Stop Texting A Guy Every Time Your Website Needs A Change | Practical AI Ep 50

Published: August 1, 2026 · Hosts: Olga Pechnenko and Chris Pearson

Table of Contents

• About This Show
• Frequently Asked Questions
• Key Definitions
• Quotable Moments
• What You’ll Get Out Of This Week
• Called It: Three Companies Confirmed The Same Prediction
• Why Meta’s “Business In A Box” Will Be Bigger Than You Think
• Buzz: AI Agents That Get Their Own Login And Sign Their Work
• OpenAI Cut The Price Of Running AI By 80%
• The Red Flag On That 80% Number
• One In Six Work Messages Is Somebody Else’s Job
• The Painter Who Built His Own Tracker
• Give People A Scoreboard And They Get Better
• Visa Cut 2,600 Jobs And Named AI In The Memo
• An AI Has Answered A Million Restaurant Calls
• Your Shared Claude Chats May Be In Google. Go Check.
• An AI Reads Your Instagram Post Before Any Human Does
• People Are Using AI To Talk To Each Other In Person
• Prompt, Then Write: The Job That’s Left
• Europe’s AI Labeling Rules Start Sunday
• One UI To Rule Them All: Who Wins, Who’s Doomed
• Verify Your AI News Before You Repeat It
• The $50,000 Website Nobody Bought
• You Vibe-Coded A Site. Now What?
• Discovery AI: The Interview That Builds Your Website
• Watch It Interview A Business From Scratch
• It Picked The E-Commerce For You. You Never Saw A Plugin.
• Why Claude Fable Makes The Best Interviewer
• The Design Brief That Becomes Your Site’s Rulebook
• Staging, Your Domain, And Who Owns The Hosting
• Three Sites Already Built This Way, Including One In Arabic
• Where To Get In Line, And What $49 A Month Covers
• Bring Your Own Keys If You Want The Most Power
• Twenty Years Later, Setting Up A Website Still Sucks
• Update Your Sales Page By Talking To It
• The Site Lands, Copy And All
• The Second Problem: Changing It Two Months Later
• The Agent That Runs Your Marketing For You
• The Robot That Cleans Up After The Build
• AI Funding: $11.7B And 64% Of Every Venture Dollar
• NVIDIA Wrote $5B To A Company That’s Never Shipped
• There Are Only Two Numbers
• Resources And Sources

What You’ll Gain

• A realistic picture of what “AI builds your website” actually means in 2026. Not a landing page generator. An interview that produces a multi-page site with a working contact form, an ordering page, a FAQ, generated imagery, a logo, and a quality pass that goes back and fixes its own contrast mistakes.
• The distinction that matters more than the build: who changes it next month. Getting a site made was never the expensive part of owning a website. Being unable to change it is. Test any website solution on the second question, not the first.
• A 30-second privacy action you should take today if you have ever pressed share on a Claude conversation.
• A week of AI news checked against primary sources, including which claims did not survive the check.
• An honest look at a live demo that ran long. The build did not finish on the show’s schedule. You see the wait, the fallbacks, and what actually landed. That is what a live show is for.
• Where $11.7 billion of AI money went this week, and why two checks were nearly three-quarters of it.

Biggest Takeaway to Implement: Look at the thing you currently have to ask another person to do, on their schedule, every time your business changes. For most small businesses that is the website. Write down what you would change this month if changing it were free. If that list is longer than zero, the constraint was never your ideas. It was the request queue.

Key Definitions

Quotable Moments

There’s no small business solution on the internet that makes any damn sense to me.

— Chris Pearson, on why hosting, email, SEO, bookings and e-commerce as five separate purchases has never worked for a small business

Every story you heard here today got checked against the source before we say anything.

— Olga Pechnenko, on why the show exists

Building things is getting easier and easier. Selling them is not getting easier, because there’s a lot more noise.

— Olga Pechnenko, on where the real difficulty moved

The less people know about what’s going on, the less anxiety they’re going to have. It just works. That’s what they know.

— Chris Pearson, on what he calls conceptual disambiguation

If your life and operations is anything more than a one, you’re doing too much.

— Chris Pearson, on consolidating everything into a single command center

It’s your responsibility as a business owner to set up that playing field for success.

— Chris Pearson, on giving people a scoreboard instead of blaming them for the result

You haven’t made a $50,000 website until you sold a $50,000 website.

— The reply under the most-saved AI post of the week, quoted on air

0:00 What You’ll Get Out Of This Week

The fiftieth consecutive Friday. Olga frames the value the way the show always does: best case, you learn something you can use to ship, build or sell with AI. Worst case, you walk out the most interesting person at dinner this weekend. The specific promise she makes is about filtering rather than volume. A great deal of what circulates as AI news does not survive contact with a primary source, and separating the real from the unreal is getting harder every week, not easier.

Chris teases the demo: a website built live on the show, from an interview, unique to the business rather than generic, and afterwards runnable by talking to an AI instead of logging into an admin.

2:30 Called It: Three Companies Confirmed The Same Prediction

On Episode 47, Chris predicted that every good company would converge on one interface where the AI lives and dispatches to everything else. Three official confirmations landed inside one week.

Microsoft confirmed on its FY2026 Q4 earnings call, on July 29, that it is merging Copilot, Cowork and GitHub Copilot into one super app spanning consumer and commercial. Perplexity turned Spaces into Projects on July 30, a persistent shared workspace with a self-improving memory layer and more than 400 tool connections, which means the chat stopped being disposable. And on the same Meta earnings call, Zuckerberg described AI agents evolving into a business in a box.

Olga’s read on all three: this is a race for where you are going to live and operate. Her own version of it is already running. She builds with Claude, connects everything she can over MCP, and has largely stopped logging into things and re-authenticating a dozen times a day. The open question is not whether consolidation happens but who you trust to hold all of it.

4:04 Why Meta’s “Business In A Box” Will Be Bigger Than You Think

Chris’s argument is about who is actually being served. An enormous number of real businesses already run on Facebook groups as their primary channel for marketing, customer communication and sales. Those owners are not reading AI newsletters and they are not the audience for developer tools. When Meta offers them a bundle inside the environment they already live in, with agents running the pieces, adoption will not look like a considered platform decision. It will look like turning on a feature.

His prediction: a large share of web businesses get run strictly through Meta’s properties, fairly soon, and it will happen without much visible discussion. Olga’s counterweight is a durable one. Meta moves quietly, does not explain itself, and the terms are theirs. Everything works until you are blocked or deplatformed for a reason nobody will explain.

6:01 Buzz: AI Agents That Get Their Own Login And Sign Their Work

Chris promised on the previous episode that he would look into Buzz. He came back with a correction and a sharper description of what it is.

The correction first: it is Block’s Buzz, not Jack Dorsey’s startup. Block is the company behind Square and Cash App, where Dorsey is CEO. The announcement, the GitHub organization and the Apache 2.0 license all sit under Block.

It was widely described as a Slack alternative, which undersells it. The real premise is accountability. Companies are already putting agents into Slack and Teams, doing everything from recording transcripts to real work, and enterprises immediately ask how they audit the chain of actions. Buzz answers that at the architecture level. Built on Nostr, it generates a cryptographic signature for every action an agent takes, producing an unbreakable ledger of chats, code contributions and anything else. Agents are members of the channel with their own logins. You can see which agent did which work and which human authorized it.

Chris’s framing: this is really a cryptographic cataloging service for agents. And it is where crypto found a use that sticks. Crypto had its moment, receded, and everyone agreed the technology was useful without finding the magical application. Using cryptography to track what AI actually did is that application.

The honest flip side, stated on air: this is a developer’s tool right now. Self-hosting, code repositories, nine days old, and nobody has migrated a real company onto it. Stars are a bookmark, not a customer.

11:29 OpenAI Cut The Price Of Running AI By 80%

The thing that stopped most small businesses from letting AI run all day was the meter. If agents are constantly watching for triggers and executing tasks, token costs compound quickly. On July 30, OpenAI cut GPT-5.6 Luna, its fastest and most affordable model, by 80%, and Terra, its balanced everyday model, by 20%. OpenAI attributes the gains to improvements in the models, the systems running them, and the harness connecting them to tools.

Olga’s read on the strategy: Anthropic has been clear that it does not want subscription plans powering third-party agent frameworks, though the API remains open. OpenAI said yes to exactly that market. Whoever gets deployed most widely inside businesses wins the next phase, and this is a move to be that.

12:28 The Red Flag On That 80% Number

Chris does not accept the stated explanation. His point is structural: internal improvements to a model do not produce an 80% price cut. That is not the shape of an efficiency gain, that is the shape of a decision. Something else is going on, whether that is a hardware arrangement, a competitive response, or a deliberate loss leader to win the agent market before it consolidates.

Worth holding as an open question rather than a conclusion. It is logged as a prediction and it is falsifiable: if the real driver surfaces later, this is the week to point back to.

13:42 One In Six Work Messages Is Somebody Else’s Job

OpenAI analyzed more than 800,000 US work-related ChatGPT messages. In 16.8% of them, the task belonged to a different occupation than the person sending it. Among occupation-specific messages the figure reaches 43.5%. Crucially, the crossover is highest at the smallest companies: 18.9% at workspaces of 2 to 5 people, against 16.3% at 100-plus.

The picture that produces is concrete. The owner who used to wait a week for a designer, a month for a bookkeeper, or who had to beg a developer for a small fix, now just asks. Chris takes it further and calls it the experience of life generalized. Anything where you defer a responsibility and wait for someone to come back to you is a slow, dependent process, and AI removes it. Olga’s counterpoint is the uncomfortable one: if enough of your job is being done by other people using AI, the conversation about whether your role is needed starts on its own.

16:16 The Painter Who Built His Own Tracker

A painting contractor with no software background built himself a real-time tracking app by talking to AI, so his crews can see where they stand against the budget and their bonus. He reports his labour cost went from an average of 45% to below 37% and trending lower.

Two honest labels on this one. It is self-reported, and he said himself he would provide his P&L in a few months if it stayed consistent, which is his own caveat, offered without being asked. There is also a skeptical reply under the post that he never answered.

Why it matters anyway: it is the same shape as everything else in this episode. Not a generic product bent to fit his business. A specific tool that does exactly what he needs, built by describing the problem, and it moved a number he cares about.

17:19 Give People A Scoreboard And They Get Better

Chris pulls out the transferable principle, and it is the most portable idea in the news block. If you give the humans operating inside your business a scoreboard, you get better results. Not a new incentive, not a new hire. Visibility.

His argument is that employees routinely operate without the specific context that would let them perform better. They do not know where things stand, what good looks like today, or what would move it. A scoreboard supplies that context continuously. And the responsibility sits with the owner, not the staff: setting up the playing field for success is the owner’s job, not something to be waved off as somebody else’s ownership problem.

18:13 Visa Cut 2,600 Jobs And Named AI In The Memo

Visa is cutting roughly 2,600 jobs, about 7% of its workforce, mostly in technology and product. The CEO’s memo says AI is helping accelerate how work gets done. Visa is profitable and growing. This is not a downturn cut, and the company frames it as efficiency and reinvestment rather than replacement.

Chris’s position is that large companies that hire at scale have overhired to a ridiculous extent, and this is a correction people are sensitive about but should not be surprised by. Olga’s angle comes from thirteen years of recruiting: the people leaving these companies carry compensation and benefit expectations set by a large employer, and smaller companies will not match them. Watching that adjustment happen is the interesting part.

20:33 An AI Has Answered A Million Restaurant Calls

Yelp’s AI phone assistant has handled more than a million calls for restaurants. It books, changes and cancels reservations, and takes takeout orders straight into the point-of-sale. Over a single Mother’s Day weekend it handled more than 40,000 calls. Separately, Meta said more than a million businesses use its AI agents every week on WhatsApp and Messenger to talk with customers and complete sales.

Chris’s take is the one small business owners should sit with. If you can have an AI answer your phone, it is almost certainly better than whatever you have now. His example is his daughter’s gymnastics facility, which does not answer during posted business hours. One ring, an immediate answer, and a real solution beats both a voicemail and a one-minute recorded menu. Olga’s pushback: an angry customer wants a human, and the workflow for handing off matters. Both are true, and the handoff is the design problem.

23:42 Your Shared Claude Chats May Be In Google. Go Check.

Hundreds of shared Claude conversations turned up in ordinary Google results, including resumes, business plans, and in some cases keys and credentials.

Olga is precise about what this is, and the precision is the point. It is not a breach. When you create an artifact or a conversation in Claude and press share, you produce a live public web page. It is disclosed, though not prominently, and Google does what Google does. Nobody broke in.

The 30-second action: open Claude on the web, go to Settings, then Privacy, then Shared Chats, and look at what is in there. Unpublish anything you would not post. And the standing rule underneath it: never paste keys or credentials into a chat at all. Have the AI read them from a file on your own machine instead.

One correction to what was said on air. Anthropic’s robots.txt has blocked crawlers since September 2025, and robots.txt only prevents crawling, not indexing. The fix deployed around July 26 was a noindex tag, which Google honored quickly while Bing lagged.

25:54 An AI Reads Your Instagram Post Before Any Human Does

Every public Reel and Feed post is now run through a language model for topic and tone before the recommendation system decides who sees it. If you post to be found, your first audience is not a person. Meta credits the change with a modest lift in Instagram sessions.

The practical consequence: the machine that decides your reach is now a reader, so what you write has to be legible to it as well as compelling to a human. Olga’s line for it is that AI is now the bouncer at the door of human attention.

27:07 People Are Using AI To Talk To Each Other In Person

The Wall Street Journal reported that young adults are using chatbots as a social coach for icebreakers, dating messages, and in some cases real-time help during a face-to-face conversation. Experts quoted in the piece cut both ways: it gets anxious people into conversations they would otherwise avoid, and it deepens the avoidance because the hard part never gets practiced.

Olga extends it to business, where it is just as real. Owners use AI to work out what to say to an upset employee or a difficult client, which is genuinely useful coaching. The risk is identical. If every difficult message is drafted for you, the live version of that skill never develops, and the live version is the one that decides outcomes.

Her rule: use AI to coach yourself, then do the hard part yourself. In a world where everyone can generate polished text, the separator becomes who can actually communicate in the room.

30:05 Prompt, Then Write: The Job That’s Left

Chris describes a working realization. Some people run many prompts in parallel across many tabs. That does not suit him, because he audits everything closely and stays on one task. Which leaves real gaps of time between prompts, and he noticed himself filling them reflexively with social media.

His answer is that the gap is where writing goes. The job becomes prompting and then writing for humans, alternating. The reasoning is worth taking seriously: as the mechanical production of things gets cheap, the highest-leverage remaining skill is the human-connection layer, which for him means being the best copywriter he can be. Prompting is a form of writing, but it is not the same as writing something a person will read and feel.

31:44 Europe’s AI Labeling Rules Start Sunday

The EU AI Act’s transparency obligations became enforceable on August 2. If you sell into Europe or make content for a living, you must tell people when they are interacting with an AI, and mark AI-generated content in a machine-readable way. That covers watermarking generated images and disclosing when a phone agent is an AI rather than a person.

Also in the week: Claude Opus 5 shipped on July 24 at unchanged pricing, and ChatGPT’s desktop apps gained voice, so you can talk to it on Mac and Windows and interrupt it mid-answer.

32:27 One UI To Rule Them All: Who Wins, Who’s Doomed

Chris’s takeaway for the week, with names attached. There is an arms race for the single interface, everyone is in it, and he is willing to call it: Microsoft is doomed, Google has a reasonable play, and Meta will win an enormous population regardless of whether anyone reading this would touch it.

Then he points the same logic at websites, which is the setup for the demo. Are you going to log into your email provider, then your host, then a form tool, then a payment processor, six separate places, to run one website? There is an arms race there too. Anywhere you currently travel to separate places to get related jobs done, expect consolidation, and expect to end up with one preferred point of entry for everything you do.

33:49 Verify Your AI News Before You Repeat It

Before handing off, Olga makes the point that defines the show. She goes through a large volume of AI news every week before air, and items that arrive through newsletters regularly fail a check against a primary source. Not nuance problems. Claims that are simply not true, published confidently, with no verification attached. Her reaction to it is not analytical, it is irritation: who goes on the internet and just lies.

She does not put a figure on how often it happens, and neither will this page. The part that transfers is the instruction, and it applies to any source including this one: before you repeat a claim, know where it came from.

35:08 The $50,000 Website Nobody Bought

The case in point, pulled up live on screen. The most-saved AI post of the week was a tutorial on creating “$50,000 cinematic websites” with Claude Fable 5, at roughly 16,800 bookmarks and 1.5 million views. The post is in Spanish and describes a 16-minute tutorial.

The workflow is real and the site is genuinely beautiful. But the $50,000 is the creator’s own valuation of the output, not a transaction. Nothing was sold. The reply that lands it: you have not made a fifty-thousand-dollar website until you have sold a fifty-thousand-dollar website. Technically the post never claims a sale, which is exactly how the framing works. It gets the attention without the assertion.

And it sets up the actual question. People clearly want this. You can generate a beautiful site. So what happens next?

36:44 You Vibe-Coded A Site. Now What?

Chris lays out the gap between a generated site and a working business, and it is the strongest three minutes of framing in the episode. You have a beautiful thing on your screen. Now you need somewhere to deploy it, which means software capable of running it, which means a server, which means hosting, which means a database and a setup process. WordPress has one-click installs. The new AI-native thing does not.

And that is only the infrastructure. You also want SEO, a custom design, e-commerce, bookings, and email to customers. Each is a separate decision, a separate vendor, a separate subscription, and then an integration problem on top. A large company puts a department on this. A small business owner is told to hire someone they cannot afford, to solve a problem they did not want to have.

38:20 Discovery AI: The Interview That Builds Your Website

The reveal. Discovery AI opens by asking whether you already have a website. If you do, it will examine it and attempt a migration, reading your existing pages and converting the whole thing. For the demo, Chris starts fresh instead, and Olga picks the business on the spot: flower delivery.

39:22 Watch It Interview A Business From Scratch

What follows is the substance of the demo, and the thing to watch is the quality of the questions rather than the speed.

It asks what the business is called and what it does. Where you are based and how far you deliver. Whether you arrange the bouquets yourselves and what sets you apart. Whether there are subscription tiers and whether one-off orders matter. Then the question that separates it from a form: when a visitor arrives, what do you want them to feel and trust in the first five seconds, with two concrete options offered as examples.

The business that emerges is monthly custom arrangements delivered as a recurring gift, with three tiers named Princess, Queen and Empress, one-off orders as a secondary path, and Instagram used for marketing but deliberately not competing with the main call to action. A viewer in the chat suggests randomizing the delivery date so it never arrives on the same day, and it goes straight into the brief.

Olga’s reaction is the useful review: it felt like being interviewed by an agency that understands your business. Chris’s framing is a build partner, which is what most people actually want. Input on the things that matter to you, and no obligation to have an opinion about the technical parts.

43:00 It Picked The E-Commerce For You. You Never Saw A Plugin.

The moment Chris stops to underline, and it is easy to miss. When the interview learns the business runs monthly subscriptions, it replies that the site can handle recurring monthly payments, built right in.

Nobody went looking for an e-commerce plugin. Nobody compared options, checked which supports recurring billing, or priced the add-on. Nobody was even told a plugin exists. The requirement was expressed in business terms and the component decision happened underneath. That is the pattern the whole product is built on, and it is what Chris later calls conceptual disambiguation: the less you have to know about the machinery, the less anxiety you carry.

47:46 Why Claude Fable Makes The Best Interviewer

A genuinely useful piece of model-selection insight. They tested several models to drive the interview. Claude Fable came out ahead, and the reason is specific: it surveys the information space around whatever you just said and comes back sounding like an expert in that field. It uses the right vocabulary and asks the questions a specialist would ask, which is what makes the process feel like an agency rather than a form.

The runner-up was Kimi K3, which Chris describes as very creative but slower, and not as strong linguistically. Fable’s edge is warmth and precision in language, and occasional humor.

The transferable lesson: when a model is doing a conversational job, general benchmarks are not the thing to optimize. Run the same task through several and compare how they behave in role.

53:33 The Design Brief That Becomes Your Site’s Rulebook

Before building, the interview handles design. It offers typography options rendered in the business’s own name, and Olga picks a serif for elevation. It offers color palettes, and when both are rejected in favor of “purple and green, with pops of color like a bouquet,” it does not fail. It proposes botanical green as the foundation with vivid floral pops, asks one clarifying question about warm accents versus jewel tones, and builds a custom palette. Asked for a logo, it makes one.

Then the part that matters most for the long run. The output is a saved design brief, and the brief becomes a living spec the AI consults every time you add anything later. That is what stops future pages from arriving as strangers. Chris’s framing: you do not want to start from a blank slate every time, because you will never get something that folds into the environment. It replaces templates, and it means new pages are generated to fit rather than selected from a limited set.

55:55 Staging, Your Domain, And Who Owns The Hosting

Olga asks the ownership questions directly, which is the right instinct and the same test the show applied in Episode 49.

The site is built on a staging server first. You get an email with admin login, and seven days to decide. If you want it, you supply your domain and it gets connected behind the scenes, so the live site is yoursite.com, not a platform subdomain. Hosting is included. Asked whether she really owns it, Chris’s answer is that you own it the same way you own any hosting: you rent the space, the site is yours, and if you stop liking the arrangement you can leave.

59:31 Three Sites Already Built This Way, Including One In Arabic

While the flower build runs, Chris shows three finished examples, and these are the strongest evidence in the segment.

A dental practice, with generated imagery, a non-linear section divider, treatment pages, a team page, contact, and a working intake form that was generated along with everything else. A bakery in Dubai, built entirely in Arabic and rendered right to left, which he flags as a genuinely annoying problem in traditional systems where a single site-wide setting interacts badly with a theme and leaves artifacts. Here it is simply correct, because the site was generated in that language from the start. And a personal site for his mechanical pencil collection, an early prototype without generated imagery, where he is candid that he did not like the homepage design but the interior pages came out well.

1:04:27 Where To Get In Line, And What $49 A Month Covers

Chris reveals the page live: aboxforyour.site. You enter an email and join a waitlist. He is explicit that access is not immediate, that you will receive exactly one confirmation and then one email when your slot opens, and that there is no marketing sequence attached.

The offer as stated on air and on the page: the design and the build are free on a fresh build. Migration from an existing site costs extra, because it is compute-heavy. Then $49 a month, flat, described as a founding price with no tier two.

The argument underneath the price is subscription consolidation. Email deliverability alone has meant a separate provider for years, along with the SMTP configuration nobody wants to do. Forms have historically meant paying a form company annually for what is a small piece of functionality. Bundling those removes four or five separate services, their individual costs, and the integration work between them. And your email list sits alongside your actual customers in one place rather than in a separate silo.

1:08:10 Bring Your Own Keys If You Want The Most Power

You do not have to use the platform’s AI. Chris is openly a fan of bringing your own keys: connect your existing Claude, ChatGPT or whatever you already pay for, and use that as the interface to your website. He does this himself, and calls it the most magical way to work.

Why this matters beyond convenience is the ownership point the show keeps returning to. If the AI operating your site is one you chose and can change, the model is a replaceable component rather than the foundation. The platform’s own tools are there so nobody is blocked at the door, but the door is not locked to them.

1:13:33 Twenty Years Later, Setting Up A Website Still Sucks

Olga admits she finds the technical side intimidating, all of it, and asks Chris to explain why it has stayed that way.

His answer is a good short history. Many people carry memories of uploading files over FTP, editing a configuration file, needing database credentials, discovering they must first create a database with their host, having already worked out the host and pointed name servers. Twenty years later those problems still exist in some form. The reason is not that nobody tried, it is that the process is spread across many independent parties, which makes it an economic and integration puzzle rather than a technical one. Where it is heading is zero configuration.

1:16:50 Update Your Sales Page By Talking To It

Olga’s own worked example, and it is the most concrete proof in the episode because it happened on a real business this week.

Chris published an article about what comes after WordPress, which changed the pitch. She noticed the beta sales page no longer matched. She opened Claude, connected to her PageMotor site over MCP, and told it to update the beta sales page with what Chris had just published. Twenty to thirty minutes of back and forth later, the page was rewritten with current information, including participant numbers, and published. She wrote none of the copy.

The part she calls out as the real gain is not the writing. The AI also told her which other pages needed updating as a consequence, so she did not have to hold the site map in her head. She approved, it shipped. No admin, no page selection, no editing code.

1:22:25 The Site Lands, Copy And All

The homepage arrives. The headline: set it up once, be the hero every single month. Custom flower arrangements delivered monthly across Austin. The three tiers are framed as “choose your level of legendary,” with the line “crown her princess, make her queen, hail her empress.” A three-step explainer reads: pick your tier, we arrange a surprise, you take the credit.

What both hosts land on is that the copy is good, and it is good because the model knew what it was writing for. The contact page asks for the buyer’s name as the hero of the story, and offers “whether it’s an anniversary, an apology, or just because.” The ordering page is built. The FAQ is built, with delivery questions and the signature-move explanation.

It is not flawless on arrival. Image placeholders sit where generated imagery will go, and the FAQ text has a legibility problem that Chris flags immediately as something he would simply ask it to fix.

1:25:22 The Second Problem: Changing It Two Months Later

This is the thesis of the episode, and it is worth separating from the demo.

Getting a first build is a real problem. It costs thousands and requires a working relationship with a designer. But the worse problem arrives after. Two months later you have a new initiative, you want a page or a form or a section, and you cannot do it. You wait for someone’s schedule, you pay again, and you hope. Then you look ahead and realize this is the arrangement for the rest of the business’s life. So you stop asking, and the website quietly stops reflecting the business.

Be precise about what was live at the time of the show. Editing through the admin works. Connecting your own AI over MCP works, and Olga does exactly that. The built-in chat for editing after the build is the stated direction, and Chris said plainly that he had not tested it and did not know whether it was live on that rig. Olga’s response was to book it as a follow-up demo for a future episode, which is the honest way to handle it.

1:30:31 The Agent That Runs Your Marketing For You

Asked what he would do if he actually ran the flower business, Chris describes a standing agent, and it is a good template for any small business.

The agent tracks incoming testimonials and places them on the site. It is told about specials, and whenever there is one it drafts social content to promote it and makes sure the offer gets prominent placement on the site. Then it emails a report with links to everything it created and drafted, so there is a review step rather than blind automation.

The general shape: define the recurring process once, let it run on a schedule, keep a human at the approval point. Not “AI does marketing,” but a named process, executed repeatedly, reported back.

1:35:07 The Robot That Cleans Up After The Build

The detail Chris says he undersold, and he is right. The build does not stop when the pages exist. A quality pass runs afterward and audits the result, including color contrast and legibility. The FAQ text problem spotted earlier gets caught and fixed by that pass, live, on camera.

Also coming: location detection, so a build run from the EU includes what EU regulation requires, and regional language conventions are applied automatically. Plus the ability to conduct the interview in your own language and receive the site in it.

The pattern worth stealing regardless of what tool you use: when you have AI generate something, have a second pass review it against explicit criteria. Generation and review are different jobs, and the reviewer catches what the builder does not.

1:39:44 AI Funding: $11.7B And 64% Of Every Venture Dollar

Week 35. AI companies raised $11.71B across 86 companies, which is 64.3% of every venture dollar raised in the world, the biggest AI share since June 18. The United States took the largest regional block, China was second at $4.0B, and Europe trailed at $874M.

The honest reading underneath the headline: two checks were $8.5B of the $11.71B. Strip the largest round out of this week and last, and the underlying market went from about $2.08B to about $3.21B, with the number of AI companies funded rising from 58 to 86. Busier week, not just bigger checks. Seed is crowded and cheap, with 35 AI seed rounds at a median of $5.7M. And the money is moving toward running models rather than training them, with AI security emerging as its own sector.

1:41:16 NVIDIA Wrote $5B To A Company That’s Never Shipped

The biggest check of the week went to Safe Superintelligence, Ilya Sutskever’s lab. He co-founded OpenAI, was its chief scientist, and sat on the board that fired Sam Altman. He left to build one thing, and the stated plan is that nothing ships before it. As of this week the company has no product, no revenue, no released model and no published research, with roughly 50 employees. NVIDIA invested $5 billion.

The rest of the top five: Moonshot AI raised $3.5B, the company behind Kimi, reportedly valuing it near $35B, backed by China’s national AI industry investment fund with Alibaba, Tencent and Meituan as earlier investors. Multiverse Computing in Spain announced a raise of up to $569.6M for CompactifAI, which uses quantum-inspired math to shrink large models so they run on a phone or a factory sensor instead of a data center. Etched raised $300M to build a chip that does exactly one thing, run transformer models, by burning the architecture into the silicon rather than using a general-purpose GPU. And HiDream.ai raised a reported $221.6M in Beijing for image generation and editing.

Two precision notes. Multiverse’s own release describes the round as “targeting up to” that figure, which is a target rather than a confirmed close. HiDream.ai’s round is reported by funding trackers with no direct company announcement found.

1:46:52 There Are Only Two Numbers

Chris closes on the theme that ties the news to the demo. In computing there are two numbers, zero and one. Zero is off. One is on. If your operations amount to anything more than a one, you have too much spread across too many places. The goal is a single command center you speak to, which reaches out to everything else on your behalf.

His description of how he and Olga actually work: Claude Code is the brain, and they build it up continuously with how they operate, what they do, and which connections it can reach. So when they speak, one instruction goes out to multiple places, but they only ever operate in one. You invest once in the brain, then run everything from there. If there is such a thing as a singularity, he argues, it exists individually, and it is the one you set up for yourself.

Olga’s closing counterweight is the sentence to take away from the whole episode. Building things is getting easier and easier. Selling them is not, because there is far more noise. That is the next hard problem, and it is the one nobody has automated.

And the reason to subscribe, stated plainly: every story in this episode was checked against its source before it was said out loud.

Resources And Sources

• BuildTheWeb.site. The page revealed on air. Get in line for a build. $49 a month, flat.
• PageMotor. The AI-native CMS every site built this way runs on.
• AI Funding Report, Week 35. $11.71B, 86 companies, every top round verified.
• Weekly AI Funding Tracker. 35 weeks, about $415.5B.
• Block: introducing Buzz. The official announcement.
• Buzz on GitHub. Stars, releases, and the Apache 2.0 license.
• OpenAI: 80% price cut on GPT-5.6 Luna. The official announcement.
• OpenAI: how AI is expanding what people do at work. The one-in-six research.
• Microsoft FY2026 Q4 earnings. The one super app confirmation.
• Perplexity: Spaces are now Projects. Persistent workspace, 400-plus tool connections.
• Shared Claude conversations in Google results. Anthropic on the record.
• Visa cuts about 7% of its workforce. CEO memo, confirmed by a spokesperson.
• Yelp Host passes 1 million calls. Now takes reservations and takeout orders.
• The painting contractor’s tracker, and the skeptical reply he never answered.
• The $50,000 website post, and the reply that lands it.
• WSJ: AI as a social coach. Both sides quoted.
• Claude Opus 5. Shipped July 24 at unchanged pricing.