iterative improvements as we go. What you see here is less than 40% of a $200 ChatGPT Pro plan weekly limit, with GPT-6 Astra on Medium settings (about seven hours total). I did a quick POC between Blender/Unity vs three.js, and quickly chose three.js, because you can see the improvements in-browser immediately. What a time to be alive.

I just released Veritas Kanban 6.1.7. This one got a lot bigger than I originally planned. Overall there were nearly 40 features, fixes, and security improvements. I started by using the latest builds on my Mac and kept finding things that worked in the browser but did not hold up inside the macOS desktop app. Every time I fixed one annoying interaction, I found another, and then another... So, I kept going. The task workspace has been rebuilt around Overview, Plan, Run, Results, and History. Settings is organized by workflow instead of being a long pile of unrelated pages. Fields, dropdowns, units, buttons, page headings, dialogs, drawers, and popouts now follow the same layout rules. Board moves now recover properly when something fails. Conflict messages explain what happened and what to do next. Consecutive card moves no longer fail because the app kept an old revision around. Memory pressure and health checks are also based on the actual process and host state instead of misleading runtime numbers. I spent quite a bit of time on Chat. Board Chat, Squad Chat, and Task Chat now use the same Workbench layout. I removed the bottom dock because it created more problems than it solved. When the window gets small, sidebars and Workbench close instead of fighting over the same pixels. Drafts now survive resizing and channel changes, and the header was reorganized and refined. The quick Command+K feature has better search, keyboard navigation, scrolling, grouping, and focus handling. Board colors now have meaning and no longer have Shadcn remnants. Primary actions and selected navigation now look and behave consistently throughout the app. Failed requests no longer wipe out the form or leave a dialog in a useless state. Task creation, templates, previews, conflict resolution, approvals, workflows, Git actions, and reviews retain enough context to retry safely. Partially created blueprints cannot simply be submitted again and create duplicates (that was annoying!). Templates keep their task fields, critical priority works consistently, optional values can actually be cleared, and workspace run budgets survive updates. Rapidly closing and reopening a task no longer lets old cleanup remove the new window. Enlarged text, mobile layouts, tooltips, Activity rows, reduced motion, and narrow desktop windows received their own fixes. Repository publishing is now tied to the exact managed worktree, repository, branch, origin, and captured commit. If that state does not match, Veritas stops instead of guessing. I also updated Electron and the affected dependencies, tightened the public documentation boundary, refreshed the screenshots and GIFs from the release build, and cleaned up the native About panel. PDF work products are downloads. Open them in Preview or whatever system viewer you use. Veritas no longer tries to be a PDF reader (this caused way more problems than it was worth). This is the final Veritas Kanban release using GPT-5.6 Sol. Over the next few days, I am doing a full code audit with GPT-6. Astra and will release that work as Veritas Kanban 6.2.0. I'll be writing up the side-by-side comparison of the outputs of Sol and Astra. Until then, install or upgrade: brew update brew upgrade --cask bradgroux/tap/veritas-kanban Release notes and details: github.com/BradGroux/veritas…

Doug Gabbard (@DougGabbard) helped train me when I joined @Microsoft. A lot of what I learned from him still shapes how I approach a problem. One of those lessons: a confident answer is not the same thing as a correct one. @RegGroux and I caught up with Doug on the latest episode of the Start Small, Think Big podcast to talk about how that applies to AI, from learning Photoshop to building agents around SharePoint knowledge. My favorite question from the conversation is the one to ask before calling a project finished: What did I miss? Read the blog post for the episode resources and follow Doug's Practicing Photography work. go.sstb.ai/E0037