AI Built 3 Websites. Only One Looked Like an Agency Made It. | Practical AI Ep 35

Practical AI: Episode 35

AI Just Did What Agencies Charge $10,000 For. I’ll Show You the Results.

Watch on YouTube

Published: April 3, 2026

TL;DR

  • OpenAI closed the biggest funding round in tech history. $122 billion at an $852 billion valuation. Still not profitable. $2 billion/month revenue, 900 million weekly users. The market is betting on AGI, not products.
  • MCP is becoming the universal standard for AI tools. Anthropic created Model Context Protocol. Now OpenAI and Google are buying in. OpenAI opened GPT plugins through Codex, making MCP the shared language across all major AI platforms.
  • Claude Mythos leaked and it is a big deal. Anthropic confirmed a “step change” above Opus after 3,000 internal documents were exposed through a CMS misconfiguration. Government cybersecurity agencies issued warnings.
  • Olga built a website three different ways to prove a point. A vibe-coded version scored 6/10. An agency framework version scored 8/10. A Claude-from-spec version scored 7/10. The agency framework won because skills encode real expertise. This is the difference between “naked vibe coding” and “framework vibe coding.”
  • AI funding hit $4.2 billion this week. 46% of all VC. Defense AI arrived. Over $2 billion went to military and cybersecurity companies in a single week. Qodo raised $70M for AI code verification.

Table of Contents

About This Show

Practical AI is a weekly live show (Fridays 11am CT) hosted by Olga Pechnenko and Chris Pearson that cuts through AI hype to deliver news, trends, and hands-on tips for builders and founders. Unlike technical AI podcasts, Practical AI focuses on business applications and ROI. What actually works, what is hype, and what you can implement Monday morning.

What You’ll Gain

  • See the real results of building a website three different ways with AI and understand why a 75-minute agency framework beat a quick vibe-coded version. The difference between good AI output and bad AI output is the quality of the skill driving it.
  • Understand why MCP matters for your business as Anthropic’s protocol becomes the shared standard across OpenAI, Google, and every major AI tool. If your tools do not speak MCP, they will be left out.
  • Learn what Cloudflare’s Emdash means for the WordPress ecosystem and why Chris Pearson (creator of the first million-dollar WordPress theme) says it copied WordPress’s architecture including its problems.
  • Hear Olga’s recruiter pushback on Jack Dorsey’s “replace managers with AI” essay and why the five things good managers do (recruiting, motivating, mentoring, coaching, accountability) cannot be automated away.
  • See where defense AI funding is heading with over $2 billion going to military drones and cybersecurity in a single week. The money is moving from consumer AI to national security.

Biggest Takeaway to Implement: Stop vibe coding naked. If you are building anything with AI, start with a spec or a framework that encodes real expertise. Olga tested three approaches head to head. The one backed by an agency framework beat the others every time. Skills that use the software are what we will pay for. The software itself wants to be free.

PageMotor and Practical AI Updates

Free, informative, and FUN!

Frequently Asked Questions

How much did OpenAI raise and what is their valuation?

OpenAI closed a $122 billion funding round at an $852 billion valuation. They have 900 million weekly users and generate $2 billion per month in revenue but are still not profitable. The market is betting on AGI, not current products. Read more below.

What is MCP and why is it becoming the standard?

Model Context Protocol (MCP) is an open standard created by Anthropic that lets AI tools connect to external services. OpenAI and Google have both adopted it, making MCP the shared language for AI integrations across platforms. Read more below.

What is Claude Mythos?

Mythos is Anthropic’s next-generation AI model, confirmed as a “step change” above Opus. Its existence was revealed when 3,000 internal Anthropic documents leaked through a CMS misconfiguration. Government cybersecurity agencies issued warnings about the breach. Read more below.

What is Cloudflare Emdash and why does it matter for WordPress?

Emdash is a WordPress competitor built by Cloudflare in one week using AI. Chris Pearson says it copied WordPress’s architecture including its problems. A plugin gold rush started immediately. Matt Mullenweg responded: “Keep the WordPress name out of your mouth.” Read more below.

What is the difference between naked vibe coding and framework vibe coding?

Naked vibe coding means prompting AI with no structure, spec, or expertise behind it. Framework vibe coding uses an agency-level framework that encodes real design and strategy knowledge into reusable skills. Olga tested both head to head. The framework version scored 8/10 vs. the naked version at 6/10. Read more below.

Can AI really replace middle managers?

Jack Dorsey thinks so. He wrote an essay calling for AI to replace middle managers after Block cut 4,000 employees. Olga pushed back hard using her 13 years in recruiting: good managers recruit, motivate, mentor, coach, and hold accountability. AI can handle logistics but not leadership. 80 out of 100 people should never be managers, but the good ones are irreplaceable. Read more below.

What is Meta TRIBE v2?

TRIBE v2 is Meta’s AI model that predicts brain activity across video, audio, and text. 70x resolution improvement. Trained on 720 subjects and 1,115 hours of MRI data. It can predict brain responses for people who have never been scanned. Meta open-sourced it. Read more below.

How much AI funding was raised this week?

$4.2 billion in AI funding, making up 46% of all venture capital. The biggest shift: defense AI arrived in force, with over $2 billion going to military drones and cybersecurity. Saronic led at $1.75 billion for autonomous military vessels. Read more below.

Practical AI Episode 35: AI Just Did What Agencies Charge $10,000 For. I’ll Show You the Results.

Key Definitions

What is MCP (Model Context Protocol)?

An open standard created by Anthropic that lets AI models connect to external tools, services, and data sources through a universal interface. Think of it like USB for AI. Instead of every AI tool building custom integrations, MCP provides one protocol that works everywhere. OpenAI and Google have both adopted it, making it the emerging standard across the industry.

What is framework vibe coding?

Building with AI using a structured framework that encodes real expertise. Skills, specs, and proven workflows. As opposed to “naked vibe coding,” which means prompting AI with no structure and hoping for the best. The framework version produces better results because the AI has expert knowledge baked into its instructions, not just a vague request.

What is future time orientation?

A concept Chris Pearson uses to describe the mindset shift AI requires. Instead of doing one task once, you set up dominoes. Build the skill, build the system, build the workflow. Then AI lets you knock those dominoes over a billion times. The upfront investment in structure pays off exponentially.

What is GEO (Generative Engine Optimization)?

Optimizing web content so AI search engines (ChatGPT, Perplexity, Google AI Overviews) can discover, understand, and cite it. Traditional SEO focused on Google’s link-based algorithm. GEO focuses on making content structurally readable by AI models that answer questions directly.

Quotable Moments

Skills are the software of the future. Software wants to be free. Skills that use the software are what we’ll pay for.

— Chris Pearson on why agency frameworks beat naked vibe coding

We can do anything with technology and we’re laser cutting ads.

— Chris Pearson on Meta’s TRIBE v2 brain prediction model

80 out of 100 people should never be managers. But the good ones? You cannot automate recruiting, motivating, mentoring, coaching, and accountability.

— Olga Pechnenko pushing back on Jack Dorsey’s manager-free vision

Set up the dominoes. AI lets you knock them over a billion times.

— Chris Pearson on future time orientation

2:40 OpenAI’s $122 Billion Round: Betting on AGI, Not Products

OpenAI officially closed a $122 billion funding round at an $852 billion valuation. To put that in context: that is larger than the GDP of most countries. They have 900 million weekly users and $2 billion per month in revenue. They are still not profitable.

The Scale of This Bet

$122 billion in a single round. $852 billion valuation. $2 billion/month revenue. Still not profitable. The market is not buying a product. It is buying the belief that OpenAI will build artificial general intelligence before anyone else does.

Chris made the key observation: this is not a product bet. Nobody is investing $122 billion because ChatGPT is a great chatbot. The investors are betting that OpenAI reaches AGI and that AGI changes everything. If they are right, $122 billion is cheap. If they are wrong, it is the most expensive experiment in tech history.

5:03 MCP Becomes the Universal Standard

Anthropic created Model Context Protocol. Now OpenAI has adopted it through Codex plugins, and Google is building on it too. MCP is becoming to AI tools what USB became to hardware. One protocol that lets any AI model talk to any external service.

This matters because it means the ecosystem is consolidating around a shared standard. If you are building tools for AI, you build for MCP. If you are connecting AI to your business, you connect through MCP. The walled gardens are opening up because nobody wants to build 15 different integrations for 15 different AI platforms.

Key Takeaway: MCP was Anthropic’s invention. The fact that OpenAI and Google are adopting it instead of building their own tells you everything about where the industry is heading. Open standards win. Proprietary ecosystems are losing.

8:52 Claude Mythos Leaked: A Step Change Above Opus

Anthropic confirmed the existence of Claude Mythos after a significant data breach. 3,000 internal Anthropic documents were exposed through a CMS misconfiguration. The documents describe Mythos as a “step change in capabilities” above Claude Opus.

The Leak

3,000 internal documents exposed. CMS misconfiguration. Government cybersecurity agencies issued warnings. Anthropic confirmed Mythos is real and in internal testing. Described as a “step change” above their current top model.

The leak itself became a cybersecurity story. Government agencies flagged it because the exposed documents contained details about AI capabilities that touch national security considerations. The irony of a safety-focused AI company getting breached through a basic CMS misconfiguration was not lost on anyone.

11:26 Meta TRIBE v2: Predicting How Your Brain Reacts

Meta released TRIBE v2, an AI model that predicts brain activity across video, audio, and text. The resolution improvement is 70x over the previous version. Trained on 720 subjects and 1,115 hours of MRI data. The breakthrough: it can predict brain responses for people who have never been scanned.

TRIBE v2 by the Numbers

70x resolution improvement. 720 subjects. 1,115 hours of MRI data. Predicts across video, audio, and text. Can predict for people never scanned. And Meta open-sourced it.

Chris had the most pointed reaction of the episode: “We can do anything with technology and we’re laser cutting ads.” The capability to predict brain responses at this level of detail, combined with the fact that Meta open-sourced it, raises questions about what happens when advertising companies can model your neural response to content before you ever see it.

Olga’s take was more nuanced. As someone who builds sales training, she sees the potential for understanding how people actually process information. But the advertising implications are the elephant in the room. Meta did not build this for neuroscience research. They built it for engagement prediction.

21:22 Zencoder’s 170% Throughput and Dorsey’s Manager-Free Vision

Two workforce stories landed back to back. Zencoder published first-party data showing 170% throughput with 80% of their original headcount. The CEO wrote it himself. This is not a consulting firm’s estimate. It is a company reporting its own results: more output with fewer people.

Then Jack Dorsey published an essay calling for AI to replace middle managers. Block had already cut 4,000 employees. Dorsey’s argument: the layers of management that exist to translate between executives and individual contributors can be replaced by AI that moves information faster and with less distortion.

Zencoder’s Real Numbers

170% throughput. 80% headcount. Written by the CEO, not a consultant. First-party data from a company that actually did it. More work getting done by fewer people using AI tools.

Olga pushed back hard using her 13 years in recruiting. She broke management down into five things: recruiting the right people, motivating them, mentoring them, coaching them through problems, and holding them accountable. Her NFL coach analogy landed: “Nobody watches a football game and says the coach is a middle manager. The coach is the one who makes the team work.”

Her verdict: 80 out of 100 people should never be managers in the first place. Most “managers” are people who got promoted because they were good at their individual job. Bad managers absolutely should be replaced. But the good ones. The ones who actually recruit, develop, and lead. Those people are more important than ever when AI handles the logistics.

Key Takeaway: AI can replace the information-routing function of management. It cannot replace the human development function. If your managers are just passing messages up and down, yes, AI will replace them. If they are actually building people, they are irreplaceable.

32:26 xAI Lost Its Last Co-Founder

Ross Nordeen, the last remaining co-founder of xAI, left the company. All 11 original co-founders are now gone. Musk described the company as being “rebuilt from the foundations up.”

Chris and Olga both noted the pattern. When every single co-founder leaves a company, it tells you something about the working environment. xAI has access to massive compute through Tesla and SpaceX infrastructure, but the brain drain is real. Grok is functional but has not kept pace with Claude or GPT on the capabilities that matter for business users.

36:52 Cloudflare Emdash: WordPress Competitor Built in a Week

Cloudflare launched Emdash, a content management system positioned as a WordPress competitor. Built in roughly one week using AI. Domain names were bought the moment it was announced. A plugin gold rush started immediately. Matt Mullenweg’s response: “Keep the WordPress name out of your mouth.”

Chris Pearson has a unique perspective here. He built Thesis, the first million-dollar WordPress theme. He spent over a decade inside the WordPress ecosystem before building PageMotor. His verdict on Emdash: they copied WordPress’s architecture, including its problems.

The WordPress Architecture Problem

Emdash inherited the same structural issues Chris spent years fighting in WordPress. The plugin model that creates bloat. The theme/plugin separation that limits what you can build. Cloudflare built a WordPress clone faster with AI, but faster does not mean better when you are copying a flawed blueprint.

The deeper story: AI makes it trivially easy to clone existing software. But cloning copies the architecture decisions, good and bad. PageMotor took a different approach by rethinking the CMS from scratch. Emdash proves you can build a WordPress competitor in a week. It does not prove you should copy WordPress to do it.

The plugin gold rush is real, though. Developers are already building Emdash plugins because Cloudflare’s distribution is massive. Whether the architecture is good or not, if Cloudflare pushes Emdash to their existing customer base, it will get adoption. The question is whether that adoption sticks once people hit the same walls WordPress users have been hitting for years.

47:45 Deep Dive: Three Websites, Three Methods, One Winner

This was the centerpiece of the episode. Olga built the Sales Uplevel website three different ways to test a theory: does it matter how you build with AI, or does the tool do all the work?

Version 1 (53:35): Boring Marketer skills (vibe-coded). Score: 6/10. This was the “naked vibe coding” approach. Olga used a set of marketing-focused skills with no agency framework behind them. The output was functional but generic. It looked like every other AI-generated website. Nothing wrong with it. Nothing special about it either.

Version 2 (56:24): Kenn’s BuildTheWeb agency framework. Score: 8/10. This was the “framework vibe coding” approach. Kenn built BuildTheWeb on top of PageMotor. The framework encodes real agency expertise. Strategy, positioning, design principles, conversion patterns. The 75-minute process produced output that felt like a real agency built it. Because in a sense, one did. The expertise was encoded in the framework.

Version 3 (1:08:54): Claude Code from a detailed spec. Score: 7/10. Olga wrote a thorough specification document and fed it to Claude Code directly. The result was better than the vibe-coded version but not as good as the agency framework. Claude followed instructions well, but a spec document is not the same as a system that has been refined through hundreds of real client projects.

The Scorecard

Naked vibe coding: 6/10. Functional but generic. Agency framework: 8/10. Felt like a real agency built it. Claude from spec: 7/10. Better than naked, worse than framework. The winner: the method that encoded the most real-world expertise.

1:12:15 Skills Are the Software of the Future

Olga then combined all three versions (1:12:15), pulling the best elements from each into a final website. The result was better than any single version. But the real insight came from what the process revealed about the future of AI-powered work.

Chris stepped back and made the statement that defined the episode: “Skills are the software of the future. Software wants to be free. Skills that use the software are what we’ll pay for.”

The big idea: AI makes the software layer free. Anyone can build a website, write code, generate content. What you cannot automate is the expertise that makes the output good. That expertise lives in skills. Frameworks. Processes built through years of real work. The agencies, consultants, and experts who encode their knowledge into reusable AI skills will own the next era.

This is the PageMotor thesis in action. Kenn’s BuildTheWeb tool is built on PageMotor. The framework encodes real agency thinking. When Olga ran the same website build through the framework vs. naked AI prompting, the framework won every time. Not because the AI was different. The same AI powered all three versions. The difference was the quality of the instructions.

Chris introduced the concept of future time orientation (1:23:00). Set up dominoes. Build the skill, build the system, build the workflow. Then AI lets you knock them over a billion times. The people who invest time in building reusable skills now will have an exponential advantage over people who keep doing one-off prompts forever.

1:37:05 AI Funding: $4.2 Billion and Defense AI Arrives

Total AI funding this week: $4.2 billion, making up 46% of all venture capital. The headline shift: defense AI is no longer a niche category. Over $2 billion went to military and cybersecurity companies in a single week.

Top 5 AI Deals This Week

Saronic $1.75B (autonomous military vessels). Rebellions $400M (AI chips, South Korea). Galaxea $290M (robots). TENEX.AI $250M (cybersecurity). 9fin $170M (financial AI).

The defense concentration is new. Saronic alone raised $1.75 billion for autonomous military drones. TENEX.AI raised $250 million for AI-powered cybersecurity. When you add the smaller military and defense deals, over $2 billion in a single week went to companies building AI for national security. This is a pivot from the consumer and enterprise AI that dominated funding through 2025.

Qodo raised $70 million for AI code verification. The thesis: as more code gets written by AI, someone needs to verify it works. Chris was skeptical. His take: you do not need a separate verification layer if you have good engineers who know how to test. Olga compared it to translation quality assurance. When machine translation first arrived, human QA was essential. As the models improved, the QA layer became less necessary. The same pattern may play out with AI-generated code.

Google shipped Veo 3.1 Lite, cutting video AI costs by 50%. This landed the same week OpenAI killed Sora. The video AI market is consolidating fast. Google is pushing costs down while OpenAI exits entirely.

Huawei’s 950PR inference chip is now being ordered by ByteDance and Alibaba. US chip sanctions were designed to keep advanced AI hardware out of Chinese companies. Those sanctions are losing effectiveness as domestic alternatives reach production scale.

Keep Learning