$105M Skio sale
Kennen Frost sold his previous company Skio for $105 million in cash before founding Icon.
EP16: Nobody's seen AI UGC work, world models & AI agents go desktop
A founder spent $12M on a domain, hit $5M ARR in 30 days, then walked back his entire AI pitch — renaming his product from "AI ad maker" to "human ad maker" because AI UGC simply doesn't work yet.
Jun 4, 2026
53:03
Difficulty: Intermediate
Played
God Mode Podcast
EP16: Nobody's seen AI UGC work, world models & AI agents go desktop
A founder spent $12M on a domain, hit $5M ARR in 30 days, then walked back his entire AI pitch — renaming his product from "AI ad maker" to "human ad maker" because AI UGC simply doesn't work yet.
Jun 4, 2026
53:03
Difficulty: Intermediate
Played
TL;DR
Three founders debate whether AI is delivering real value, anchored by Icon.com's dramatic pivot from "AI ad maker" to "human ad maker." The conversation covers why AI UGC still can't replace human creators, a practical media-buying playbook contrasting $1 vs $30 CPM strategies, Fei-Fei Li's framework for world models as the next frontier beyond LLMs, the AI bubble and looming IPO risk, and the new wave of desktop AI agent apps. The single biggest takeaway: human UGC at $1 CPM crushes paid Meta ads at $30 CPM — and no AI platform has closed that gap yet.
#AI UGC
#AI video generation
#world models
#AI agents
#desktop AI apps
#media buying
#CPM strategy
#Facebook pixel seasoning
#AI bubble
#open source business model
#vibe coding
#AI slop design
#Fei-Fei Li
#Nous Research Hermes
#Icon.com pivot
#Icon.com
#human ad maker
#CPM
#Facebook pixel
#Hermes
#Nous Research
#desktop agent
#AI slop
#open source
#freemium
#UGC creators
#retargeting
#LLMs
Four months in, all three hosts accidentally wore white shirts — then spent an hour arguing about whether AI actually works. Topics include Icon.com's pivot from AI ad maker to human ad maker, a deep dive on whether AI UGC actually works, a media-buying playbook, Fei-Fei Li's world models, the AI bubble, and AI agents going desktop.
Chapter list
- UGC
- User-Generated Content — short videos made by real people (rather than brands) that are used as authentic-feeling advertisements, typically on TikTok or Instagram Reels.
- CPM
- Cost Per Mille — the price an advertiser pays per 1,000 ad impressions; a key metric for comparing the efficiency of paid vs organic distribution.
- Facebook Pixel
- A snippet of Meta tracking code placed on a website that cookies visitors and reports back to Meta, enabling retargeting and lookalike audience creation.
- Pixel seasoning
- The practice of accumulating enough user data on a Meta pixel (typically ~1,000 events) before it becomes statistically useful for retargeting and lookalike audience generation.
- Lookalike audience
- A Meta advertising feature that analyses the profile of your existing pixel audience and finds new users with similar demographic and behavioural characteristics.
- LLM
- Large Language Model — an AI system trained on massive text corpora to predict and generate text; examples include GPT-4 and Claude. Used broadly as a synonym for 'current-generation AI'.
- World model
- An AI system that learns the statistical structure of physical space and time — not just text — enabling it to simulate physics, geometry, and actions; proposed as the next step beyond LLMs.
- AGI
- Artificial General Intelligence — a hypothetical AI system capable of performing any intellectual task a human can do, often used as the long-term goal of AI research.
- ARR
- Annual Recurring Revenue — a measure of predictable yearly revenue from subscriptions; a key SaaS growth metric.
- VPS
- Virtual Private Server — a cloud-hosted virtual machine that stays always-on, allowing AI agents to run autonomously without needing the user's local computer to be active.
- qPCR
- Quantitative Polymerase Chain Reaction — a lab technique for measuring gene expression levels; mentioned by Luca in the context of his stem cell research background.
- IDE / ADE
- Integrated Development Environment / AI Development Environment — software tools that combine a code editor, terminal, and AI agent interface; examples include Cursor, Codex, and Claude Code.
- Retargeting
- Serving ads specifically to users who have previously visited your website or interacted with your content, using pixel or cookie data to re-engage warm prospects.
- Open source
- Software whose source code or model weights are publicly released, allowing anyone to inspect, modify, or run it; increasingly used as a top-of-funnel strategy by AI companies.
- AI slop
- Colloquial term for AI-generated content or design that is formulaic, recognisably machine-made, and low on originality; the podcast discusses visual AI slop patterns like purple gradients and misplaced 'live' badges.
- Vibe coding
- The practice of building software almost entirely through natural-language prompts to AI tools, with minimal manual coding; associated with rapid prototyping and characteristic visual design patterns.
- Two-sided marketplace
- A platform business model that connects two distinct user groups — e.g. brands and creators — and derives value from facilitating transactions between them.
- Latency
- The delay between a user's input and a system's response; in the context of AI voice models, sub-200ms latency is considered the threshold for natural-feeling conversation.
Chapter 1 · 00:00
Cold Open & White Shirt Gang Intro
Kennen Frost sold his previous company Skio for $105 million in cash before founding Icon.
$12M domain purchase
Kennen Frost reportedly spent $12 million to acquire the domain icon.com.
Chapter 2 · 02:44
Icon.com: From AI Ad Maker to Human Ad Maker
He spent $12M on a domain and $5M ARR in 30 days — then quietly admitted AI ads don't work.
Icon.com's Full Retreat: AI Ad Maker Becomes Human Ad Maker
Icon.com swapped one phrase on its website — 'AI ad maker' to 'human ad maker' — and that single change reveals the state of AI video advertising. AI UGC doesn't work yet, and the platforms betting on it are quietly admitting defeat.
$5M ARR in 30 days
Icon claimed to go from zero to $5 million ARR within a single 30-day period after launch.
AI UGC → Human ad maker pivot
Icon relaunched by changing just one phrase on its website — swapping 'AI ad maker' for 'human ad maker' — signalling a full retreat from AI-only video.
80% AI, 20% Human: The Winning Product Formula
Icon tried to be 100% AI from start to finish, and it failed. The winning formula is 80% AI with a human in the loop for taste, editing, and final execution. The human layer isn't a weakness — it's what makes the product credible.
80% AI / 20% human loop
Video creation is roughly 80% achievable with AI today, but the final 20% requires a human in the loop, especially for authentic UGC.
Chapter 3 · 10:40
Media Buying Playbook: UGC vs Paid Ads & CPM Math
Go viral on TikTok, pixel everyone, then retarget on Meta — this is the full-funnel playbook.
The Retargeting Flywheel: TikTok Virality to Meta Precision
Go viral on TikTok or X, pixel every visitor, then retarget that warm audience through Meta's precision targeting. Luca's version: dance on TikTok until 30% of Malta knows your face, then run ads to people who already recognise you. CPM drops accordingly.
Malta organic reach: 30% of population
Luca's locally-targeted Reels were reaching approximately 30% of Malta's population organically, potentially warming audiences before paid retargeting.
Chapter 4 · 18:00
Facebook Pixel Seasoning & Meta AI Targeting
Most AI founders have never heard of a Facebook pixel — and it's costing them a fortune.
Facebook Pixel Seasoning: The Foundation Layer Every Builder Misses
Before you run a single paid ad, you need at least 1,000 users cookied on your Meta pixel. Luca was shocked to meet AI builders who had never heard of pixel seasoning — the foundational move that unlocks retargeting and lookalike audiences.
Pixel seasoning: 1,000 users
A Facebook pixel needs at least 1,000 cookied users before it is considered 'seasoned' enough to enable effective retargeting and lookalike audiences.
Chapter 5 · 21:10
Fei-Fei Li's World Models: What Comes After LLMs
The next frontier beyond LLMs isn't more text — it's physics, space, and time.
Fei-Fei Li's World Models: What Comes After LLMs
LLMs only capture the statistical structure of text. Fei-Fei Li's world model taxonomy adds renderers, simulators, and planners — and the simulation layer is the underrated linchpin that could unlock true AGI and slash the cost of AI video.
Fei-Fei Li's 3 world model buckets
Fei-Fei Li's taxonomy divides world models into renderers (pixels), simulators (geometry/physics), and planners (actions) — with simulation as the underrated linchpin.
Human UGC at $1 CPM vs Meta Ads at $30 CPM
A viral UGC post can deliver $1 CPM. The same content boosted through Meta paid ads in New York costs $30 CPM. That's a 30x difference — and no AI platform has closed that gap.
Claude Opus 4.8: 70-80% task coverage
Claude Opus 4.8 can reportedly complete 70–80% of the tasks that a typical desk worker performs.
$1 vs $30 CPM
Organic UGC campaigns can achieve roughly $1 CPM while paid Meta ads in premium US markets cost up to $30 CPM — a 30x difference.
UGC creator pay: $400–600 per 1M views
Ben pays UGC creators a base fee plus a bonus of $400–600 when a video reaches one million views.
One second of video = 25–30 frames
Generating one second of AI video requires producing 25 to 30 individual frames, illustrating the computational cost of AI video generation.
AI video: ~$1/minute to generate
Generating AI video currently costs approximately one dollar per minute, and quality is not guaranteed.
Chapter 6 · 27:40
Open-Sourced Voice AI at 110ms Latency
110ms latency: a new voice AI just crossed the threshold where it feels human.
Voice AI at 110ms: Below the Threshold of Awkwardness
Natural conversation needs a response time under 200 milliseconds. A new open-sourced voice model clocks in at 110ms — one tenth of a second. That's not just fast, it's the threshold that makes AI voice feel genuinely human.
110ms AI voice latency
An open-sourced voice model demonstrated 110 milliseconds of latency — below the ~200ms human-acceptable threshold for natural conversation.
Chapter 8 · 34:00
The AI Bubble: Is the ROI Actually There?
The 1990s fiber glut took decades to absorb — the AI compute boom might too.
The AI Bubble: Fiber Cable Déjà Vu
The 1990s fiber build-out overran by years and crushed investor returns — even though it eventually worked. The AI compute boom may be heading the same way, with companies like Uber and Microsoft already cutting AI spend just as Anthropic and OpenAI eye public markets.
Fiber overinvestment parallel
Luca drew a parallel between the AI compute boom and the 1990s US fiber glut, where overbuilding took far longer than expected to be absorbed but ultimately benefited consumers.
Chapter 9 · 37:10
Hermes Desktop Agent & the Model-Agnostic Future
All AI agent apps now look identical — but Hermes lets you pick any model underneath.
Hermes Desktop Agent: The IDE That Lets You Pick Your Own Model
Hermes just launched a desktop app that looks identical to Claude, Codex, and Cursor — because they all do. The real unlock is that you can swap any model into the backend. Vendor lock-in in AI agents is ending.
Chapter 10 · 43:00
Why OpenAI & Anthropic Haven't Matched Hermes Agency
Why won't OpenAI and Anthropic let you run agents the way Hermes does?
Why OpenAI and Anthropic Haven't Matched Hermes-Level Agency
Hermes lets you run a self-improving agent on a VPS over Telegram or a desktop app. OpenAI and Anthropic keep you locked to user-facing apps. Ben wonders: is that a safety decision, a business model decision, or just a failure of ambition?
Chapter 11 · 45:00
AI Slop Design: The Four Horsemen of the Vibe-Coded Apocalypse
Google is probably already deprioritising apps with these four AI design tells.
AI Slop Design: The Four Horsemen of the Vibe-Coded Apocalypse
There's a Wikipedia article on signs of AI writing, and Ben had all his apps rewritten to remove the traces. The four visual tells of vibe-coded apps are becoming a liability — Google and Bing are likely already deprioritising them.
Chapter 12 · 49:45
AI-Assisted Writing & Luca's Book Project
He wrote a movie plot in 2016 and forgot about it — now AI is turning it into a TV series.
Luca's AI-Assisted Book Project: From 2016 Journal to TV Series
In 2016, Luca wrote a movie plot in his journal about a million-person social experiment that proves even a perfectly balanced society still finds reasons to discriminate. Now he's using AI to write it, trailer it, and potentially turn it into a TV series.
No indexed bits in this chapter.
Show stoppers
Icon.com's Full Retreat: AI Ad Maker Becomes Human Ad Maker
Icon.com swapped one phrase on its website — 'AI ad maker' to 'human ad maker' — and that single change reveals the state of AI video advertising. AI UGC doesn't work yet, and the platforms betting on it are quietly admitting defeat.
Human UGC at $1 CPM vs Meta Ads at $30 CPM
A viral UGC post can deliver $1 CPM. The same content boosted through Meta paid ads in New York costs $30 CPM. That's a 30x difference — and no AI platform has closed that gap.
Fei-Fei Li's World Models: What Comes After LLMs
LLMs only capture the statistical structure of text. Fei-Fei Li's world model taxonomy adds renderers, simulators, and planners — and the simulation layer is the underrated linchpin that could unlock true AGI and slash the cost of AI video.
Snapshots ()
Key Quotes ()
This episode
Claims & Sources
2 / 15 cited (13%)
Factual claims made this episode, and whether a source was named.
Kennen Frost sold his previous company Skio for $105 million in cash.
Kennen Frost spent approximately $12 million to acquire the domain icon.com.
Icon claimed to go from zero to $5 million ARR within a 30-day period after launch.
Icon's website relaunch changed just one phrase — from 'AI ad maker' to 'human ad maker' — with everything else essentially the same.
Fei-Fei Li published a functional taxonomy of world models on her Substack on June 3, 2026.
World models learn the statistical structure of space and time, whereas language models only learn the statistical structure of text.
Claude Opus 4.8 can complete 70 to 80 percent of the tasks that most people in a desk job perform.
An organic UGC campaign can achieve a CPM of approximately $1, versus $20–30 CPM for paid Meta ads in premium US markets.
Ben pays UGC creators a base fee plus a $400–600 bonus when a video reaches one million views.
Generating AI video currently costs approximately one dollar per minute using current LLM-based approaches.
A new open-sourced voice model achieved 110 milliseconds of response latency, below the ~200ms threshold for natural conversation.
Meta's built-in AI audience suggestions are outperforming Lookalike Audience targeting, according to media buyers at a recent coffee meetup.
Luca's locally-targeted Reels are reaching approximately 30% of Malta's population organically.
All five major AI desktop agent apps — Claude Code, Cursor, Codex, Antigravity, and Hermes — have nearly identical user interfaces.
Cluely, which started as a cheating tool, now derives a significant portion of its revenue from operating as a UGC agency.
This episode
Cast
-
Founder of Icon.com who previously sold Skio for $105M and spent $12M on the icon.com domain.
-
AI researcher and World Labs CEO who published a taxonomy of world models on Substack in June 2026, cited by the hosts as the next frontier beyond LLMs.
-
AI ad platform founded by Kennen Frost that pivoted from 'AI ad maker' to 'human ad maker' after going dark in March 2026.
-
Social media and advertising platform discussed extensively in the context of Meta ads, CPM rates, Facebook pixel retargeting, and lookalike audiences.
-
AI company behind Claude, discussed in the context of the AI bubble, potential IPO, and its conservative approach to agentic use cases.
-
AI company behind ChatGPT and Codex, discussed for its lack of support for autonomous cloud-hosted agent use cases versus Hermes.
-
Venture capital firm that signal-boosted Fei-Fei Li's world models taxonomy and is discussed as doubling down on the thesis that AGI requires more than LLMs.
-
Track
Discussed as an example of a major AI spender cutting back, and for integrating Copilot AI into Windows, OneDrive, SharePoint, and Teams.
-
AI research company that released the Hermes agent desktop app, enabling model-agnostic AI agent use from a consumer-friendly interface.
-
AI voice provider noted as sounding 'too professional' — a quality Ben sees as a limitation when trying to emulate casual or contextual human speech.
-
Kennen Frost's previous company, sold for $105 million in cash before he founded Icon.
-
Cited as an early example of a major company publicly questioning its AI ROI and cutting AI spend, catalysing broader industry doubt.
-
Nous Research's self-improving AI agent that launched a desktop app allowing users to swap backend models freely.
-
Anthropic's AI model, discussed in the context of Claude Opus 4.8's task coverage, Claude Code, and the Hermes agent using it as a backend.
-
Short-form video platform discussed as a primary channel for organic UGC distribution and top-of-funnel audience building before Meta retargeting.
-
AI-native code editor compared visually to Hermes, Codex, and Claude Code in a live demo showing how similar all AI desktop agent UIs have become.
-
AI music generator cited as a good example of the freemium model, offering older model outputs free while gating commercial use behind a paid tier.
-
The location where two of the three hosts are meeting in person for this episode, described as an ideal small-market test environment for ad experiments.
Stats
