Rex sits down with Garrett in person in Denver for a wide-ranging conversation that starts at ETHDenver and ends at the future of AI agents as economic actors. They unpack the mood on the ground at a smaller, more subdued ETHDenver, debate whether Ethereum’s rollup-centric roadmap is being corrected or abandoned, and discuss why stablecoins may still be Ethereum’s clearest product-market fit. From there, the conversation pivots into AI: agentic coding, always-on tools like OpenClaw, what it means to build products for agents (not just with AI), and how financial rails may become essential infrastructure for machine users. Thoughtful, skeptical, and optimistic in equal measure.

This week’s episode is a little different: instead of an interview, I’m reading an essay I wrote called “Sitting at a Slot Machine,” a candid story about falling headfirst into “vibe coding,” building an increasingly elaborate AI Context System, and realizing that unlimited execution can be just as dangerous as it is exhilarating. I unpack the dopamine loop of constant progress, the slow creep of complexity disguised as productivity, and the hard pivot back to simplicity... where the real craft becomes pruning, not piling on, and a few tight, well-chosen habits beat any sprawling framework.

Crypto’s melting down, so Rex sits down with Dan and Gerrit for AI Tools: Round Three—a conversation about what’s actually changing in day-to-day work when models ship, agents run in parallel, and “sessions” start to feel like a lifestyle.

They react to a big model-release day (Opus 4.6 + ChatGPT 5.3), compare Claude Code vs Codex for real coding work, and unpack why Claude feels so sticky: better UX, more glazing, and a dopamine-loop quality that’s hard to ignore once you notice it.

From there it gets practical: managing context windows with dashboards and handoff files, building bespoke internal tools (like Rex’s “notification hub”), and watching weird new ecosystems form — Moltbook-style bot social networks, and even “Tinder for bots.”

Are we building useful developer tools—or just feeding an addiction?

In the second episode of our AI tools series, Rex, Dan, and Garrett dig into the internal systems they've built around Claude Code: markdown session logs, pre-commit hooks, context management strategies, and notification hubs. Dan shares war stories from running AI on helicopters inspecting power lines (including the time an agent changed his root password without asking). Garrett walks through his approach to scaling ten concurrent projects. And Rex asks the uncomfortable question: is all this meta-tooling actually helping, or are we just tinkering because it feels productive?

The conversation moves to local LLMs—Qwen, DeepSeek, Mistral—and whether a $15-20k home lab makes sense when Claude iterates faster than anyone can keep up. The hosts wrestle with context window limits, the ROI of refactoring, and what it means that these tools are specifically designed to make you feel like you're accomplishing more than you are.

Rex digs into how AI coding agents are changing the way real teams build software. Taylor Savage and Gerrit Hall share the workflows, tools, and guardrails that actually work: from custom IDEs and inbox-driven agents to multi-pass code reviews, collaboration at “agent speed,” testing that matters, and keeping data safe. A grounded look at what to automate, what to supervise, and how to ship faster without losing the plot.

Rex brings on his longtime friend Taylor Savage — a lifelong “techie” turned software engineer and product manager — for a candid, outside-the-crypto-bubble conversation about what crypto actually looks like from the broader tech world. From an early Bitcoin birthday gift (and an early sale…) to the last decade of fraud “speed-runs,” they debate irreversibility, regulation, institutional adoption, and whether crypto’s real promise is narrow-but-real or mostly drowned out by hype and extraction.

Rex sits down with Sam McCulloch (Growth Lead at USD.ai, formerly Leviathan News and Flywheel DeFi) to talk about what crypto’s been optimizing for—and what it should optimize for next. Sam argues the last couple of cycles rewarded extractive “attention finance” (meme coins, short-horizon speculation), and that the industry needs to reclaim a more constructive narrative: build real products, connect to real-world balance sheets, and make the rails useful outside of crypto itself.

In the back half, they go deep on USD.ai’s thesis: bringing DeFi-style, programmatic lending to GPU infrastructure. Sam walks through how tokenized “title” to GPUs can stay productive inside data centers while changing hands—using familiar commercial-law concepts like warehousing and documents of title—so liquidations look more like transferring ownership than physically moving hardware.

Based rollups have been floating around Ethereum discourse for a couple years now, but the concept remains confusing even to people who understand rollups generally. The pitch sounds almost contradictory: use Ethereum's decentralized validator set to sequence rollups, but somehow still get the fast, smooth UX of centralized sequencers like Base or Arbitrum.

Jason Vranek works on Fabric, building the infrastructure to make based rollups actually work. In this episode, he walks through the architecture piece by piece: what "based sequencing" actually means, why the original "total anarchy" designs had terrible UX, and how pre-confirmations and gateways solve that without re-centralizing everything.

The conversation gets into the PBS (proposer-builder separation) analogy—just as builders abstract away MEV sophistication from validators, gateways can abstract away pre-confirmation complexity. Jason explains the transaction flow from user to blob, how proposer commitments enable coordination without requiring validators to become sophisticated, and why the real unlock isn't necessarily replacing centralized rollups, but enabling L1-to-L2 composability that doesn't exist today.

They also discuss where things actually stand: Tycho is live on mainnet with a rotating sequencer set, the Fusaka hard fork just shipped a critical EIP for deterministic lookahead, Fabric is deploying a universal registry contract for proposer collateral, and real-time proving has gone from "years away" to "usable building block." Plus: why the end state probably isn't "all rollups become based," but rather a spectrum of designs that pick their tradeoffs deliberately.