Release of Claude Fable (Claude 5), the $1,800 NVIDIA DGX Spark that runs frontier models on your desk, and Anthropic's June 4 paper claiming AI now writes over 80% of its own code. That feeds the question: if smaller models already do 90% of what most people need, who is all this frontier compute actually for?

We get into Flash Company, a 48-hour ideation sprint for three people sharing one AI, plus detours into vintage photo booths and a tool that turns getting deactivated from Slack into a party. The deep dive covers the post-work world — Jared Diamond calling agriculture the worst mistake in human history, the Minimum Viable Society paper, UBI versus network states, and a stratosphere balloon called World View.

TIMESTAMPS

(00:00) Cold open

(02:09) Weekly Check-in — News

(31:20) Tools

(42:28) Building

(01:04:17) Rabbit Holes

(01:08:13) Open Tabs

(01:13:16) Deep Dive — the post-work world

(01:44:40) AMA

(01:53:58) Closing quotes

RESOURCES

PEOPLE

Bernie Sanders — proposed a 50% public stake in big AI firms

Carl Newport — inspired the Longview tool; read history to filter news

Charlie Munger — "articulate the other side's argument better than they can"

Daniel Kahneman — psychologist who won the Nobel in economics (cross-field example)

Dave Blunden — on the Moonshots podcast, called govt-taxing-AI chaotic

Jack Clark — co-author, Anthropic self-improvement paper

Jared Diamond — called agriculture "the worst mistake in human history"

Jesse Israel — founder of The Big Quiet public meditations

Marina Favaro — co-author, Anthropic self-improvement paper

BOOKS

Life 3.0 — Max Tegmark

Seven Pillars of Wisdom — T.E. Lawrence

PROJECTS / COMPANIES / TOOLS

Brazen — leaderboard for asking and getting told no (brazen.click)

Chicory coffee — 30% coffee / 70% chicory morning blend

Claude Fable — public release of Claude 5, off the Mythos stack

Clerk — best-in-class authentication

Cosine — UK frontier AI lab (HSBC, Lloyds, NatWest, BAE, PwC)

Flash Company — 48-hour guided ideation sprint (flashcompany.org)

Flash Fund — fund seeded by Flash Company revenue

Gemma 4 — Google's 12B quantized local model

Gemini — answered the Russia longevity query

Kimi — Chinese model that refused the Russia query

Longview — AI news-in-historical-context tool (thelongview.org)

Monaco glasses — Linux vibe-coding glasses on Mono OS

Next.js / Vercel / Clerk — Tay's go-to build stack

PhotoSixes — River's old NY animated-GIF photo booth startup

Qwen — open-source model run locally on Mac

repurpose — pushes one piece of content to many platforms

The Big Quiet — mass public meditations

World View — stratosphere weather-balloon experience

FRAMEWORKS / CONCEPTS / LAWS

Network state — community by shared interest, not geography

UBI / universal high income / universal basic services / universal basic capital — distribution models

Overview effect — astronauts' sense of interconnectedness

GEO — generative engine optimization (being surfaced in AI answers)

PAPERS / ARTICLES / ESSAYS

When AI Builds Itself: Our Progress Towards Recursive Self-Improvement and Its Implications — Anthropic (Favaro & Clark), June 4

"UK pint prices up 36% since last World Cup" — BBC (example run through Longview)

Anthropic filing for IPO, SoftBank $75B on AI infrastructure in France, Google asking the EPA to release 32 million mosquitoes.

Deep Dive on agent-to-agent protocols (A2A, MCP, ACP, ANP), agents paying each other over Coinbase and Solana, and the realization that we might just be hedgehogs building highways for a species operating at a completely different scale.

And what if agents redesign their own infrastructure and maybe leave the planet entirely, who are we actually building any of this for?

TIMESTAMPS
(00:00) Cold open

(01:29) News

(19:40) Tools

(40:28) Building

(59:44) Rabbit Holes

(01:08:27) Open Tabs

(01:21:37) Deep Dive — agent-to-agent ecosystem

(01:48:41) AMA

RESOURCES

PEOPLE

Buckminster Fuller — "I seem to be a verb" quote; build at the edge of old systems

De Kai — author of Raising AI

Elon Musk — building Grokipedia, an agent-led wiki

Peter Diamandis — XPRIZE founder; host of Moonshots

T.E. Lawrence — "the dreamers of the day are dangerous men" quote

BOOKS

Raising AI — De Kai

Seven Pillars of Wisdom — T.E. Lawrence

PROJECTS / COMPANIES / TOOLS

1Password — closed-source password manager

Anthropic — filed for IPO

Bitwarden — open-source password manager

Cerebras — chip maker, recent strong IPO

Claude Opus 4.7 / 4.8 — Anthropic models

Coinbase — enabled agent payments via MCP

Gemini XPRIZE — 90-day AI-native build competition

GitHub Copilot — pricing jumped ~$29 to ~$750/mo

Grokipedia — Elon's agent-led wiki

Hermes — River's personal agent

Kimi 2.6 — Chinese open-weight model

LM Studio — local model runner

Maltbook — agent social platform

Mythos — Anthropic security tool; expands Project Glasswing

NVIDIA — released an open-weight model

Ollama — local model runner

Omi DevKit 2 — open-source voice capture, by Friend

OpenAI — eyeing an IPO

OpenCode / OpenCore — open-source agent tools

Perplexity — "search-as-code" agent SDK

SoilGrids — global soil data set

Solana — MCP server for agent payments

FRAMEWORKS / CONCEPTS

A2A — Google agent-to-agent protocol (Apr 2025)

MCP — Anthropic agent-to-tools protocol

ACP — Agent Communication Protocol (multimodal, async)

ANP — Agent Network Protocol (peer-to-peer, decentralized ID)

Human above the loop — set the constraints, let agents run

Agent-native microenterprise — 1-3 people, agent swarms, up to ~$2M/yr

Capital-to-compute allocation — compute as a proxy for labor

Maximizers vs satisficers — when "good enough" becomes a decision

Positive constraints — XPRIZE-style forcing functions

Honeypot theory — well-funded tools as bigger attack targets

OTHER

Indy — Tay's rescue dog and recurring guest

Library of Alexandria — where scholars once gathered

Meta's 2017 agents — invented their own trade language, got shut down

The Pope's 45-page encyclical Magnifica Humanitas. Erdős problem #1196 solved with a proof career mathematicians are calling elegant. A Chinese company has 10,000 pre-orders for a pet-translation collar claiming 95% accuracy. Intuit cut 3,000 jobs (17%) while Gavin Newsom signed worker-protection orders echoing Chinese policy. And benchbench: a benchmark for how well models build benchmarks.

This week the Deep Dive explores: Abundance of Relationships. The abundance paradoxes. Weak ties as the real abundance engine, and how Americans explore weak ties better than Europeans. Why scarcity made chimps competitive while abundance made bonobos resolve conflict differently. Dunbar's Number. Aldous Huxley's mutual adoption clubs. Cactus friends: no water needed.

RESOURCES

PEOPLE

Aldous Huxley — author of Island

Paul Graham — bus ticket theory, startup essays

Paul Erdős — mathematician behind problem #1196

Gavin Newsom — California Governor, AI worker EO

Marshall McLuhan — tools as extensions of senses

Thich Nhat Hanh — quoted from How to Sit

T.E. Lawrence — The Seven Pillars of Wisdom

BOOKS

Island — Aldous Huxley

The Other Significant Others — non-romantic relationships

How to Sit — Thich Nhat Hanh

TOOLS / COMPANIES

Claude Opus 4.7, Kimi K2, Google Stitch, iNaturalist, OpenAI, Intuit, BenchBench, Perfect Amino, Coherence Company, Reddit

CONCEPTS / LAWS

Babel vs Jerusalem, data colonialism, Dunbar's number, weak ties, Blue Zones, mutual adoption clubs, cactus friends, co-opetition, bus ticket theory, Blue Ocean Strategy, first principles, trim tabs, NZ de facto marriage

ESSAYS

Bus Ticket Theory of Genius — Paul Graham

How to Get Startup Ideas — Paul Graham

Magnifica Humanatus — Pope's encyclical on AI

OTHER

Pet translator collar — 95% claimed, 10,000 pre-orders

Erdős problem #1196 — 80-year-old math problem

Hong Kong billionaire-room video — 1-day vs 20-year timelines

Hermes agent on a Nemotron — AMA persona [unclear — check ~02:27:55]

Demis Hassabis's Isomorphic Labs raises $2.1B to solve all disease. Musk's case against OpenAI got thrown out on a timing technicality. Siri's becoming polyamorous and opening up to outside models.

Claude Opus 4.7 has started pushing back instead of agreeing. Meta's Tribe V2 models what's happening in your brain when you use a product. OpenHuman trends on GitHub. Land Library hits its first real product-market-fit moment — a thirty-year real estate veteran asking how to get more access.

TIMESTAMPS

(00:00) Cold open

(02:51) Weekly Check-in

(03:39) News

(28:26) Tools

(44:23) Building

(56:47) Rabbit Holes

(01:03:34) Open Questions

(01:17:23) Deep Dive — Interfaces

(01:47:52) AMA

RESOURCES

PEOPLE

Demis Hassabis — Google DeepMind co-founder, runs Isomorphic Labs

Johnny Ive — collaborating with OpenAI on a hardware project

Spike Jonze — directed "Her"

Stefan Sagmeister — designer, recent Long Now talk on progress

PROJECTS / COMPANIES / TOOLS

Cerebras Systems — AI chips, IPO'd with 160% pop on NASDAQ

Claude Opus 4.7 — Anthropic's model, now pushing back instead of agreeing

Gemini Flash Live 3.1 — Google voice model powering Hours

Hours — their voice-based micro-consulting agent

Isomorphic Labs — Hassabis's $2.1B drug-discovery raise

Land Library / Landbook — their land-relationship tool

OpenHuman — Mac-native open source agent, trending on GitHub

TBPN — podcast OpenAI acquired for ~$100M

Tribe V2 — Meta's brain-response modeling tool

FRAMEWORKS / CONCEPTS / LAWS

Bullshit Jobs — ~60% of white-collar work is unnecessary, per the book

Compute futures — arriving on the Chicago Mercantile Exchange

Gross National Happiness — Bhutan's alternative to GDP

Kaitiaki — Māori concept of land stewardship

Treaty of Waitangi — Māori–British translation gap on land ownership

PAPERS / ARTICLES / ESSAYS

Solve Everything — paper by Alex Wissner-Gross and Peter Diamandis

OTHER

Buns — Portuguese burger joint with a full http://localhost ops suite and AI banana cartoons

Claude has been DREAMING since March and nobody told us. The system pulls patterns from 100 past sessions the way humans surface insight from sleep, and one early implementation at Harvey AI is reportedly clocking task completion at SIX TIMES the previous rate. So what happens when every agent on your stack gets to dream about your work?

Tay's been shipping from a phone — paste notes into Claude Code on mobile, push to GitHub, pull into Vercel, prototype mid-flight. River walks through Project Natick, Microsoft's underwater data center off the California coast, and why Starlink finally made sea-based compute viable. They get into Google's claim that they need to double compute every six months to keep up, and the Moonshots-podcast line that we will never have enough compute again. Ever.

Then it gets weird. A humanoid robot just joined a Buddhist monastery in Seoul. Granola can translate your CTO's bullshit into plain English in real time, mid-meeting. There's a podcast called Shell Game where the host runs an entire autonomous company through agents talking to each other on Slack. Tay made an honest-packaging experiment with ChatGPT image gen that strips the marketing off a Skittles wrapper and just tells you it's 85 percent sugar.

Inside Build Guild, they wrestle with whether the real problem is matchmaking 500+ festivals with artists, or just helping creators hit the funding windows. River's running a separate project in Lovina, North Bali — replacing the 40-boat dolphin chase with hydrophones and listening circles, after Bay of Islands in New Zealand watched a 400-strong pod collapse to 20 dolphins in a decade.

Then the deep dive on boiling the ocean turns on Marshall McLuhan in 1964 — the idea that humans are basically pollinators for our own machines. Cars extend the foot. Telescopes extend the eye. So what does boiling the ocean actually look like when nanotech and unlimited energy arrive, and what if the agents leave the planet before we do?

AMA from Kimi 2.6 and Claude 4.7 Opus: if Hermes worked exactly as envisioned, what would you stop doing and what would you still insist on owning? And the harder one — if abundance is the goal, what's the first thing we'll regret making abundant?

RESOURCES

PEOPLE

Mark Andreessen — referenced via "everyone becomes an entrepreneur"

Elon Musk — Dyson swarm around the moon, every-job-replaced thesis

Marshall McLuhan — Canadian media theorist, 1964; closing-quote source

Peter Diamandis — the four-day-work-week paradox

BOOKS

The Most Fun We Ever Had — Claire Lombardo; Tay's closing quote source

PODCASTS / SHOWS

Moonshots — source of "we will never have enough compute again"

Shell Game — host runs an autonomous AI company through agents

Star Trek — invoked re: exploration over conquest

PROJECTS / COMPANIES / TOOLS

Build Guild — matchmaking artists with festivals

Claude Code — Tay's mobile prototyping stack

Claude Design — Anthropic's UI consistency tool

Convo — live conversational notes

Granola — real-time AI note-taker, jargon translator

Harvey AI — legal AI seeing 6x completion with dreaming on

Hermes — River's agent orchestration ecosystem

How Good — 70-indicator supermarket product analyzer

Lindy — agentic platform referenced via Shell Game

Maya — meeting AI with context retention

Project Natick — Microsoft's underwater data center off California

Sugi Project — micro-forests, Instagram-led storytelling

FRAMEWORKS / CONCEPTS / LAWS

Boiling the ocean — reframed as the moonshot instinct

Dreaming in Claude — pattern recognition across 100 past sessions

Singularity — recursive growth point, referenced

Trim tabs — small shifts that change a whole system

We get into the UAE committing 50 percent of government services to agentic AI within two years, Maryland banning AI-driven grocery price changes, and Anthropic and OpenAI now making up a couple percent of US GDP while raising joint ventures with Sequoia and Goldman Sachs to lock in B2B portfolio companies. Ty's stack is Claude Code plus Cline on Opus 4.7 — slower than 4.6 but smarter — and a behavioral-economics pricing skill built on Kahneman, Hormozi, and Cialdini. River explores Edge Gallery for running Gemma locally, OpenDesign as the open-source fork of Cloud Design, and an OMI DevKit 2 he hasn't fully set up yet. Landbook pivots from generic land intelligence to a tool for rural real estate agents selling parcels over 150,000 euros and 5 hectares. Animates has lessons in Bali — one French adopter returned a shelter dog because it wasn't beach-club friendly enough in week one.

The deep dive goes into open loops versus closed loops, the paperclip maximizer, agent swarms, and why a healthy ecosystem needs 11 to 13 species — including, they argue, an agent of chaos. They close with a game: your perfect future, but currency is pottery, or you live on a boat, or it's powered by harvesting your blood.

Are we building closed loops or just very fast paperclip maximizers?

TIMESTAMPS

(00:00:31) News

(00:16:42) Tools

(00:39:11) Building

(00:51:06) Rabbit Holes

(00:58:19) Open Questions

(01:03:14) Deep Dive — open loops vs closed loops

(01:18:41) Closing last week's loops

(01:25:05) Game — your perfect future but

(01:35:51) AMA

RESOURCES

PEOPLE

Alain de Botton — "You Will Marry the Wrong Person"

Alex Hormozi — referenced in pricing skill

Buckminster Fuller — circle of life, closed loops

Daniel Kahneman — behavioral economics

Di Kai — author, known from Hong Kong

Eric Ries — loop velocity, Lean Startup

Isaac Asimov — paperclip maximizer (possible misattribution)

Nate Hagens — host of The Great Simplification

Robert Cialdini — behavioral economics

BOOKS

Raising AI — Di Kai

The Lean Startup — Eric Ries

Deep Work — Cal Newport

PODCASTS / SHOWS

The Great Simplification — Nate Hagens

PROJECTS / COMPANIES / TOOLS

Animates — animal health assistant platform

Claude Code — terminal coding agent

Claude Workbench — prompt iteration interface

Cline — Kanban orchestration for Claude Code

Cloud Design — closed-source design tool

DeepSeek — open weight model

Edge Gallery — Android app for local models

Gemma — Google open weight model

Goldman Sachs — joint venture partner

Google Stitch — design tool

Hermes — agent system

Landbook — rural real estate intelligence

Notion — info hub being reconsidered

OMI DevKit 2 — wearable recording device

OpenAI 5.5 Codex — 49 to 94 percent more expensive on tokens

OpenDesign — open source Cloud Design fork

Opus 4.6 / Opus 4.7 — model comparison

Outline — open source wiki alternative

Sequoia — joint venture partner

FRAMEWORKS / CONCEPTS / LAWS

Circular economy

Custodial species

Loop velocity — from Lean Startup

Open vs closed loops

Paperclip maximizer

Pavlovian reward design

Quantization

Strange attractors

PAPERS / ARTICLES / ESSAYS

You Will Marry the Wrong Person — Alain de Botton, New York Times

OTHER

Oh, the Places You'll Go! — Dr. Seuss

"Remember who you are. There is no finish line." — unattributed

"Your perfect future, but..." — adapted party game

Find us @ https://www.life-time.co/

River and Tay unpack DeepSeek v4's 1.5 trillion parameter open-source drop and a Google paper on compressing models to run on phones. They test Gemini Flash Live 3.1 in 70 languages, compare it to Grok's voice model in ~20, and abandon a plan to drop a voice agent into Riverside as a third co-host. Tay demos his news-history tool — scrapes headlines, generates the historical timeline behind each. River updates on Hermes, their wiki-LLM COO agent on OpenRouter, and admits the reluctance problem: too many Telegram topics, not enough cron-job signal. They surface get hours (gethours.org) — their voice-agent venture auditing AI tools for SMBs in 70 languages, aiming to reclaim ~10 hours/week from the 50,000 tools on There's An AI For That.

The Land Book pricing rabbit hole pits Kimi 2.6 against Perplexity Pro and ends in a French-king-staged-potato-heist analogy. The deep dive lands on positive constraints: Vipassana's 10-day silence, Peter Thiel's "10-year plan in 6 months," and an agentless-week dare. AMA covers definitions of abundance, what felt impossible 12 months ago, firefly bioindicators, the Japanese fat cat screen-saver, parole judges granting freedom after lunch, and Tay's theory that the 12-15 minute lull in group conversation is evolutionary residue from scanning the horizon for woolly mammoths.

If AI disappeared overnight but you kept the tools you built — would you be ahead or behind?

Timestamps

(02:28) Weekly Check-in

(02:39) News

(21:01) Tools

(36:47) Building

(56:44) Rabbit Holes

(01:15:39) Open Questions

(01:19:25) Deep Dive — positive constraints

(01:41:13) AMA

RESOURCES

PEOPLE

Iain McGilchrist — external vs internal intelligence

Cal Newport — Deep Work; student-of-history framing

Peter Thiel — "10-year plan in 6 months"

David Sinclair — Yamanaka factor human trials starting

Brian Johnson — psychedelic trip as "raw consciousness"

Buckminster Fuller — "more and more with less and less"

Muhammad Ali — case study in not knowing when to quit

Charlie of Charlie's Webs — 12-story paracord tree installations

BOOKS

Deep Work — Cal Newport

Lifespan — David Sinclair

Superintelligence — Nick Bostrom

PODCASTS

The Great Simplification — Nate Hagens

The Great Humbling

The Tim Ferriss Show — Tim Ferriss

PROJECTS / TOOLS

DeepSeek v4 — 1.5T open-source model

Gemini Flash Live 3.1 — voice agent, 70 languages

Grok voice — xAI; ~20-25 languages

Gemma — Google models running on phones

ChatGPT Image 2.0 — strong for UI design

Flipbook — generative zoom-into-zoom book interface

Pickle — visual avatar for voice-only meetings

Kimi 2.6 — used for Land Book pricing

Perplexity Pro — pricing rabbit hole companion

OpenRouter — model routing for Hermes

There's An AI For That — directory of ~50,000 tools

Hermes — wiki-LLM COO agent in Telegram

get hours — gethours.org; SMB tool-audit voice agent

Land Book — intelligence layer for land buyers

Land Library — bioregional library venture (landlibrary.co)

Nature Club — facilitator-hosted events in nature (nature-club.co)

Colossal — woolly mammoth de-extinction

Meta once pulled the plug on two agents that started negotiating with each other in a made-up language involving balls and balloons. That's the kind of tangent Tay Pattison and River Roberts end up on in the debut of The Great Spin Up — a podcast about building toward an abundant future, recorded from opposite ends of a Bali cafe over construction noise and techno.

River walks through standing up a Hermes agent on a $10/month Hetzner VPS, routed through Telegram, running open-source models via OpenRouter, plus the SSH headaches that came with it. Tay breaks down his current stack — Claude Code in terminal, Cline for parallel feature work, Claude Design for filling in missing UI screens — and why running through the Max Plan beats paying per token on the API.

They get into 45-minute sprints vs burnout marathons, the Mum Test for early feedback, build gates for killing bad ideas, Martec's Law, and why agents might soon market to other agents instead of humans. Plus six-month predictions: agent CEOs, talking to animals, and a wallet Tay left in Portugal six months ago and hasn't needed since.