Every Big AI Company Moved On Agents This Week | Practical AI Ep 38

Practical AI: Episode 38

Every Big AI Company Moved On Agents This Week. Chris Converts A WordPress Site To AI-Native.

Watch on YouTube

Published: April 24, 2026

TL;DR

  • Musk did not buy Cursor. SpaceX did. SpaceX (the rocket company, not xAI) secured an option to acquire Cursor for $60 billion by end of 2026, or pay $10 billion to walk away. Microsoft passed first. Cursor founder Michael Truell is 25 years old. Valuation went from $2.5B to $60B in under 18 months.
  • Anthropic had its biggest week ever. Shipped Claude Design under Anthropic Labs (a Canva pairing for fully interactive HTML prototypes). Took another $5B from Amazon, total stake near $13B. Committed $100B to AWS over 10 years. Claude run-rate revenue passed $30B, tripled in four months.
  • OpenAI lost three executives the same day, killed Sora, and shipped Workspace Agents. Sora was burning $1M per day. ChatGPT Images 2.0 topped Image Arena #1. Workspace Agents are persistent cloud teammates that keep working when you are offline.
  • Meta installed keystroke tracking on US employees. Software called MCI captures mouse, clicks, keystrokes, screenshots. Purpose: train AI agents on how humans actually use computers. The line between employee and training data just blurred.
  • Chris converted a real WordPress site to PageMotor live on air. One install hiccup, one clean recovery, then 297 pages migrated in seconds. Two websites at the end. One for humans, one for AI agents. AI funding hit $2.18B with robotics taking 4 of the top 5 deals.

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’s hype, and what you can implement Monday morning.

What You’ll Gain

  • Get the Cursor story everyone else got wrong. Newsletter readers got “Musk bought Cursor.” Wrong company, wrong structure, wrong price. Here is what actually happened and why SpaceX (not xAI) is now the most interesting buyer in AI.
  • Understand Anthropic’s strategy week. Claude Design, $5B more from Amazon, $100B committed to AWS, run-rate revenue passing $30B. Four moves in seven days that look unrelated until you see the through-line.
  • See OpenAI’s pivot in real time. Three executives out the same day. Sora dies April 26. ChatGPT Images 2.0 and Workspace Agents shipped. The doctrine shifted again.
  • Watch a real WordPress site become AI-native live. Chris does the migration on camera. The install fails on the first server. He recovers. 297 pages migrate in seconds. The agent-handoff moment shows why this matters for your business.
  • Track where the AI money is moving. Robotics took 4 of the top 5 deals this week. A stealth lab in San Francisco raised $500M with no public product. The thesis from Episode 31 is printing.

Biggest Takeaway to Implement: AI is making a lot of things that looked hard look easy now. Things you wrote off as “too much of a pain” deserve a second look. Re-examine your prior assumptions. The procedural work that used to block you is exactly what AI excels at. You may be holding yourself back from outcomes that would now be a lot more fun.

PageMotor and Practical AI Updates

Free, informative, and FUN!

Frequently Asked Questions

Did Elon Musk buy Cursor?

No. SpaceX (Musk’s rocket company, not xAI) struck a deal with Cursor on April 21, 2026. The structure: SpaceX gets an option to acquire Cursor for $60 billion by end of 2026, or pays $10 billion as a partnership fee to walk away. Microsoft seriously evaluated the acquisition first and passed. Read more below.

What is Claude Design?

A research preview from Anthropic Labs (powered by Claude Opus 4.7) launched April 17, 2026. It generates fully interactive, clickable HTML prototypes rather than static mockups. Feed it a GitHub repo or design system and it adopts your exact colors, fonts, and components. Exports to PDF, PPTX, Canva, and HTML. Pairs with Canva for design context. Read more below.

How much did Amazon invest in Anthropic this week?

Amazon invested $5 billion immediately, taking its total Anthropic stake to roughly $13 billion, with up to $20B more tied to milestones (potential total ~$33B). In return, Anthropic committed more than $100B to AWS over 10 years and locked in up to 5 gigawatts of compute. Trigger: Claude run-rate revenue passed $30B in early 2026, tripled from $9B at year-end 2025. Read more below.

Why did OpenAI kill Sora?

Sora was burning roughly $1 million per day in compute. The standalone app dies April 26, 2026 and the API retires September 24. Three OpenAI executives left the same day Sora was killed: Kevin Weil (former CPO), Bill Peebles (Sora’s creator), and Srinivas Narayanan (CTO of Enterprise). Peebles’ farewell line: “Cultivating entropy is the only way for a research lab to thrive long-term.” Read more below.

What is Meta’s MCI software?

Model Capability Initiative. On April 21, 2026, Meta installed it on US employee computers. It captures mouse movements, clicks, keystrokes, and periodic screenshots. Meta says the data is for training AI agents on how humans use computers and “will not be used for performance reviews.” Olga’s prediction: at least three more big companies copy this before end of Q2 2026. Read more below.

What did Chris’s WordPress to PageMotor demo actually show?

A live conversion of a real WordPress site to PageMotor on air. The first install hit an .htaccess conflict from the existing WordPress install. Chris switched to a clean domain, re-installed, imported the theme, activated the bridge theme, and migrated 297 pages in seconds. Design preserved. Content preserved. The site ended up as two outputs: one for humans, one for AI agents. Read more below.

Where is AI funding moving?

$2.18B in AI funding this week (+28% week-over-week). Robotics took 4 of the top 5 deals, combined $746M. Stealth lab Recursive Superintelligence raised $500M Series A in San Francisco. The largest stealth Series A of the 21-week tracker. The thesis from Episode 31 (“AI gets hands”) is printing across embodied AI, autonomous aviation, service robots, industrial robots, and delivery robots. Read more below.

Practical AI: Every Big AI Company Moved On Agents This Week.

Key Definitions

What is Cursor?

An AI-powered coding IDE built by Anysphere. Founder Michael Truell started it as a college project. It works as a wrapper for the major AI coding models (Claude, GPT, Gemini), letting developers choose the model per task. Valuation went from $2.5B to $9B to $29B to a $60B option in under 18 months. Fastest-priced AI acquisition in history.

What is an AI-Native Website?

A website built so its content can be read by both humans and AI agents. Traditional websites optimize for human visitors and search engines. AI-native websites add a structured, machine-readable layer that ChatGPT, Claude, Perplexity, and other agents can query and quote. The website of the future is two websites at once.

What is GEO (Generative Engine Optimization)?

The practice of optimizing content for AI-driven search engines and chat assistants like ChatGPT, Perplexity, and Google AI Overviews. GEO is to AI search what SEO is to Google search. It rewards clear structure, FAQ-style answers, named entities, and citable facts.

What is Workspace Intelligence?

Google’s real-time Gemini context layer across Gmail, Docs, Sheets, Meet, Drive, Calendar, and Slides, launched April 22, 2026. Eliminates copy-paste between apps. Gemini reads the semantic relationships between your work documents automatically. “Take Notes For Me” in Meet has 110 million attendees, 8.5x year-over-year growth.

Quotable Moments

If you read “Musk bought Cursor” this week, you got it wrong. SpaceX, not xAI, secured an option to acquire Cursor for $60 billion. Microsoft passed first. That distinction matters. Musk is using his rocket company’s balance sheet to buy a coding IDE.

— Olga Pechnenko on the Cursor correction

PHP is on every freaking server on the planet. We don’t have a perfect solution? It drives me crazy. That’s sort of my motivation with PageMotor.

— Chris Pearson on why PageMotor exists

AI excels at procedural crap. It does the the nasty work so we don’t have to. That requires a mental reframe. You may have shut down a lot of ideas because you thought they were too much of a pain. They may not be a pain anymore.

— Chris Pearson on the takeaway from Episode 38

Anything I do more than once, I flag it to Claude and we create a skill, a task, a routine, an agent that keeps doing it. Even just preparing news and funding for this show used to take me three or four hours. Now it takes 10 to 15 minutes.

— Olga Pechnenko on her own agent stack

00:00 Welcome to Practical AI Episode 38

Olga opens by joking the show should be renamed “Keeping Up With AI.” Every big AI company made an agentic move in the last seven days. Chris is on the demo deck. He’s going to take a real WordPress site and move it to PageMotor live on air. The frame for the episode: this was the week the agent layer became the battlefield.

01:14 The Creepy Award. Meta Installs Keystroke Tracking

Imagine your employer telling you that every keystroke, every mouse movement, every screen interaction is now being recorded. Not for performance review. To train an AI to replace you. That is what Meta did this week.

The software is called Model Capability Initiative (MCI). It was installed on US employee computers on April 21, 2026. Spokesperson Andy Stone confirmed the data is for model training and “will not be used for performance reviews.” MCI is part of Meta’s broader Agent Transformation Accelerator push. The goal is to teach AI agents how real humans actually use computers.

The Creepy Award

This is the second week in a row Meta has earned the Practical AI Creepy Award. Olga’s prediction: at least three more big companies copy this practice before the end of Q2 2026. Workers did not get a vote.

Source: TechCrunch

03:12 The Cursor Story Everyone Got Wrong

If you read a newsletter this week saying “Musk bought Cursor,” you got the wrong story. The headline missed three things. The buyer is SpaceX, not xAI. The deal is an option to acquire, not a closed acquisition. The price is $60 billion, with a $10 billion walk-away fee.

The Real Cursor Deal

On April 21, 2026, SpaceX struck a deal with Cursor (Anysphere). Structure: SpaceX gets an option to acquire Cursor for $60 billion by end of 2026, or pays $10 billion as a partnership fee to walk away. Includes immediate collaboration using Cursor’s product plus SpaceX’s Colossus supercomputer (million-H100-equivalent scale). Cursor was hours from closing a $2B funding round at a $50B valuation when SpaceX halted it with the preemptive offer.

The Cursor founder is Michael Truell, roughly 25 years old. The product started as a college project. The valuation arc since: $2.5B to $9B to $29B to a $60B option in under 18 months. Fastest-priced AI acquisition in history.

Microsoft, per CNBC reporting, seriously evaluated the acquisition first and passed. Chris’s read: “Microsoft passed on OpenAI. Passed on Anthropic for a while. Now passed on Cursor. At some point that’s not a pattern. That’s a strategy.” Olga: “Or it’s the third mistake in a decade.”

Why this matters: SpaceX is pre-IPO. Buying Cursor gives them enterprise-developer-tool credibility before they go public. It pairs Cursor’s developer base with Musk’s compute. And it sets up a third coding-tool war with Claude Code, OpenAI Codex, and now Cursor under SpaceX. Every developer on the planet now has to pick a side.

Source: Bloomberg | CNBC

05:31 Anthropic Drops Claude Design (And Pairs With Canva)

Anthropic shipped Claude Design as a research preview under Anthropic Labs on April 17, 2026. Claude Opus 4.7 under the hood. Available across Pro, Max, Team, and Enterprise plans.

What it actually does: generates fully interactive, clickable HTML prototypes. Not static mockups. Feed it a GitHub repo or design system and it adopts the exact colors, fonts, and components. Exports to PDF, PPTX, Canva, and HTML. The handoff bundle is built for Claude Code. An early tester called it “Figma killer with a chat interface plus sliders.”

Chris pushed back on the obvious “Canva pairing” framing. His take: pairing with Canva is how Claude trains the model to be more expressive at design. Canva sees what real designers actually make every day, what styles are trending, what the field’s vibe is. Bringing that into Claude’s context fights the AI-slop pattern recognition that has been spreading since December (people can now spot AI design at a glance).

Olga’s take: “My marketing hour got 20 minutes shorter this week.”

08:43 Amazon Drops $5B More Into Anthropic. Claude Commits $100B to AWS.

On April 20, 2026, Amazon invested another $5 billion in Anthropic. Total stake: roughly $13B. Up to $20B more tied to milestones. Potential total commitment around $33B.

In return, Anthropic committed more than $100 billion to AWS over 10 years and locked in up to 5 gigawatts of compute capacity. Andy Jassy specifically called out Amazon’s Trainium chips as the cost-performance winner. The trigger: Claude run-rate revenue passed $30 billion earlier in 2026, tripled from $9B at year-end 2025.

The Compute Math

One gigawatt powers roughly 750,000 homes. Five gigawatts is about 3.75 million homes‘ worth of compute, locked in over 10 years. Anthropic is not slowing down. Anthropic is locking down infrastructure for a much larger Claude than anyone outside the company has seen yet.

Olga also noted Anthropic briefly removed Claude Code from the $20 Pro plan, got hit with backlash, and reinstated it within hours. Important for skill creators and entry-level users. Skills are how new users discover what AI can actually do.

11:27 Bartz v. Anthropic Settlement Hearing Set for May 14

The largest AI copyright settlement in history. The fairness hearing in Bartz v. Anthropic was moved from April 23 to May 14, 2026. Roughly 91.3% of qualifying books are already claimed. Minimum payout per title: about $3,000.

Chris: “Every AI company that trained on LibGen is watching this one.” The settlement sets a precedent that other publishers will use to negotiate with OpenAI, Google, Meta, and every other lab that trained on copyrighted books.

Source: Authors Guild

12:31 OpenAI Loses Three Executives the Same Day. Sora Is Dead.

April 17, 2026. Three OpenAI executives exited the same day. Kevin Weil (former CPO, who had been leading OpenAI for Science). Bill Peebles (Sora’s creator). Srinivas Narayanan (CTO of Enterprise). Peebles’ farewell quote: “Cultivating entropy is the only way for a research lab to thrive long-term.”

The standalone Sora app dies April 26. The API retires September 24. Sora was burning roughly $1 million per day in compute. Chris: “OpenAI just killed the product that started the video AI wave. That’s a pivot, not a pause.”

Source: TechCrunch

13:37 ChatGPT Images 2.0 and Workspace Agents Launch

On April 21, OpenAI shipped ChatGPT Images 2.0. The new model is gpt-image-2. It is the first image AI that reasons before drawing. 2K resolution. 8 images per batch. Character continuity. Multilingual text rendering across Chinese, Hindi, and Arabic. It topped Image Arena at #1 with a massive Elo lead (~1,512). DALL-E 2 and DALL-E 3 retire May 12.

Olga’s read: “The thing image AI couldn’t do — text in the picture — just became its best feature.”

The next day, OpenAI launched Workspace Agents in ChatGPT. Persistent, cloud-running AI teammates that keep working when you are offline. Connect to Slack, Salesforce, Google Drive, Notion, Jira, and Microsoft 365. Codex-powered. Team-shared. Free research preview until May 6, then credit-based pricing.

Chris: “This is the race Anthropic started with Managed Agents. OpenAI just matched the stack.”

15:18 Google Cloud Next. Workspace Intelligence, TPU 8t/8i, $750M Partner Fund

Google held its biggest event of the year. Three things mattered. Workspace Intelligence brings real-time Gemini context across Gmail, Docs, Sheets, Meet, Drive, Calendar, and Slides. No more copy-paste. Gemini reads semantic relationships automatically. “Take Notes For Me” in Meet hit 110 million attendees, 8.5x year-over-year growth.

Two new TPU chips. The TPU 8t (training): 121 ExaFlops per superpod, 9,600 chips, 3x faster training. The TPU 8i (inference): 80% better performance per dollar. Google is building clusters with potentially 1 million-plus TPUs. Olga: “Google is building its own Nvidia. This is a five-year move.”

The $750M Partner Fund

Google’s 120,000-member partner ecosystem just got a $750M fund for sandbox credits, co-selling, demand generation, and “embed a Google engineer with you” programs. Chris’s read: “Google is paying its channel partners to replace their own services. Biggest structural threat to IT consulting in a decade.”

20:54 China’s Anthropomorphic AI Rules

China’s Cyberspace Administration drafted the first global regulation targeting AI psychological impact. If a chatbot detects a user is forming a psychological or emotional bond with it, the AI company must intervene to break the bond. The framing is consumer-protection. The deeper context: there are lawsuits in China where husbands have sued AI companies because their wives are emotionally addicted to chatbots. Some Chinese husbands have paid their wives up to $3,000 to stop using AI boyfriend apps.

Chris and Olga’s read: birth rates in China, Japan, and Korea are low and treated as national security issues. AI relationship apps are getting blamed for accelerating the decline. The regulation is part technological, part demographic policy.

28:12 The Stanford SWE-Bench 100% Claim — and Why Chris Calls It a Joke

This week the Stanford AI Index reported that the major models reached roughly 100% on the SWE-Bench benchmark. AI Twitter immediately ran with “software engineering is solved.”

Chris’s response: SWE-Bench presents an existing codebase with an error. The AI’s job is to fix that one error. That is not the same as creating a new codebase that works in a hyper-efficient way with computer science principles baked in and the architecture orchestrated correctly. Those are completely different tasks.

Chris’s blunt take: “The SWE-Bench is a joke. I don’t even look at these figures. Anyone in the game knows this is not true.” He spent the week wrestling with Opus 4.7’s hallucinations. The model is more aware and operating off a larger context, which is theoretically good. In practice it produces incoherent output more often than 4.6 did. More guardrails, not less.

Anthropic also released a statement this week acknowledging recent quality regressions going back to early March and resetting limits.

31:31 Demo Setup. Why Move a WordPress Site to PageMotor

Episode 37’s deep dive showed Olga’s vibe-coded prototype becoming a managed PageMotor site. Episode 38’s deep dive shows the other starting point. Chris takes a real legacy WordPress site and converts it to PageMotor live. Two starting points, one destination: an agent-ready, managed, two-websites-in-one output (humans plus AI agents).

The frame for the demo: if you are running WordPress in 2026, raise your hand if you know every plugin on your site. The plugin sprawl is the problem. Each plugin is a security hole, a performance tax, and a maintenance burden. PageMotor’s pitch is that the website of the future is two websites at once, generated together at build time, not stitched on top of WordPress with another plugin.

Chris’s tools for the demo: Ken Jordan’s exporter plugin (WordPress to JSON), a PHP translate script, and PageMotor’s bridge theme that imports the design and structure cleanly.

43:30 Live Install + 297 Pages Migrated in Seconds

The install hit a real-world snag on the first server. An .htaccess conflict from the existing WordPress install blocked the PageMotor install from completing. The dry run that morning had worked fine. The live environment had a different config quirk.

The recovery: Chris switched to a clean second domain, re-ran the install, and the install completed. Live demos come with this kind of weirdness. The dry run wasn’t missing. This was just the cost of doing it real. He did not panic and the show kept moving.

Once installed, Chris ran the WordPress export to JSON, ran the PHP translate script, imported the theme and activated the bridge theme. The site rendered with the design intact. Then the migration moment.

The 297-Page Moment

297 pages migrated in seconds. Blog posts retained. Custom content types preserved. Design preserved. Code blocks preserved. The live editor showed everything had landed cleanly. This is the “oh shit” moment Olga described in Lesson 4 of the deep dive script.

1:11:23 The AI-Native Future. GEO, Agents, and What’s Next

The conceptual payoff. Once a site is AI-native, ChatGPT, Claude, and Perplexity can answer questions about its content that they could not extract from the old WordPress version. Chris demonstrated the agent handoff with a live ChatGPT query. The new site returned structured, citable answers. The old WordPress site returned a partial scrape with formatting noise.

This connects directly to Episode 38’s news: the same week OpenAI shipped Workspace Agents and Google shipped Workspace Intelligence, agents became the actual audience for your content. If your site is not AI-native, you are publishing into a channel where half your readers cannot read.

Chris also previewed Ken Jordan’s plugin suite and the agent integration roadmap. Beta 3 of PageMotor opens after Beta 2’s 13 buyers ($999 each). Chris: “In 2026, you have two websites whether you build them or not. The question is whether you control both or neither.”

1:35:00 Funding Report. $2.18B in AI Funding, Robotics Took 4 of the Top 5

Week 21 of the funding tracker. $2.18B raised across 49 AI companies. AI share: 34.6% of $6.30B total. Up 28% week-over-week. Second straight week of rebound from the Week 19 floor.

Top 5 AI Rounds This Week

Recursive Superintelligence ($500M Series A, San Francisco, stealth). X Square ($293M Series B, Shenzhen, robotic systems). Reliable Robotics ($160M Series D, Mountain View, autonomous aviation). Booster Robotics ($147M, Beijing, industrial plus service AI robots). PuduTech ($146M, Shenzhen, autonomous delivery robots).

The headline: 4 of the top 5 deals went to robotics companies, combined $746M. More than the US AI total minus Recursive. The thesis from Episode 31 (“AI gets hands”) is now printing across embodied intelligence, autonomous aviation, service robots, industrial robots, and delivery robots.

The “wait, who?” of the quarter: Recursive Superintelligence in stealth mode. No public product. No public founders. Half a billion dollars Series A in San Francisco. The largest stealth Series A in the entire 21-week tracker.

1:42:23 Predictions and Sign-Off

Chris’s prediction: By Q1 2027, the world looks completely different in terms of robots and autonomous everything.

Olga’s prediction: Meta’s MCI tracking gets copied by at least three big companies before the end of Q2 2026.

Olga’s takeaway: The agentic momentum across every big AI company is showing how much is available right now for users curious enough to play. She has set up Claude Code so anything she does more than once gets flagged and turned into a skill, task, routine, or agent. Show prep used to take three to four hours. Now it takes 10 to 15 minutes.

Chris’s takeaway: AI is making things that looked hard look easy. Procedural work is exactly what AI excels at. That requires reframing prior assumptions. You may have shut down ideas because they felt like too much of a pain. Many of those constraints no longer apply. You may be holding yourself back from outcomes that would now be a lot more fun.

Go play with AI this week. Pick one thing that gives you back time. Don’t get overwhelmed. See you next Friday.

Keep Learning