The One Tweak That Unlocks Data You Already Have | Practical AI Ep 41

Practical AI: Episode 41

How To Get More Out Of The Data You Already Have. One Tweak.

Watch on YouTube

Published: May 17, 2026

TL;DR

  • An AI found a 27-year-old bug with a three-word prompt. Anthropic’s Mythos, running inside Project Glasswing, was told “go find bugs” and surfaced a flaw that sat invisible in foundational internet code for nearly three decades. The hunt is just getting started.
  • The interface layer is now the battleground. Apple’s leaked iOS 27 lets you pick which AI runs each task; Google rebuilt Android around Gemini the same week. Opposite bets, one fork in the road.
  • Anthropic changed pricing for the sixth time in a year. Starting June 15, every paid plan gets a separate monthly credit for programmatic (agentic) use. Chat is unchanged. Claim the credit or your automated workflows move to API pricing.
  • The deep dive: one tweak. Stop saving good presentation for special occasions. Tell Claude to turn the boring markdown file you’ve been avoiding into HTML you actually want to read.
  • Funding hit $9.4B with 60% going to AI. Anduril $5B, Isomorphic Labs $2.1B, and a direct callback to Chris’s Episode 35 prediction that biotech AI was grossly underused. He was right.

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

  • Understand the pricing shift coming June 15 so you know exactly which of your Claude workflows are about to cost money and how to claim the free credit before they do.
  • Learn the one-sentence tweak that turns the data you already have into something you actually want to look at, without writing code or being a designer.
  • Discover why the phone is becoming a model marketplace as Apple and Google place opposite bets on who picks your AI.
  • See what a non-developer’s control panel looks like when every report across four businesses becomes a single visual dashboard.
  • Gain a read on where capital is moving as defense AI, biotech, and off-grid compute pull the week’s biggest checks.

Biggest Takeaway to Implement: Find the ugliest, most boring file in your week — the one you keep avoiding — open Claude, drop it in, and say “make this into HTML I can actually enjoy reading, lead with the headline numbers, make it scannable in 60 seconds.” Then react to what comes back and iterate.

PageMotor and Practical AI Updates

Free, informative, and FUN!

Frequently Asked Questions

What is Anthropic changing about pricing on June 15, 2026?

Starting June 15, every paid Claude plan gets a separate monthly credit only for programmatic use — Claude Code automation, claude -p in scripts, cron jobs, SDK apps, remote scheduling, and similar. Regular chat at claude.ai is unchanged. The credit must be claimed (watch your email and spam folder); unclaimed, your automated usage moves to pay-as-you-go API pricing. Read more below.

What bug did Anthropic’s Mythos find?

Given a three-word prompt to find bugs, Mythos — running inside the internal Project Glasswing program — surfaced a security flaw in foundational, widely-used internet infrastructure code that had gone unnoticed for roughly 27 years. The hosts framed it as a show of force: this is the small one they’re willing to talk about. Read more below.

What is the “one tweak” in the Episode 41 deep dive?

Stop treating polished HTML as a special-occasion thing for other people. Set a rule in your Claude config so that anything you have to read and decide on gets rendered as HTML instead of markdown. Olga’s short version: “Anything I send to a human or have to review, my preferred way is HTML.” Read more below.

How are Apple and Google approaching AI on phones differently?

Apple’s leaked iOS 27 introduces “Extensions” — you pick which AI handles which task (Claude for writing, Gemini for images), each with its own voice, bring-your-own-keys. Google went the opposite direction the same week and rebuilt Android around Gemini Intelligence, no model choice. Read more below.

What was the biggest funding story of the week?

AI took roughly 60% of $9.4B in venture funding. Anduril raised $5B (now valued ~$61B), and Isomorphic Labs raised $2.1B for AI drug discovery — a direct hit on Chris’s Episode 35 prediction that biotech AI was grossly underutilized. Read more below.

What is Anthropic Dreaming?

A background process for managed agents that reviews past sessions, looks at what worked and what failed against your feedback, and rewrites the agent’s own memory to self-improve. Harvey, the legal AI, reported a 6x increase in task completion using it. Read more below.

Practical AI: How To Get More Out Of The Data You Already Have

Key Definitions

What is programmatic (agentic) use?

Any Claude usage driven by automation rather than a human typing at a keyboard — scripts, cron jobs, SDK apps, remote schedulers, GitHub Actions, agent loops that run on an interval. Anthropic distinguishes this “unbounded” usage from “bounded” usage (a human at Claude Code maxes out around 10 prompts an hour; an automated system runs 24/7).

What is Project Glasswing?

Anthropic’s internal program housing Mythos, the bug-finding capability the company chose not to release publicly. Glasswing is where the controlled, internal vulnerability-discovery work happens before anything is shared.

What is “bring your own keys”?

An architecture where the user supplies their own API credentials for the AI model they prefer, and apps or operating systems pass requests through to it. Apple’s leaked iOS 27 is built around this; Chris’s standing prediction is that the apps and platforms that support it win and the ones that don’t lose.

What is the HTML “one tweak”?

A config rule that makes your AI default to rendering decision-grade information as HTML instead of markdown — because markdown files are boring enough that you avoid reading them, and HTML (tables, color, charts, links) is information you actually engage with.

Quotable Moments

I stopped saving a good experience for once a quarter, once a month, once a week. I’m bringing the nice china and I’m using it for every meal of the day, not just when we have guests over.

— Olga Pechnenko on why she switched her data to HTML

The cost of entry was too high. The barrier is now nothing.

— Chris Pearson on what changed

This is the portal from which you control all of your business activities. You can access the reports you need. You can run the processes you need to run. You can see which systems are active and which are stale.

— Chris Pearson on Olga’s cross-business dashboard

How does it feel to be quoted by AI?

— Chris Pearson, after his Episode 35 biotech prediction hit in the funding numbers

Your data deserves to be seen. It deserves to be a visual experience for you to work with and experience.

— Olga Pechnenko on the real point of the deep dive

02:43 AI Found a 27-Year-Old Bug With a Three-Word Prompt

The Show of Force

Anthropic’s Mythos, running inside the internal Project Glasswing program, was given a three-word prompt — “go find bugs” — and within hours surfaced a security flaw in foundational internet infrastructure code that had been invisible for roughly 27 years. More are coming.

The hosts read this less as a single CVE and more as a statement. The flaw itself wasn’t catastrophic — which is exactly the point. As Chris put it, they probably picked one of the smaller findings to make public. The real question is what they are not telling us they found. Decades of audits missed something sitting in plain sight; an AI with a vague prompt did not.

The strategic takeaway for builders: the way modern software gets assembled “promotes sloppiness,” in Chris’s words, not airtight code. As AI is pointed at the popular software everyone depends on — he named Microsoft Word as an example — the expectation should be that bugs are everywhere, and the tooling to find them is now cheap and fast. This connects directly to the broader Glasswing debate about disclosure.

05:44 Apple’s iOS 27 Leak: Pick Your AI Per Task

Per a Bloomberg leak, Apple’s iOS 27 turns the iPhone into something like a fantasy roster for AI: you assemble specialized models for specific jobs. Claude for writing, Gemini for images, your call — each potentially with its own voice. The system is reportedly called “Extensions,” and it decentralizes intelligence away from Siri trying (and failing) to do everything.

The Strategic Question

If your home screen stops being a grid of apps and becomes three connections — Claude, Gemini, Grok — the apps that expose a bring-your-own-keys pass-through to your agent win. The ones that don’t, in Chris’s framing, lose. WWDC is in June, reportedly with a new presenter.

The deeper shift Chris flagged: people underestimate how strongly behavior moves toward the models you actually like. If your phone supports your preferred model, you stay; if it doesn’t, you’ll ditch software you’ve used for twenty years. Preferences win — that’s the bet.

09:53 Google’s Opposite Bet: Gemini Everywhere

The same week Apple opened up, Google did a total 180. At the Android show it unveiled “Gemini Intelligence” — rebuilding the entire OS around one model that understands context and executes cross-app tasks without you ever picking the model. Two operating systems, opposite strategies, announced days apart.

Key Takeaway: Chris stayed consistent with his prior call — tying an OS down to a single model is the wrong bet. Everyone building in the AI expansion is leaning toward bring-your-own-keys and flexibility. This is the third or fourth show in a row where “the UI is disappearing” has been the connective theme.

14:48 OpenAI Moves in Three Directions at Once

OpenAI moved on enterprise services, advertising, and defensive cyber in the same week. The Deployment Company (“Deploy Co”) got more detail: highly technical OpenAI people go into Fortune 500s, map how data flows through old systems like SAP, identify the manual workflows, and replace them with agentic loops. Chris’s blunt summary: data comes in, gets proceduralized, goes out — that’s a definite candidate for replacement, and OpenAI becomes the nervous system of those companies. Olga’s harder point: the layoffs follow.

OpenAI’s Ad Ambition

OpenAI is moving away from the minimalist chat interface toward search-style contextual CPC ads inside responses, targeting $2.5B in ad revenue this year and $100B by 2030 — putting it on par with Google search. The hosts were skeptical: bolting old-internet patterns onto the new internet “just doesn’t feel like a fit.”

Third move: GPT-5.5-Cyber (“Daybreak Cyber AI”) rolled out to EU defenders, with the Pentagon already holding access — OpenAI’s answer to the same ground Anthropic’s Mythos covers.

20:04 Musk v. OpenAI Goes to the Jury

Closing arguments wrapped Thursday. Sam Altman testified that Elon Musk wanted 90% equity early on and proposed merging OpenAI with Tesla. The most-quoted detail of the week was the lawyers’ theater — Musk’s attorney asking the jury to imagine crossing a bridge “built on Sam Altman’s version of the truth.” The 2023 board ouster came back up: a unanimous no-confidence vote, somehow reversed after letters of support from hundreds of Microsoft engineers — Microsoft being OpenAI’s number-one financial partner.

Olga’s prediction (logged): “I don’t think Elon Musk is going to win. There’s too much money at stake. OpenAI will pay some fines, paid in the form of something to a nonprofit arm.” The jury began deliberating Monday — this one resolves fast.

25:04 Anthropic’s Pricing Change: What June 15 Means For You

This is the most actionable item in the episode. For the sixth time in twelve months, Anthropic is changing pricing. Starting June 15, every paid plan gets a separate monthly credit usable only for programmatic use — Claude Code automation, claude -p in scripts, cron jobs, SDK apps, remote scheduling, certain GitHub Actions, connections to tools like Zapier, Cursor, Zed, Cline. Regular chat limits stay the same.

Bounded vs. Unbounded Usage

A human at Claude Code tops out around 10 prompts an hour — bounded, no concern. An automated system runs 24/7 with no sleep — unbounded, and that’s what’s being metered separately. Anthropic can detect automated calls by their nature and interval, and is sharpening compute limits ahead of a presumed fall IPO.

Do this now: Paste Anthropic’s pricing release into Claude Code and say “read this and run through everything I have going on — tell me what’s affected June 15.” When the credit email arrives, check spam and claim it. Unclaimed, your agentic workflows move to API pay-as-you-go.

The bigger pattern: Olga attributed it directly to Chris’s standing prediction — the all-you-can-eat subscription buffet for agentic use is ending, and the industry is moving to outcome-based pricing. Expect companies with flat baked-in offerings to reprice before June 15. It may be cheap now; the open question is what an always-on agent costs when servers are overloaded two years out.

34:02 Anthropic Dreaming: Agents That Improve Themselves

Harvey’s 6x

Anthropic Dreaming runs scheduled background processes where a managed agent reviews past sessions. Harvey, the legal AI, reported a 6x increase in task completion using it.

Chris’s read: this is the same principle he uses building skills — a core instruction plus a tracked log of failures and user preferences, periodically reviewed so the system rewrites its own memory to get better. He expects this self-improvement feedback loop to become a standard part of AI, not a feature.

36:21 The One HTML Tweak That Unlocked My Data

The week before, Olga read Thariq’s article, “The Unreasonable Effectiveness of HTML” — past 12 million views. His argument: markdown was great when humans wrote and read the files. Now agents write them, the files are huge and boring, and you procrastinate on reading them. HTML shows what markdown can’t — tables, color, charts, layout, links — so you actually read it, share it, and act on it.

The reframe that landed for Olga: she used to think HTML was the “nice china” you bring out for guests (the show’s funding reports, presentations). Now she uses it for every meal. She locked a rule in her Claude config, installed Netlify so she can publish in one command, and made “publish this” a single sentence that does it all. Her short version of the rule: anything the AI works on lives in markdown (cheaper tokens, the AI doesn’t care); anything she has to look at and decide on — state that changes over time — gets HTML by default, made automatically without her asking.

Pro tip: Add a publish guardrail. Olga’s rule runs three filters before anything goes public — nothing sensitive, internal-only, or wrong-domain. It caught and flagged competitive information before she published it, without her thinking about it.

48:43 Chris’s Turn: A 2,900-Line File Made Readable

Chris is building a PageMotor API with a feature he calls “developer on demand” — tell Claude Code you want a landing page or a contact form on your site and it happens. Planning it produced a 2,900-line markdown spec. “Might as well be 2,900 miles,” he said — he was never going to read it. So he invoked the same tweak: make this an HTML document I can actually look at.

What the HTML Surfaced

The rendered document showed 29 new API actions (bringing the total surface to 92), two new tools to build, and 12 sections of the software affected — with workflows, security model, and error messages laid out visually. A developer using HTML to see information in a better way, not just to present it to others.

The meta-point both hosts converged on: we think about websites too rigidly. Some pages come from inside the CMS; some can just be static HTML you made this way and dropped in. Chris asked Claude to bump that capability up the PageMotor roadmap on air.

52:47 The Design-System File and What’s Coming Next

Olga’s next build is a design-system file that locks her brand visually — colors and fonts — so she stops re-explaining them. Work inside one company’s folder and the HTML comes out in that company’s scheme; switch folders, switch identity. After that: an interactive editor with drag-and-drop for candidates, prospects, and guests; sliders and knobs; SVG diagrams for her “empire map” and flywheel. Chris’s one-liner: “You’re building a little UI” — for herself, on her own terms.

59:00 The Takeaway: Find Your Ugliest File and Make It Talk

Key Takeaway: Find the ugliest, most boring file in your week — the one you’ve been avoiding. Open Claude, drop the file, say “make this HTML I can actually enjoy reading, lead with the headline numbers, make it scannable in 60 seconds.” React to what comes back. Iterate. It’s not just markdown — feed it a CSV or a 200-row Crunchbase export and let it make the data explorable.

The payoff Olga showed: a single visual portal across all four of her companies — agents’ status, what’s still to catalog, personal projects — color-coded (green refreshed in 48 hours, yellow aging, purple public). Chris called it the villain’s master control panel from the 80s cartoons, except she’s actually using it. The point isn’t the dashboard. It’s that data you can see is data that inspires you to act.

66:30 Funding This Week: $9.4B, 60% to AI

The Week’s Numbers

$9.4 billion in venture funding, 10 mega-rounds, 50 AI companies funded out of 209 total — under 25% of companies, but over 60% of the dollars. Average AI round: $189 million. US took 66%; Europe had a strong week.

The top rounds: Anduril $5B (Palmer Luckey’s defense robotics, now ~$61B); Isomorphic Labs $2.1B (DeepMind drug discovery, UK, first human trials cited for late 2026); Graphcore $457M (UK AI chips, the Nvidia backup plan); Mind Robotics $400M (Rivian’s RJ Scaringe building factory robots); Cowboy Space $275M (orbital data centers — rockets whose top half stays in space as a solar-powered, naturally-cooled data center).

Prediction Hit: Episode 35

Olga surfaced Chris’s Episode 35 call on air: “AI and biotech is grossly underutilized — biggest gains for human life expectancy and quality of life, and it’s all happening in Europe, not America.” Isomorphic’s $2.1B, the UK sovereign AI fund as a first-time investor, AI drug discovery becoming a $2B lane — a clean hit. Chris: “How does it feel to be quoted by AI?”

Three patterns: defense AI is now a top-three capital lane; the UK is on the map via AI drug discovery and its sovereign fund; and compute keeps migrating off-grid — ocean data centers, the Anthropic–SpaceX Colossus deal, Cowboy Space. Olga’s framing for why funding closes the show: it shows where things are moving three months before everyone else notices. “We talk about the robots, then three months later they’re everywhere.”

73:01 Why HTML Makes PowerPoint and PDF Obsolete — and Your AI a Thinking Partner

Chris’s close: we lived through a time when good visuals were expensive — skill-intensive or not worth it — and we didn’t see HTML as the medium. You thought PowerPoint for presentations, PDF for documents. HTML is a superior stand-in for both: interactive, linked, with depth a 2D slide can’t have. One page can open into the biggest rabbit hole on the internet.

Key Takeaway: Olga’s bigger point — easy-to-see data doesn’t just inform you, it changes what you do. She’s moving from “AI makes my reports” to “AI tells me every morning what to adjust.” The reports were step one; the thinking partner is step two. Freeing the mental space spent generating reports is what lets her go do the thing she’s actually good at.

The send-off, at 76:12: this works in chat, in Codex, anywhere — your files, go make them tell you a story. Go ship something fun this week, and come back next Friday and show what you did.

Keep Learning