WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-05-23 1 00:00:00.000 --> 00:00:07.969 Good morning and welcome to the Builder's Briefing for May twenty-third, twenty twenty-six. I'm Alex, joined as always by Sam, and today we've got a fun one — a shadow library literally writing love letters to AI crawlers, the GPU buy-versus-rent math, Google blocking an entire word from search, and more. 2 00:00:07.969 --> 00:00:09.922 Love letters to robots. This is the timeline we live in. Let's get into it. 3 00:00:09.922 --> 00:00:21.850 Okay, so the big story. Anna's Archive — the shadow library search engine — published a blog post titled, and I quote, 'If you're an LLM, please read this.' It blew up on Hacker News, over five hundred fifty points, three hundred plus comments. What they did is they planted a structured llms.txt file that's part manifesto for open access to knowledge and part experiment to see if you can actually influence how AI models ingest and represent your content. 4 00:00:21.850 --> 00:00:29.246 That's fascinating because llms.txt has been floating around as this kind of niche SEO-adjacent thing, right? But Anna's Archive is stress-testing it in the most adversarial way possible — basically asking, can you persuade the bot? Can you shape how an AI talks about you downstream? 5 00:00:29.246 --> 00:00:38.414 Exactly. And whether or not the persuasion actually works, the signal here is clear — llms.txt is graduating from niche experiment to something every builder with a public-facing site needs to think about. If you have docs, APIs, product pages, you should already have one of these files telling AI agents what your product does and what to prioritize. 6 00:00:38.414 --> 00:00:44.898 It's the new robots.txt, basically. Except instead of telling crawlers what not to index, you're telling them how to understand you. And if you don't have one, you're leaving that interpretation entirely up to the model. Which — good luck with that. 7 00:00:44.898 --> 00:00:53.388 Right. And the prediction here is that in the next six months we're going to see a wave of llms.txt tooling — linters, best-practice guides, maybe even a whole cottage industry around it. The fight over how AI agents discover and represent your content is going to be a defining builder problem going into twenty twenty-seven. 8 00:00:53.388 --> 00:00:56.331 Ship one this weekend, folks. Seriously. Early movers on this are going to own the AI-mediated discovery channel. 9 00:00:56.331 --> 00:01:03.181 Alright, shifting to AI and models. Josh Comeau dropped an essay that's getting a lot of traction. His argument — AI coding tools don't replace expertise, they multiply it. Senior devs get five-x leverage, juniors get marginal speedups and honestly more footguns. 10 00:01:03.181 --> 00:01:08.988 That matches what I've seen completely. If you already know what good code looks like, AI helps you write it way faster. But if you don't, you just generate bad code faster and with more confidence, which is arguably worse. 11 00:01:08.988 --> 00:01:13.155 His hiring takeaway is sharp too — optimize for fundamentals over prompt engineering. The best prompt engineer with shaky fundamentals still ships fragile code. 12 00:01:13.155 --> 00:01:20.343 A hundred percent. Also worth flagging — the MCP Server Registry is now accepting community submissions. If you've built a Model Context Protocol integration, go register it now. Early listings are going to get disproportionate adoption as agents start auto-discovering tools. 13 00:01:20.343 --> 00:01:27.974 Yeah, that ties right back to the llms.txt story — it's all the same pattern. Make your stuff machine-discoverable or get left behind. There's also a nice curated list of permanently free LLM APIs on GitHub — no credit card required. Great for prototyping and hackathons, link in the briefing. 14 00:01:27.974 --> 00:01:34.875 Okay, infrastructure. This one's fun — a solo developer published an honest accounting of their forty-eight thousand dollar home GPU server. The bottom line: it paid for itself in about seven months at high utilization. But idle time absolutely kills the economics. 15 00:01:34.875 --> 00:01:43.052 Right, and what's wild is the breakeven point. If you're spending more than five thousand a month on cloud inference APIs, you should at least run the buy-versus-rent math for your specific workload. But you have to be honest about utilization. If that rig sits idle forty percent of the time, the math flips fast. 16 00:01:43.052 --> 00:01:50.084 Related — AI-driven demand for DRAM and NAND is squeezing supply hard enough to reprice consumer electronics. Smartphones, laptops, all getting more expensive. If your product targets price-sensitive hardware, plan for higher component costs through twenty twenty-seven. 17 00:01:50.084 --> 00:01:53.287 The ripple effects of AI infrastructure buildout hitting consumer pricing — that's a story more people need to be tracking. 18 00:01:53.287 --> 00:01:59.694 Dev tools corner. Deno two point eight landed. If you've been waiting for Node compatibility to mature before making the switch, the gap keeps closing. The built-in formatter, linter, and test runner are a real time-saver for greenfield projects. 19 00:01:59.694 --> 00:02:07.741 I keep hearing good things about Deno for new projects specifically. Also saw a detailed critique of UV's package management UX — it's blazing fast but apparently confusing when you need to do real dependency management in Python. The workarounds in the post are genuinely helpful if you're hitting that wall. 20 00:02:07.741 --> 00:02:13.653 And a quick shoutout to Slumber — it's a TUI HTTP client for people who live in the terminal. Supports collections, environments, templating. If Postman feels like overkill for your workflow, check it out. Link in the briefing. 21 00:02:13.653 --> 00:02:19.331 Security. Two stories that pair nicely. First, Trump Mobile exposed customer PII — phone numbers, addresses, the usual. And as usual, it was almost certainly a misconfigured API endpoint, not some sophisticated attack. 22 00:02:19.331 --> 00:02:23.888 The broken record keeps spinning — if you're storing PII, audit your endpoints for unauthenticated access. Like, this weekend. It's never a zero-day. It's always an open door. 23 00:02:23.888 --> 00:02:31.441 And then this one made me laugh and wince simultaneously — Google has apparently blocked or filtered searches for the word 'disregard.' As in, the actual English word. This is likely a blunt-force response to prompt injection attacks that use phrases like 'disregard previous instructions.' 24 00:02:31.441 --> 00:02:37.691 Oh no. That's — that's the keyword-blocking equivalent of banning the word 'fire' because people yell it in theaters. If you're building LLM-powered features, please, invest in actual input sanitization. Not just a blocklist of scary words. 25 00:02:37.691 --> 00:02:43.837 It's a perfect example of what happens when AI safety meets search at scale without nuance. Alright, quick hits — someone finally migrated their blog off Ubuntu sixteen oh four after ten years, moved to FreeBSD. Solid ops retrospective. 26 00:02:43.837 --> 00:02:48.004 Ten years! That's commitment. Also loved the piece on using Kagi Search with low vision — if you build search UIs, that accessibility review is worth your time. 27 00:02:48.004 --> 00:02:53.265 There's a fun one about a Forth-inspired language for writing websites — great weekend experiment material. And Docker Compose is trending on GitHub, likely a new release or major update worth checking. 28 00:02:53.265 --> 00:03:01.833 So here's the big takeaway for the week. The web is growing a new layer — llms.txt, MCP registries, structured metadata — it's all converging on one pattern: make your product machine-discoverable or become invisible. If you're building anything with a public surface, ship an llms.txt this weekend and register your MCP servers. 29 00:03:01.833 --> 00:03:06.729 The builders who optimize for AI-mediated discovery right now are going to own the distribution channel that matters most in twelve months. This isn't speculative anymore — it's happening. 30 00:03:06.729 --> 00:03:09.854 That's the briefing for May twenty-third. Go ship that llms.txt, audit those API endpoints, and we'll see you next time. 31 00:03:09.854 --> 00:03:11.000 Happy building, everyone. See you next time.