Intelligence Snacks

What comes after AI chat? Pete and Andy explore why traditional chat interfaces struggle as AI agents take on more work, and how visual workspaces, interactive canvases and personalised interfaces could offer more natural ways to think, plan and collaborate with agents.

Pete shares his latest experiments with Ambulando, building a spatial canvas where documents, task lists, whiteboards and small applications sit alongside each other, connected to the same underlying information. The discussion turns to how AI is changing the economics of work. When agents can build something faster than humans can read its specification, traditional approaches to planning, documentation and project management start to make less sense.

They also explore how individuals might develop their own ways of working with AI, from voice-driven agents to visual planning environments, and what this means for businesses. Could companies become dramatically smaller as agents take on more execution? And if automated production becomes available to everyone, do human judgment, originality and trusted relationships become more valuable?

Key Moments:

• [01:40] Pete introduces visual interfaces and the challenge of managing multiple AI agents.

• [03:41] From terminal sessions to Autopilot and Ambulando, evolving how agents are managed.

• [08:30] Experimenting with spatial canvases and small applications called naplets.

• [11:06] Combining structured data, whiteboards and task lists in one visual workspace.

• [13:37] Bringing live documents and project information into a canvas through Nostr Stack.

• [17:01] Andy on the limitations of chat and his experiments with agent-accessible visual planning.

• [19:48] Exploring interface designs, comparing versions and editing directly on a canvas.

• [23:00] Why AI-generated documentation is becoming harder for humans to keep up with.

• [25:00] Which documents still matter when building is quicker than specifying?

• [28:13] How AI is changing the timescales behind planning and software development.

• [31:11] Why visual schematics can make complex systems easier to understand.

• [38:43] Temporary workspaces and the challenge of managing multiple projects.

• [40:44] Turning comments into conversations with agents directly inside a canvas.

• [43:51] One set of information, multiple interfaces, including voice-driven agents.

• [46:00] Why long-running chat threads become difficult to manage.

• [49:02] Personalised AI workflows versus standardised enterprise software.

• [51:30] Automating photo editing through Claude without changing familiar workflows.

• [54:58] Should organisations prescribe how people work or focus on outcomes?

• [57:41] Different interfaces for dispatching, discussion and strategic planning.

• [59:55] Building prototypes as a way of thinking through ideas.

• [01:02:33] How multiple humans collaborate when everyone has their own AI agents.

• [01:04:42] Could AI make four- or five-person companies the optimal size?

• [01:08:34] Why automated marketing production may lose value as it becomes ubiquitous.

• [01:11:32] Authenticity, originality and trust in an internet filled with AI-generated content.

• [01:13:15] Kevin Kelly's 1,000 True Fans and the value of established relationships.

Projects and ideas mentioned: Ambulando, Autopilot, Wingman, Flight Deck, Nostr Stack, naplets, Rick's Artifacts, Bitcoin Bush Bash, Excalidraw, Miro, Claude, ChatGPT, Lightcraft, Matt Pocock's Grill Me and Wayfinder, Pieter Levels and Kevin Kelly's 1,000 True Fans.

Find out more: ambulando.io, intelligencesnacks.com

Closing thought: "Maybe this is what I was building to all along." — Pete

Pete and Andy welcome Skyler from Finite.Computer to discuss building agents that give people more room for human work: exercising judgment, developing taste, finding meaningful problems and connecting with others.

Skyler introduces Finite.Computer, the company he is building with Paul Miller, and its hands-on agent camps for freedom-focused organisations.

He describes his shift from design work to building AI products, and why taste still matters when software becomes cheap to create. Knowing what is good, explaining why and deciding whether something should exist remain skills worth practising.It then takes a philosophical turn into embodiment, faith, beauty and what gives work meaning.

“Humanmaxing” becomes the organising idea: let computers handle the digital drudgery while people lean into the experiences and relationships that make them human.

**Key Moments:**

- [00:05] Welcome to Episode 76 with special guest Skyler.

- [01:13] Introducing Finite.Computer: agents built toward sovereignty, paired with in-person training.

- [03:44] Early agent camps, working with Paul Miller and turning the training into a product.

- [06:04] Skyler’s move from client design work to experimenting with AI and building a company.

- [07:58] Defining taste: knowing what is good or bad and being able to explain your reasons.

- [10:55] Novelty, differentiation and why good taste also means understanding the context.

- [13:33] Enterprise software and the economics that pushed products toward standardisation.

- [18:55] Could an agent become a protective barrier between you and distracting interfaces?

- [21:10] Pete introduces Ambulando: managing agents while you go out walking.

- [24:03] Reviewing proposed changes through live visual prototypes and comments before implementation.

- [26:00] Andy on computer use, personal preferences and turning repeatable work into code.

- [29:27] Beyond the builder bubble: does everyone actually want to make their own software?

- [31:30] Pete’s case for a more local software industry with accessible providers.

- [33:16] Finite’s human approach: training, shared meals and helping people work through ideas.

- [35:11] Turning an HRF presentation into a shared canvas where comments become agent-made edits.

- [41:27] Real-world experience and community help people discover which problems need solving.

- [43:09] Teaching design thinking and the five whys alongside agent skills.

- [46:31] Skyler questions a picture of humanity centred entirely on intelligence.

- [51:11] Cathedrals, beauty and building for something beyond immediate economic rewards.

- [52:39] How everyday practices shape what people value and desire.

- [57:48] Could AI help bring back the individuality and community of the old high street?

- [59:50] Humanmaxing: freeing people from digital drudgery and leaning into human abilities.

- [01:02:52] Why ownership matters when a business becomes software and agents.

- [01:04:44] Open-weight models for everyday work and the pull of frontier capabilities.

- [01:06:15] Andy makes the case for raising ambition alongside more capable models.

- [01:09:53] Pete introduces Nostr Stack and using a shared identity across applications.

- [01:11:42] Authentication, key management, backups and recovery remain difficult design problems.

- [01:15:31] Where to find Finite.Computer and a preview of its shared organisational apps.

**Projects and ideas mentioned:**

Finite.Computer, agent camps, Human Rights Foundation, Paul Miller, Ambulando.io, Rick’s Artifacts, Nostr Stack, Tinfoil, open-weight models, the five whys, Bitcoin Bush Bash and *Desiring the Kingdom* by James K. A. Smith.

**Find out more:**

finite.computer, ambulando.io

**Closing thought:**

“Lean into your unique abilities as a human and let the computer do the computer work.” — Skyler

If AI can imitate a person on a video call, what happens to trust online? Pete and Andy discuss the promise of an infinitely patient language tutor alongside the obvious opportunities for scammers. That leads to a broader question about the dead internet - as public feeds fill with bots and disposable content, could more of our useful online lives move into group chats, trusted communities and real-world connections?

Andy introduces Sally, a small MacBook notch app that lets him post and schedule tweets without opening the feed. Pete shares his own work on shipping Wingman builds and finding better ways to direct agents. Editable flow diagrams, whiteboards and small custom editors offer ways to review work, show what needs to change and catch misunderstandings before implementation.

The conversation turns to how much documentation fast-moving work actually needs. Pete and Andy weigh lasting reference material against plans that become stale within an hour, and discuss why quicker iterations can help people make better decisions.

They finish with Andy’s experiment using Jev to review AI writing: flagging abstraction, unnecessary language and signposting so the drafting agent can improve its writing quality.

**Key Moments:**

- [00:02] Welcome to Episode 75, recorded back at the beach.

- [01:25] A human interaction model prompts two immediate thoughts: personal tuition and convincing scams.

- [03:00] Dead internet theory, public feeds and the value of connecting with people you trust.

- [03:23] Nostr and the difference between a community of connections and a global megaphone.

- [06:16] The retreat into group chats and moderated spaces where people can reject unwanted AI content.

- [07:28] AI could amplify both productive work and passive consumption.

- [08:21] The appeal of experiences where conversations are neither recorded nor broadcast.

- [09:51] Andy introduces Sally: posting to X from his MacBook notch without entering the feed.

- [13:51] Rescheduling posts across time zones and being more deliberate about social media use.

- [15:56] Exercising the shipping muscle; Pete describes automated Wingman builds for TestFlight and Zapstore.

- [18:36] Rick’s Artifacts becomes a place to build small tools for editing, comparing versions and reviewing work.

- [20:32] Flow diagrams make components, actions and changes of state easier to inspect.

- [24:47] Agree on the current state, draw the future state, then give the agent a clear implementation target.

- [28:31] Context trees and wiki-style documents help agents find relevant projects, processes and references.

- [33:07] How much documentation do you still need when building and changing software becomes cheap?

- [37:38] Working as an individual with agents, and why visible wiki links can be easier to trust than a complex graph.

- [40:50] Turning exploratory conversations into focused implementation threads across separate systems.

- [43:58] Longer agent runs can weaken the feedback loop; Pete wants to see what the finished change will do.

- [47:08] Andy experiments with Jev as a separate reviewer for the Intelligence Snacks writing pipeline.

- [51:29] Specific review questions and confidence scores help identify passages that need editing.

- [52:58] Could generating several drafts and comparing them produce better writing?

- [54:24] The aim for the website and newsletter: summaries that are accurate and enjoyable to read.

**Projects and ideas mentioned:**

Sally, X, Nostr, Wingman, Flight Deck, Tower, Autopilot, Rick’s Artifacts, TestFlight, Zapstore, Obsidian, visual flow diagrams, context trees and Jev.

**Closing thought:**

“What matters more is that you can connect to people, not necessarily that you are connected to everyone.” — Pete

Intelligence Snacks 74 Normie Mode

AI model rankings are easy to find. Knowing which model to use for a spreadsheet, a design task or a language lesson is harder. Andy introduces Normie Mode, a new site that turns benchmark data into task-based recommendations that weigh capability against cost. It is an early release, and Andy wants feedback on whether the comparisons are useful to people choosing a model for everyday work.

Pete and Andy then ask where agents actually help. A story about an agent negotiating a marketplace sale and sharing its owner’s home address shows what can happen when an assistant has too much freedom. They make the case for letting people choose the outcome they want, then handing agents the tedious steps: completing a booking, finding a meeting time or checking an unwanted subscription.

The conversation moves from chat and spreadsheets to documents, custom software and changing business models. It closes with a debate about buying traditional businesses to automate them. Pete and Andy see an opportunity, but argue that industry knowledge, staff relationships and the slow work of change cannot be skipped.

Key Moments:

• [00:06] Welcome to Episode 74 and the unusually early choice of a title.
• [00:27] Andy introduces Normie Mode: model comparisons built around jobs people actually do.
• [02:24] Why a useful recommendation has to account for cost as well as capability.
• [04:58] How the site combines public comparison data with its own task-based tests—and where the early data has gaps.
• [07:50] Could agents use Normie Mode to choose models for their own work?
• [09:40] The growing gap between rapid AI progress and what everyday users can reasonably keep up with.
• [11:27] Why a harness tailored to your work can matter as much as the model inside it.
• [13:43] A marketplace agent accepts a low offer and shares its owner’s address.
• [15:17] The case for keeping human contact and judgment in negotiations and plans.
• [19:29] Let an agent handle the administrative steps after you have made the important choice.
• [25:33] Documents and slides become first-class AI workspaces; Pete argues for editing and commenting directly on the work.
• [29:35] Spreadsheets as an accessible programming tool—and whether custom software will replace some of their uses.
• [37:15] Personal agents could expose subscription businesses that rely on customers forgetting to cancel.
• [42:58] When AI makes polished promotional video cheap, does its value rise or fall?
• [47:31] Andy considers the advantage of creators who have spent years building audience trust.
• [50:50] The pitch: buy an established business, automate its workflows and capture the margin.
• [53:36] The missing detail: industry knowledge, key people and the difficulty of changing a business already in motion.
• [59:43] Buy or build? The opportunity is real, but patience and listening to operators matter.
• [01:04:30] Pete’s preference for owners and staff using AI to solve the problems they know firsthand.

Projects and ideas mentioned: Normie Mode, Wingman, Flight Deck, personal agents, AI documents, spreadsheets and AI-assisted business acquisitions.

Closing thought: “As always I think AI just rewards agency.” — Andy

Pete and Andy unpack a week where AI models seemed to multiply overnight: Anthropic updates, OpenAI’s cheaper models, Meta’s Muse, Grokbot, and the more interesting arrival of Jev.

The big question is whether we need ever-bigger frontier models, or whether the real unlock is cheaper, faster, more specialised intelligence that can be sprinkled through workflows.

The centre of the episode is Jev: a “system one” decision model built for fast classification, confidence scoring, routing and prioritisation.

Pete and Andy explore why that matters for business processes, GraphRAG, tool calling, coding agents, fraud checks, refunds, traffic systems, learning graphs and app economics. Then they zoom out to consumer agents: Meta, WhatsApp, network effects, ad businesses, personal automation and whether people will hand their lives to Zuckerberg or choose to own their own agents.

**Key Moments:**

- [00:06] Welcome to episode 73, now with faces on YouTube.

- [01:20] A big week in model land: Anthropic, OpenAI, pricing cuts and model-name weirdness.

- [02:58] Are new models genuinely better, or are we comparing them against nerfed old ones?

- [05:29] Pete argues for consistency, speed and cost over “more intelligence.”

- [06:26] Jev enters the chat as the week’s most interesting model release.

- [06:52] Andy tests Jev through OpenRouter on existing workflows.

- [08:16] SEO signals, prioritisation and replacing vibe-based confidence scores.

- [10:06] RLHF versus reinforcement learning via calibrated decisions.

- [12:27] Why structured JSON decisions matter for automation.

- [15:46] GraphRAG, context selection and cutting minutes of agent prep down to seconds.

- [16:40] Refunds, fraud checks and business workflows where confidence changes the routing.

- [18:28] Human escalation, graph traces and feedback loops for better decisions.

- [19:33] The economics: why frontier models are overkill for simple sorting and routing.

- [21:57] Jev as automation-native, not chat-native.

- [23:45] Browser use is still painful when every mouse move costs time and tokens.

- [24:08] Why Jev makes old ideas like Beacon feel interesting again.

- [25:44] Open-source Jev-style models and the Intelligence Snacks thesis coming back around.

- [27:35] IRL systems: traffic lights, waste, energy and transit routing.

- [32:22] What happens to small language models when decision models exist?

- [35:12] Intelligence Snack primitives: classification, prioritisation, translation and research.

- [38:50] Jevons paradox and near-free judgment everywhere.

- [41:47] “Fast is good”: why latency may matter more than bigger benchmarks.

- [43:05] Learning graphs for codebases, agents and human understanding.

- [47:29] Tool calling, agent loops and where expensive frontier reasoning is wasted.

- [50:13] Code graphs, agent context and asking better internal questions.

- [52:52] Meta Muse, Grokbot and the mainstreaming of personal agents.

- [56:13] Agents versus consumer inertia: insurance, subscriptions and renegotiation.

- [59:46] Do network effects beat capability?

- [01:02:30] The Meta reservation: agents inside an ad business.

- [01:04:43] Torrenting, distribution failures and extractive business models.

- [01:08:27] Why own your agent instead of giving everything to Meta or Elon?

- [01:12:21] Email is not really about communication; it’s about audit trails and accountability.

**Friends of the Pod:**

Jared Grigg, John, Justin Moon, Paul

**Projects Mentioned:**

Jev, TypeSafe, OpenRouter, Anthropic, OpenAI, Astra, Fable, Meta Muse, Grokbot, Instinct, Hugging Face, Beacon, GraphRAG, Autopilot, WhatsApp, Instagram, Amazon, Netflix

**Concepts:**

Decision models, system one models, calibrated confidence, RLHF, RLCD, model routing, tool calling, GraphRAG, small language models, Jevons paradox, local agents, consumer inertia, ad-driven AI

**Quote:**

“What you actually want out of these things is very, very quick inference.” — Pete

Pete and Andy dig into FIPS, the Free Internetworking Peering Service, and why it could make the internet feel personal again. Instead of treating your laptop and phone as viewers of cloud servers, FIPS makes devices, apps and agents first-class participants in a private, addressable network.

Then the conversation pivots into AI doom, open-source regulation and whether “AI will kill us all” is really a safety argument, or just another attempt by big institutions to keep power centralized.

**Key Moments:**

- [00:12] Welcome to episode 72, from the beach.

- [01:09] A listener request from John turns into the FIPS explainer episode.

- [02:12] FIPS as networking tech: reconnecting computers instead of turning them into cloud viewers.

- [04:03] Why today’s internet pushes people into domains, tunnels, accounts and API keys.

- [07:07] FIPS addresses, Nostr-style identity and private device networking.

- [08:07] Self-organising mesh networks and transport-agnostic routing.

- [11:10] Why FIPS gets interesting when complex networking becomes invisible to users.

- [11:56] The Google Drive problem: why upload files you already own locally?

- [14:14] A demo of browsing files from one computer on another.

- [16:30] Andy sees the agent angle: “a game changer for agents.”

- [19:14] Flight Deck, Tower and Autopilot as a real Wingman use case.

- [21:41] FIPS-addressing apps so they can run locally and be reached from anywhere.

- [23:57] The Wingman app browser on phone, laptop and desktop.

- [25:43] Scan a QR code instead of buying domains and wiring up Cloudflare.

- [26:48] Nostr-based Git and agent-native developer infrastructure.

- [32:02] “Apps that aren’t apps”: familiar interfaces over data that stays where it belongs.

- [35:46] Organisations as distributed computers made from local machines.

- [46:34] Pete’s draft NIP for `window.fips` transport.

- [47:56] Nostr plus FIPS as enough to rebuild a better internet.

- [52:57] “Dropbox killer”: virtual file systems without central infrastructure.

- [54:10] The pivot: Dario, Anthropic and the “AI will kill everyone by 2030” debate.

- [56:03] Pete’s proposed AI safety fix: make murder illegal.

- [57:32] Andy calls the regulatory push classic cartel behaviour.

- [01:02:16] Pushback from Jensen and the open-source AI argument.

- [01:07:30] Existing antitrust and anti-competitive laws already matter.

- [01:11:06] Are models plateauing, and is regulation a moat?

- [01:13:21] Paperclips, recursive self-improvement and defensive AI.

- [01:14:14] If the labs are serious, why not regulate themselves?

**Friends of the Pod:**

John on Nostr, Dan (ngit), Jack, Paul

**Projects Mentioned:**

FIPS, Nostr, Wingman, Flight Deck, Autopilot, Tower, Google Drive, Cloudflare, GitHub, Forgejo, NGit Workshop, GRASP, DaVinci, Blender, FFmpeg

**Concepts:**

Private mesh networking, Nostr identity, FIPS addresses, local-first software, agent-native infrastructure, self-hosting UX, device keys, browser transports, open-source AI, AI regulation, cartel behaviour, recursive self-improvement

**Quote:**

“Nostr plus FIPS, it’s enough. We can rebuild everything now.” — Pete

Paul Itoi, founder of Stakwork, thinks the AI industry has the model wrong. Reliable enterprise AI will depend less on agents doing everything and more on workflows, shared memory and systems that can turn discovered work into repeatable software.

Paul joins Pete and Andy to explore factored cognition, recursive AI systems, graph memory and Stakwork’s forthcoming open-source work. They also get into what happens when AI removes the easy parts of work, why successful automation could actually mean less agent use, and where value accrues in a world of agent marketplaces and Bitcoin micropayments.

**Key Moments:**

- [00:05] Paul returns to the pod with an update on Stakwork, AI orchestration and Lightning.

- [01:28] The “fourth epoch of AI psychosis” and why Paul thinks the industry is wrong.

- [05:19] Model main character syndrome: why memory, actors and workflows deserve more attention.

- [07:38] Agents are great at discovery; workflows are great at repetition.

- [09:54] Show your working: why spreadsheets, evidence and auditable steps matter more than a polished answer.

- [12:49] The runtime question: put intelligence inside a process, rather than asking an agent to run everything.

- [15:01] Stakwork’s forthcoming open-source approach to continual learning, large context and recursive improvement.

- [16:21] The robotaxi analogy: once an agent finds the route, stop paying it to rediscover the way home.

- [19:34] Workflows that build tools: legal redlines, document timelines and shared graph memory.

- [24:32] How big should an intelligence snack be? Factored cognition, nutrition labels and salmon farming.

- [29:19] Human checkpoints define how much freedom to give an agent.

- [32:00] The contrarian outcome: successful automation could mean less agent use as repeatable work becomes software.

- [37:19] When AI removes the easy work, humans are left with the hardest decisions.

- [40:53] Pete explains Flight Deck, Autopilot and working with agents away from the desk.

- [46:36] Voice input, visual feedback and a working day built around short bursts of focused judgment.

- [49:44] From company initiatives to individual code changes: organising work through a zoomable roadmap.

- [53:19] Where does the value go? Small specialist models and marketplaces for knowledge and reasoning.

- [59:02] Agents buying information from agents: micropayments, competition and the economics of reuse.

- [01:02:20] Stop trying to convince everyone. Build around your expertise, interests and the marketplace you want to serve.

**Friends of the Pod:** Paul from Stakwork

**Projects Mentioned:** Stakwork, Hive, Sphinx, Harvey, Wingman, Flight Deck, Autopilot, Tower, OpenClaw, Buzz, Excalidraw, Bitcoin, Lightning

**Concepts:** Factored cognition, recursive language models, shared graph memory, dynamic workflows, auditable AI, human checkpoints, cognitive load, agent marketplaces

**Quote:** “Agents are great at discovery. Workflows are great at repetition.” — Paul Itoi

Back in the van by the beach, Pete and Andy explore how longer-running agents are changing the way people work. They move from sovereign infrastructure and the new Intelligence Snacks production system to AI-assisted writing, purpose-built agent workflows and the return of the interface. The conclusion: AI is less a finished consumer product than a capital good whose value comes from applying intelligence to real businesses and problems.

**Key Moments:**

- [00:03] Episode 70 begins in the newly panelled van, with less gaffer tape and another fine Perth winter day.

- [02:06] Pete is refactoring Wingman and adding a Forgejo Git backend to Tower, with Nostr identities controlling agent access.

- [03:25] FIPS could make Autopilot, Tower and local WApps securely accessible without exposing machines or configuring domains and DNS.

- [05:29] Npub-based sharing could let friends review apps running directly on a local development server.

- [10:53] Andy introduces the new Intelligence Snacks format and Snack Studio pipeline for transcripts, summaries, voxel artwork and publishing.

- [12:20] A podcast knowledge graph tracks recurring themes, changing opinions and contradictions across episodes.

- [20:13] AI helps write the snacks, but every output remains tethered to the hosts’ original words.

- [21:15] The real problem with “AI slop” is not the tool—it is publishing without care, judgment or review.

- [29:54] Agent jobs that once took a minute can now run for hours, creating a new coordination and attention problem.

- [32:37] Andy’s Seneca WApp grew from an SEO task into a potential AI chief marketing officer with audits, tickets and follow-up reviews.

- [35:42] Flight Deck notifications and purpose-specific channels separate quick dialogue, bounded work and broader architectural thinking.

- [39:02] Tasks solve defined problems, while interactive experimentation helps reveal the more ambitious system worth building.

- [43:53] Pete revises the idea that AI will kill the interface: apps remain useful visual containers for work and human attention.

- [47:04] Snack Studio demonstrates malleable software—detailed controls can build trust first, then collapse into a simpler workflow.

- [52:12] Testing every new AI product matters less than using capable tools to create commercial value.

- [55:41] AI behaves like a capital good: it accelerates people who already possess judgment, domain knowledge or a good business.

- [61:00] As intelligence becomes commoditised, value moves towards specialised solutions shaped by niche, industry, geography and need.

**Friends of the Pod:** Deadman, Johnathan Corgan, Rod, Gigi, Stan, Jared, Skyler, DHH, Mario Zechner

**Projects Mentioned:** Intelligence Snacks, Snack Studio, Wingman, Tower, Forgejo, FIPS, Autopilot, Flight Deck, Shoshow, Seneca, Buzz, OpenClaw, GrokBot, Omarchy, Pi

**Concepts:** Nostr-mediated infrastructure, podcast knowledge graphs, human-curated AI writing, long-running agents, malleable interfaces, AI as a capital good

**Quote:** “The goal is to be the person that uses the AI more than the AI itself.” [56:03]

Should AI agents have wallets, and what happens when they can earn, spend, rent infrastructure and act without asking a human to approve every step?

Pete and Andy sit down with Piers from Mineracks to explore the collision between AI agents, Bitcoin, Nostr and sovereign infrastructure. They get into agent payments, self-funded compute, abuse, identity, tax, cybersecurity and liability.

They also explore a practical principle for building with AI. Use intelligence where judgment is needed, then turn repeatable work into deterministic software.

**Key Moments:**- [00:00] Piers describes moving from Bitcoin mining and data-centre infrastructure into vertically integrated AI services.- [05:20] Agents pay Lightning and on-chain invoices, move eCash and buy physical goods from natural-language requests.- [07:00] Hardware wallets can keep humans in the loop for consequential agent payments.- [08:14] HTTP 402 and L402 return as payment rails for machines excluded from permissioned fiat systems.- [09:50] Roland from Alby built an agent that rented a virtual machine, moved itself and could pay to stay alive.- [12:43] MineRacks Cloud reduces launching a Bitcoin-enabled shop from roughly 90 days to about 90 seconds.- [17:54] Anonymous infrastructure creates an asymmetry. A customer spends a few sats while the physical host carries enormous legal exposure.- [18:38] Useful agents need delegated financial authority, much like assistants and employees already do.- [25:37] Nostr login provides cryptographic identity and potential web-of-trust signals through social graphs, zaps and NIP-05.- [28:44] Relatr and encrypted storage offer ways to assess trust and reduce the liability of holding readable customer data.- [38:21] Hosted Nowhere flips the hosting model by encoding an entire site into its URL and using Nostr for encrypted interactions.- [44:04] Piers’s hybrid system uses local AI to reduce 10,000 log lines to 50 outliers before sending those to a frontier model.- [47:11] The Intelligence Snacks principle - apply probabilistic intelligence where judgment adds leverage; script the rest.- [52:39] Delegated authority raises unresolved questions about counterparty risk, fraud, tax, insurance and agent liability.- [58:56] Open-ended agents can turn benign goals into dangerous actions; constrained tools and deterministic workflows reduce the risk.- [63:28] Bitcoin may be the canary for AI-assisted attacks, but its exposed surface could harden faster than legacy financial software.- [71:35] Piers closes with Wallet Playground, MineRacks Cloud, sovereign iCloud backup and a Wi-Fi QR utility built on a plane.**Friends of the Pod:** Earthmum, Haza, Gigi, Jesús, Alekos, Rob Hamilton, Sovereign Engineering, planes with Starlink**Projects Mentioned:** [MineRacks Cloud](https://mineracks.cloud), [Wallet Playground](https://walletplayground.com), [Soap 21](https://soap21.com), [Alby Hub](https://getalby.com), [Relatr](https://www.relatr.xyz), [Hosted Nowhere](https://hostednowhere.com), BTCPay Server, WooCommerce, Enclavia, Maple, Nostr, Blossom and Cashu**Concepts:** agents with wallets, delegated financial authority, L402, agent-funded compute, Nostr web of trust, secure enclaves, hybrid private/frontier AI, deterministic agent workflows**Quote:** “Use the intelligence to design the process … then turn it into software rather than token burn.” [46:58]

Hosts: Pete and Andy · Guest: Rod, founder of Shoshow

What changes when one founder with a fleet of AI agents can build what once required venture capital, a technical co-founder, and a large engineering team? Pete and Andy sit down with Rod to explore the rise of the sole entrepreneur: keeping equity and control, making more bets, attacking bloated incumbents, and building deeply for small communities. Rod grounds the theory in Shoshow, his Nostr-based live-commerce platform, and closes with a first look at Livelier.live, an open-source bridge bringing self-hosted streams into Nostr.

**Key Moments:**

- [00:00] Rod joins the campfire and introduces himself as a Sydney technology entrepreneur, former listed-company CEO, and founder of Shoshow.

- [02:43] The old startup path meant giving away equity for a technical co-founder, employees, and repeated capital raises—until the founder increasingly became an employee of the company.

- [07:22] Generative AI threatens the traditional technology moat: software built with tens of millions in capital may soon be cheap for a competitor to reproduce.

- [09:13] Coding agents move the founder from “I cannot build this alone” to one agent, then credible multi-agent teams, in barely more than a year.

- [12:56] Code may be approaching free, but human judgment, product taste, and deciding what is good enough remain scarce.

- [14:09] If small teams no longer need large amounts of capital to build software, the conventional venture-capital model loses much of its purpose.

- [16:09] Instead of renting generic SaaS, a small business can build software around its exact niche, workflows, geography, and regulatory needs.

- [22:24] Pete separates a business’s value stream from its support functions and argues that AI returns value from expensive software middlemen to the people producing the actual product.

- [24:32] Rod’s optimistic case: more people can test ideas, incumbents lose extractive margins, producers become more efficient, and consumers receive better products at lower prices.

- [30:54] The upstart playbook is to find a hated, high-margin incumbent with an obsolete technology moat, replicate the useful product, and compete with radically lower overhead.

- [32:38] Andy makes the case for going deep rather than wide: serve a small community that knows and trusts you, spend time with customers, and build a durable real-world relationship.

- [36:08] Rod explains how Shoshow emerged from a chance encounter with live-streaming commerce and why Nostr offered an enthusiastic early community plus ready-made content and discovery.

- [40:40] Live commerce combines entertainment, trust, and low-friction selling—especially for collectibles, art, limited inventory, and creator-made physical products.

- [49:11] AI lets founders make a portfolio of small bets instead of staking an entire career on one company, although Andy warns that cheap experimentation still needs focus.

- [52:19] Founders without investors or legacy customers retain the freedom to pivot, stop, or run multiple ideas in parallel; incumbents cannot take those risks as easily.

- [57:18] The group returns to “digital lemonade stands”: small online businesses powered by AI, payments, community, and high-agency operators rather than giant centralized platforms.

- [64:37] Rod previews Livelier.live, an open-source bridge designed to bring Owncast and other independent live-streaming directories into Nostr’s discovery and chat network.

**Friends of the Pod:** Mighty Penguin, Greenie, Deadman, Thomas from No Good Studio, Gigi

**Projects Mentioned:** Shoshow, Livelier.live, Nostr, Owncast, Supply Drop, Digital Lemonade Stands.

**Concepts Introduced:** the sole entrepreneur, founder dilution, free software, vertical micro-SaaS, founder portfolio bets, live-streaming commerce, value returning to the edge

**Quote:** “The code is free, but the judgment still isn’t and the review time isn’t.”