Fallthrough

Steve is back to talk JJ (Jujutsu version control) and the related product, ChangeSet, that he works on at East River Source Control. Kris and Steve trace why the GitHub monoculture is finally cracking, what JJ does that Git can't, and Steve's hypothesis that AI agents are pushing companies toward monorepos. Then the pair discuss the Opus 4.7 regression debate, the shift from "always use the frontier model" to using LLMs as one tool among many, and a quick discussion of the Tim Cook to John Ternus handoff at Apple.We've got supporter content, of course! This week that includes Steve's broader thesis that it's easier to scale a big tool down than scale a small one up, why GitHub's pull request model warped how people use Git, auto-rebase and conflicts as first-class citizens in JJ, AI subsidies, and a Bun being ported to Rust. Not a supporter yet? Fix that today by heading over to https://fallthrough.fm/subscribe where you'll get not only extra content but also higher quality audio. Sign up today!If you prefer to watch this episode, you can view it on YouTube.Thanks for tuning in and happy listening!Show Notes:JJTable of Contents:Prologue (00:00:00)Chapter 1: Episode 70, New Branding, and Catching Up (00:01:14)Chapter 2: What is JJ (Jujutsu) and Why It Exists (00:03:43)Chapter 3: Change IDs, Tangled, and the Federated Forge Wave (00:08:59)Chapter 4: Healthy Diversity After the GitHub Monoculture (00:14:33)Chapter 5: Scaling Big Tools Down: Bazel, Buck, Kubernetes, Symfony [Extended] (00:18:03)Chapter 6: JJ Is Simpler AND More Powerful: No Index, No Stash (00:18:10)Chapter 7: The JJ Workflow: Snapshots, Watchman, and JJ Undo (00:26:06)Chapter 8: GitHub Warped Git: Why Patches and Gerrit Are Better [Extended] (00:32:14)Chapter 9: Auto-Rebase and Conflicts as First-Class Citizens [Extended] (00:32:14)Chapter 10: Getting Started with JJ: Tutorials and Workflows (00:32:29)Chapter 11: East River Source Control and ChangeSet: A Forge for Mono-Repo Scale (00:42:28)Chapter 12: Why AI Forces Companies into Monorepo Scaling Sooner (00:50:01)Chapter 13: AI Subsidies and the Scaling Wall [Extended] (00:54:55)Chapter 14: The Opus 4.7 Regression Debate, Goodhart's Law, and Custom Styles (00:55:24)Chapter 15: LLMs as Tools, Not Solutions: Local Models, Gemini, and Custom Pipelines (01:04:49)Chapter 16: Bun Rewriting in Rust: A Branch That Wasn't Supposed to Be News [Extended] (01:15:48)Chapter 17: The Apple CEO Transition: Tim Cook to John Ternus (01:16:08)Epilogue (01:24:55)Hosts Kris Brandow - Host Steve Klabnik - Host Socials:WebsiteBlueskyThreadsX/TwitterLinkedInInstagramChangelog ZulipGophers Slack (00:00) - Prologue (01:14) - Chapter 1: Episode 70, New Branding, and Catching Up (03:43) - Chapter 2: What is JJ (Jujutsu) and Why It Exists (08:59) - Chapter 3: Change IDs, Tangled, and the Federated Forge Wave (14:33) - Chapter 4: Healthy Diversity After the GitHub Monoculture (18:03) - Chapter 5: Scaling Big Tools Down: Bazel, Buck, Kubernetes, Symfony [Extended] (18:10) - Chapter 6: JJ Is Simpler AND More Powerful: No Index, No Stash (26:06) - Chapter 7: The JJ Workflow: Snapshots, Watchman, and JJ Undo (32:14) - Chapter 8: GitHub Warped Git: Why Patches and Gerrit Are Better [Extended] (32:29) - Chapter 10: Getting Started with JJ: Tutorials and Workflows (42:28) - Chapter 11: East River Source Control and ChangeSet: A Forge for Mono-Repo Scale (50:01) - Chapter 12: Why AI Forces Companies into Monorepo Scaling Sooner (54:55) - Chapter 13: AI Subsidies and the Scaling Wall [Extended] (55:24) - Chapter 14: The Opus 4.7 Regression Debate, Goodhart's Law, and Custom Styles (01:04:49) - Chapter 15: LLMs as Tools, Not Solutions: Local Models, Gemini, and Custom Pipelines (01:15:48) - Chapter 16: Bun Rewriting in Rust: A Branch That Wasn't Supposed to Be News [Extended] (01:16:08) - Chapter 17: The Apple CEO Transition: Tim Cook to John Ternus (01:24:55) - Epilogue

Kris and Ian dig into the slow collapse of GitHub, starting with Ghostty off the platform after years of reliability problems. From there they trace Gary Bernhardt's old observation that we took a decentralized source control system and immediately put it behind a single point of failure, then widen the lens into AI as the engine of enshittification. The episode lands on a more optimistic note: maybe AI is also the tool that lets individuals rebuild the apps they used to have to buy.We've got supporter content, of course! This week that includes Kris's "AI was trained on the median of human code" argument, the institutional-memory thesis for why AI can't just replace people, a tour of the local-AI hardware landscape with Zig and llama.cpp, Ian's relatable "what is Cursor? what is Warp?" rant, a policy pitch on share-flipper voting rules, and a deep dive on typography as an anti-AI signal. Not a supporter yet? Fix that today by heading over to https://fallthrough.fm/subscribe where you'll get not only extra content but also higher quality audio. Sign up today!If you prefer to watch this episode, you can view it on YouTube.No episode of Break this week. We'll have more aftershow episodes soon! In the meantime, catch up on previous episodes at https://break.show.Thanks for tuning in and happy listening!Table of Contents:Prologue (00:00:00)Chapter 1: GitHub Is Falling Over: Ghostty Leaves and the Merge Bug (00:01:46)Chapter 2: Centralizing the Decentralized: Git, Forges, and the Linux Kernel Workflow (00:08:04)Chapter 3: Self-Hosting vs. The Cloud Tax (00:16:23)Chapter 6: AI as the Engine of Enshittification (00:32:02)Chapter 9: Epistemic Enshittification: Hedges, Hype, and Sci-Fi Brain (00:37:10)Chapter 10: Self-Checkout, Junior Devs, and Broken Hiring (00:46:52)Chapter 12: Fallthrough Rebrand and the Claude Design Experiment (00:54:16)Chapter 13: Typography as Anti-AI Signal: Buying Real Typefaces [Extended] (01:04:33)Chapter 14: Bleeps, Shorts, and Producing With AI (01:04:36)Chapter 15: AI as the De-Enshittification Tool (01:08:08)Epilogue (01:10:22)Hosts Kris Brandow - Host Ian Wester-Lopshire - Host Socials:WebsiteBlueskyThreadsX/TwitterLinkedInInstagram (00:00) - Prologue (01:46) - Chapter 1: GitHub Is Falling Over: Ghostty Leaves and the Merge Bug (08:04) - Chapter 2: Centralizing the Decentralized: Git, Forges, and the Linux Kernel Workflow (16:23) - Chapter 3: Self-Hosting vs. The Cloud Tax (32:02) - Chapter 4: AI Code Quality and the Mediocre Mean [Extended] (37:10) - Chapter 9: Epistemic Enshittification: Hedges, Hype, and Sci-Fi Brain (46:52) - Chapter 10: Self-Checkout, Junior Devs, and Broken Hiring (54:16) - Chapter 12: Fallthrough Rebrand and the Claude Design Experiment (01:04:33) - Chapter 13: Typography as Anti-AI Signal: Buying Real Typefaces [Extended] (01:04:36) - Chapter 14: Bleeps, Shorts, and Producing With AI (01:08:08) - Chapter 15: AI as the De-Enshittification Tool (01:10:22) - Epilogue

Jamie returns, co-hosting with Kris for quite an eventful episode. They start with the Vercel breach, the Axios attack, nvim-treesitter, and Gorilla Mux. Kris draws parallels between the current AI hype cycle to everything that came before: Photoshop was going to destroy photographers, DAWs were going to destroy musicians, and now Claude Code is going to destroy software engineers.Just like the last episode, this one is filled with supporter only content. It's actually an extra episode and a half! This includes a deep dive on OAuth scope design and why consent screens need the Let's Encrypt treatment, Anthropic locking out third-party harnesses, sandboxing LLMs with sandbox-exec and agent-safehouse, Jamie on "no is a complete sentence" and Renovate's unusually pro-maintainer code of conduct, Kris's pitch for $100M of LLM spend going to help maintainers triage backlogs instead of Mythos-style vulnerability hunting, the real complexity behind finance departments and why $250M wire transfers need entire treasury teams, the math on a $800 used 3090 running Qwen 3.5 at Sonnet-level capability, and Anthropic's pivot toward non-engineers with Claude Design and Claude Cowork. Not a supporter yet? Fix that today by heading over to https://fallthrough.fm/subscribe where you'll get not only extra content but also higher quality audio. Sign up today!If you prefer to watch this episode, you can view it on YouTube.No episode of Break this week. We'll have more aftershow episodes soon! In the meantime, catch up on previous episodes at https://break.show.Thanks for tuning in and happy listening!Table of Contents:Prologue (00:00:00)Chapter 1: Eventful Spring, Banter and Setup (00:01:20)Chapter 2: The Vercel Breach: Roblox to Production Via One Employee (00:04:25)Chapter 6: Deepfakes, the Axios Attack, and the $5 Wrench (00:06:49)Chapter 7: The XZ Utils Attack and nvim-treesitter Gets Archived (00:11:22)Chapter 8: GorillaMux, the Go Standard Library, and Drive-By Forkers (00:14:13)Chapter 11: The Hype Cycle: Opus Got Dumber, and Every Tool Was Supposed to Replace Us (00:20:44)Chapter 13: Claude Code Going Max-Only, Copilot Pro Losing Opus (00:30:56)Chapter 16: Kris the Former Hater, and the Slop Zone [Extended] (00:34:12)Chapter 17: Deep Blue, Collaboration, and Why LLMs Aren't Replacing Us (00:38:33)Epilogue (00:44:03)Hosts Kris Brandow - Host Jamie Tanna - Host Socials:WebsiteBlueskyThreadsX/TwitterLinkedInInstagram (00:00) - Prologue (01:20) - Chapter 1: Eventful Spring, Banter and Setup (04:25) - Chapter 2: The Vercel Breach: Roblox to Production Via One Employee (06:49) - Chapter 6: Deepfakes, the Axios Attack, and the $5 Wrench (11:22) - Chapter 7: The XZ Utils Attack and nvim-treesitter Gets Archived (14:13) - Chapter 8: GorillaMux, the Go Standard Library, and Drive-By Forkers (20:44) - Chapter 11: The Hype Cycle: Opus Got Dumber, and Every Tool Was Supposed to Replace Us (30:56) - Chapter 13: Claude Code Going Max-Only, Copilot Pro Losing Opus (34:12) - Chapter 16: Kris the Former Hater, and the Slop Zone [Extended] (38:33) - Chapter 17: Deep Blue, Collaboration, and Why LLMs Aren't Replacing Us (44:03) - Epilogue

They dissect post-Mythos hype and counter-narratives about access and tokenized models. The conversation shows why existing supply-chain weaknesses already enable many attacks. They argue for signatures, layered attestations, cooldowns and better tooling UX over forking or vendoring. The show ends with an infinite-mindset view that supply-chain security is ongoing work.

They unpack an AI model that uncovered dozens of deep zero-days and a 27-year-old OpenBSD bug. They wrestle with what widespread exploit generation means for supply chain security and memory-safe languages. They debate cooldowns for package upgrades, recall large-scale vulnerability remediation stories, and explore AI as a personal assistant and scalable source-control ideas.

They debate a leaked Claude code dump and whether agent code should be open. They warn about AI hype repeating social media’s mistakes and question LLMs’ lofty claims. They pivot into finance: why engineers should learn taxes and bookkeeping, and how double-entry accounting is an original distributed system. Short rants about Mac choices, corporate culture, and who really does the work round out the conversation.

A lively panel probes fears around superintelligence, tracing them to colonial and alien-anxiety narratives. They debate whether AI is already oppressive and how algorithms subtly shepherd behavior. The conversation examines hierarchy, collective power, leadership failures as real risks, and whether intermediate smarter-but-flawed systems pose unique dangers.

They nerd out over migrating NeoVim to Lua, swapping shells to Nushell, and using LLMs to clean config cruft. Terminal emulators, LSP and TreeSitter get compared alongside custom keybinding strategies. Hardware chatter ranges from Framework desktops and Mac Studios to 10G networking, NAS setups, and headless GPU discoveries. They also cover Wi Fi with per‑device keys and tinkering with home lab routers.

A lively breakdown of a proposed UUID API for Go and why it sparked strong debate. They dissect survey flaws, API design problems, and whether GitHub is even fit for technical deliberation. The conversation widens to community dynamics, enforcement double standards, and practical paths forward for shipping or experimenting with standards.

A spirited debate about Go's identity as a rebellious, simple language and when it should rethink fundamentals. A call for the community to stop relying on the language team and build a community-owned ecosystem. A provocative proposal to deprecate the error interface and treat errors as typed return values with multiple returns.