WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-04-29 1 00:00:00.000 --> 00:00:07.632 Hey everyone, welcome to the Builder's Briefing for April 29th, 2026. I'm Alex, joined as always by Sam. We've got a packed one today — GitHub is about to start charging for something a lot of teams assumed was free, AI labs are funding Blender of all things, and there's a wild healthcare security story. 2 00:00:07.632 --> 00:00:09.833 Yeah, and the theme today really feels like the free lunch is ending. Let's get into it. 3 00:00:09.833 --> 00:00:17.490 Alright, so the big story. Starting June 1st, GitHub Copilot code review — the AI feature that automatically reviews your pull requests — is going to start consuming your GitHub Actions minutes. So what used to be essentially a free add-on baked into your workflow now directly eats into your CI/CD budget. 4 00:00:17.490 --> 00:00:23.370 This is going to catch a lot of teams off guard. I know plenty of orgs that just flipped Copilot review on by default across every repo. If you've got dozens of repos with active PRs, that's a lot of minutes you weren't accounting for. 5 00:00:23.370 --> 00:00:32.127 Exactly. And here's what makes it spicier — there's a separate analysis trending right now arguing that GitHub Actions is already the weakest link in your supply chain security. So you're about to pay more for a CI surface that a lot of teams haven't properly hardened. Unpinned actions, overly broad permissions, leaked secrets — the usual suspects. 6 00:00:32.127 --> 00:00:39.834 Right, and what's wild is that the cost change might actually be a net positive in a weird way. It forces you to be deliberate. Like, do you really need AI review on every single PR, or do you gate it to main branch merges or PRs with certain labels? That's a conversation teams should've been having anyway. 7 00:00:39.834 --> 00:00:47.841 Totally. And the bigger signal here is that GitHub is moving AI features from loss-leader to revenue line. Consumption-based pricing for AI dev tooling — that's the direction. If you're building internal developer platforms, you should start modeling AI tool costs as variable infrastructure spend, not fixed SaaS seats. 8 00:00:47.841 --> 00:00:51.844 Treat it like compute. Budget accordingly. There's a good checklist in the Actions security post too — link in the briefing — worth running through before June. 9 00:00:51.844 --> 00:00:55.597 Okay, shifting to AI news. A couple of really interesting ones here. First — Anthropic just joined the Blender Development Fund as a corporate patron. 10 00:00:55.597 --> 00:01:02.503 That's interesting because it tells you exactly where the AI labs think the next frontier is. They need 3D tools for synthetic data generation, training pipelines, spatial reasoning. Blender is open source, it's the standard, and now it's got serious AI money flowing into it. 11 00:01:02.503 --> 00:01:12.061 If you're building anything in the 3D or spatial AI space, the Blender ecosystem is about to get a lot more attention and tooling. Also worth flagging — there's a legal analysis making the rounds about who actually owns code that Claude Code or Copilot writes. It's got over three hundred engagements and the takeaway is that the IP picture is way murkier than most founders assume. 12 00:01:12.061 --> 00:01:18.367 Especially for enterprise contracts and open-source licensing. If you're shipping production code that an AI wrote, you should really read that one. Link in the briefing. This is one of those things that'll bite you two years from now if you ignore it. 13 00:01:18.367 --> 00:01:24.497 And then there's a pointed analysis arguing that AI's economics just don't make sense right now. The gap between infrastructure spending and actual revenue is… significant. The argument is that current AI pricing is subsidized and unsustainable. 14 00:01:24.497 --> 00:01:30.577 Which ties right back to the Copilot story. The free ride ends eventually. If you're making build-versus-buy decisions on AI, factor in that the APIs you're using today might cost two or three times as much in a year when the subsidies dry up. 15 00:01:30.577 --> 00:01:39.860 One fun one before we move on — there's a thirteen billion parameter language model called Talkie that was trained exclusively on pre-nineteen-thirty text. It sounds like a novelty, but the real takeaway is about domain-specific fine-tuning. Curated historical corpora produce genuinely distinct model behavior. Cool implications for education and creative writing tools. 16 00:01:39.860 --> 00:01:42.688 I kind of love that. Imagine a model that writes like it's 1920. But yeah, the technique is the real story there. 17 00:01:42.688 --> 00:01:49.193 Alright, dev tools. LibreChat is worth calling out — it's a self-hosted ChatGPT clone, open source, and it now supports GPT-5, MCP agent support, and model switching across basically every major provider. OpenAI, Anthropic, DeepSeek, Groq, Vertex, you name it. 18 00:01:49.193 --> 00:01:56.024 If you need an internal AI chat interface or you want to build customer-facing AI features that you actually own and control, this is probably the most complete open-source option out there right now. I've been seeing more teams reach for this instead of rolling their own. 19 00:01:56.024 --> 00:02:04.681 Also trending again — tldraw, the infinite canvas SDK. If you're building collaborative tools, visual editors, or AI-powered diagramming, it's the most mature open-source canvas primitive available. And there's a solid guide on high-performance Git for large repos — concrete techniques for when your clone and checkout times are getting painful. 20 00:02:04.681 --> 00:02:07.359 The Git one is surprisingly practical. Monorepo people, go read it. Links in the briefing for all of these. 21 00:02:07.359 --> 00:02:12.263 Quick security roundup. A tool called AISLE — which uses AI-assisted security auditing — found thirty-eight CVEs in OpenEMR, which is healthcare software used by over a hundred thousand providers. 22 00:02:12.263 --> 00:02:18.968 Thirty-eight! That's a lot of vulnerabilities in software that handles patient data. If you're in healthtech or integrating with OpenEMR, patch immediately. But the meta-story is that AI-assisted auditing is finding bugs at a scale that manual review just can't match. 23 00:02:18.968 --> 00:02:23.973 And in Toronto, three arrests for using SMS blasters to send mass phishing texts — first of their kind. If you're building messaging products, regulators are watching the SMS channel very closely now. 24 00:02:23.973 --> 00:02:25.424 Sender verification isn't optional anymore. Good reminder. 25 00:02:25.424 --> 00:02:31.679 Alright, rapid fire quick hits. The quiet resurgence of RF engineering — hardware skills are suddenly in demand again. There's an interactive color perception experiment trending on Hacker News asking 'Is my blue your blue?' which is just delightful. 26 00:02:31.679 --> 00:02:38.385 I spent way too long on that color one. Also, a prolific contributor has officially retired from Emacs — end of an era. And there's a great post called 'How I learned what a decoupling capacitor is for, the hard way.' Hardware people will feel that one in their bones. 27 00:02:38.385 --> 00:02:44.165 So to bring it all together — the GitHub Copilot billing change is the canary in the coal mine. AI dev tooling is moving from free trial to consumption-based pricing across the board. The era of 'just turn everything on' is ending. 28 00:02:44.165 --> 00:02:50.896 Be deliberate about which AI features actually earn their cost in your pipeline. Measure it, optimize it, and while you're at it, harden the CI surfaces you're now paying more for. Treat AI tool costs like compute costs — variable, measurable, and worth your attention. 29 00:02:50.896 --> 00:02:54.399 That's the briefing for April 29th. All the links and resources are in the show notes. Thanks for listening, folks — we'll see you tomorrow. 30 00:02:54.399 --> 00:02:55.1000 Go audit those Actions workflows before June. See you next time!