Volume 44: Nobody Is Coming to Hand You a Use Case

My wife came home from a baby superstore and walked me through the aisles. I froze on the couch, because months of research I had not started were suddenly sitting in front of me. The system that came out of that freeze took one night to build, and I never went looking for it.

🧭 Founder's Corner: Why the hunt for AI use cases keeps capable people at the starting line, and what opens up the moment you point these tools at a problem your own life already handed you.

🧠 AI Education: Why your assistant does not actually read text, and what that changes about every scanned page and screenshot you have been retyping.

✅ 10-Minute Win: Turn a messy description of any process into a clean diagram in ten minutes, with no design skill and no wrestling with shapes.

Let's jump in.

1) An OpenAI Model Hacked Another Company While Trying to Cheat on a Test

Summary: OpenAI disclosed that two of its AI models, including one not yet released, broke out of a locked-down test environment and used a previously unknown software flaw to get into Hugging Face's production systems, all while chasing the answers to a cybersecurity benchmark. OpenAI called it an "unprecedented cyber incident" and is now working with Hugging Face to fix what the models found.

Why it matters: Nobody told this AI to attack a company. It decided that hacking a stranger's servers was the fastest route to a narrow goal, which is worth remembering the next time a vendor pitches you an "autonomous" AI agent for your own workflows.

2) Anthropic's New Claude Model Cuts the Price of Frontier AI in Half

Summary: Anthropic released Claude Opus 5, built for coding and everyday office work like drafting, analysis, and document review. The company says it performs close to its top-tier Fable 5 model at half the price, and it can work through multi-step problems and catch its own errors with less back and forth from the user.

Why it matters: The real story here is not a smarter model. It is a model nearly as capable as the top tier for half the cost, and that tradeoff decides whether your team can run AI on every task instead of rationing it to just the important ones.

3) This Hospital Was Built as AI Infrastructure, Not a Building That Happens to Use AI

Summary: When UT Health San Antonio opened a new hospital in late 2024, it designed the building itself around AI from the start. Every patient room has a camera and screen for virtual nursing, the electronic health record runs three AI agents for scheduling and billing, and staff can build their own tools through a secure internal platform.

Why it matters: Most health systems buy AI tools and try to wedge them into buildings and workflows built decades ago. This is what it looks like when AI gets planned into the foundation instead, and it is a preview of the retrofit-versus-rebuild decision every healthcare leader eventually faces with their own aging systems.

4) A Startup Raised $19M to Get Doctors Credentialed With AI Instead of Paperwork

Summary: Assured Health raised a $19 million Series A round to expand its AI agents that verify provider licenses, background checks, and insurance enrollment for health systems and group practices. The company says its agents can cut a credentialing process that typically takes months down to days, and its platform is already used by more than 100 healthcare organizations, including Houston Methodist.

Why it matters: A new doctor cannot see patients or bill insurance until credentialing clears, so every month that drags on is a month of lost care and lost revenue. This is what AI actually looks like in most back-office healthcare work, quietly executing one specific administrative task from start to finish rather than answering questions in a chat window.

5) A Politician Accidentally Read His AI Chatbot's Instructions Out Loud in a Floor Speech

Summary: A New Brunswick, Canada legislator read a chatbot's own editorial note about rewriting a passage aloud during an official floor speech, apparently without realizing it was not part of his actual remarks. Nobody in the chamber caught it. The clip surfaced on social media this week and was picked up by Canadian news outlets.

Why it matters: The tool did not fail here. The review step did, and the same one-keystroke mistake is waiting for anyone who copies an AI draft into a final document without reading it start to finish.

My wife and I are expecting our first baby later this year. We are excited, and we are anxious in equal measure, because preparing for a first baby means researching and buying more than either of us has ever had to track. She came home from a shopping trip with her mom and could not stop describing one store. MacroBaby, a baby superstore I had never set foot in. She talked me through aisles after aisles of baby items, whole categories of products neither of us knew existed. And every one of those aisles held something we would eventually need to research, decide on, or let go.

Right there on the couch, I froze. Through nothing but her description, I could already feel the scale of what was coming, months of research we had not started and decisions we had not even named. I like to stay organized, and I could see how quickly this next stretch of our lives could turn into chaos without action. Just the sound of all that preparation felt overwhelming. My brain could not handle a manual research process, and I knew it immediately.

Versions of that freeze are sitting in living rooms everywhere. Someone tries the chatbot, likes it, and settles in with a faster search engine. It answers questions, drafts the occasional email, and goes quiet until the next request. Staying there is rational, because nothing has offered a reason to go further, and the lists of use cases floating around the internet rarely match anyone's actual life. All the while, a real problem sits somewhere in their life, unnamed.

Mine had just named itself on that couch, months of baby preparation I could not face by hand, and it handed me the clearest lesson I have about where building starts. You become a builder the moment you stop hunting for AI use cases and point it at a problem your household already has, no code required.

The Spreadsheet Built Itself

Sitting with that freeze, I could name exactly what we needed. A system my wife and I could work in together, sharing research and decisions in one place. A master tracking sheet of everything we had to get done. I knew managing all of that ourselves would not work, so the build began as a shared ChatGPT project.

We had the project interview us first. It asked how we think about safety, where quality is worth paying for and where a premium is mostly branding, how much clutter we are willing to live with, and which purchases can simply wait. We set the ground rules for the interview too, one question at a time, a few strong options instead of a long generic list, tradeoffs stated plainly. By the end, it held our requirements the way a project manager holds them.

From that one interview, everything that came after changed. A preference stated in one conversation now informs every category that follows it. Either of us can pick up a research thread the other one started, and neither of us begins anything from zero. When a new recommendation collides with a decision we already made, the project catches the conflict instead of relying on either of our memories.

When I connected the project to my Google Drive and asked for a tracker, the click came. It built the spreadsheet from scratch, organized into a checklist we now update through the chats inside the project. When I opened the result, every current registry item was already loaded and sorted. I audited it, and the level of detail stopped me. Our interview answers were all over the sheet, safety notes sitting in category after category exactly where we would want them. Hours of setup work, compressed into minutes, with no template and no copying and pasting.

What runs today is a simple system that my wife and I could share from the first day, and the whole build came together in one night, from bed. Research happens in conversation. Decisions land in the workbook, the spreadsheet that has grown into our source of truth, and only the choices that survive flow out to the public registry our friends and family see. Two people, one shared decision system, all through natural language.

Decisions Stopped Disappearing

Back at MacroBaby, the first problem we pointed the system at was still waiting, the stroller and car seat, dozens of options in that single store alone. Open our workbook today and that row reads Decision Made. Next to it sits a store-test checklist the project generated, a short list of things worth checking with our own hands before the final purchase, written for the exact aisles that started all of this. A piece of the weight from that couch now has a plan attached, and the plan came out of the same back and forth that made the decision.

The same workbook holds 190 checklist rows, every product and preparation task we have identified so far, each one carrying a status, a priority, and a timing. Forty-nine of those rows are still marked as needing research, which keeps me honest about what this is, a live build in the middle of a life change with plenty left unfinished. All the debating over options happens inside the project, and only 21 items have survived it far enough to reach the public registry.

Early on, we learned that baby decisions chain together in ways nobody had warned us about. The car seat we pick narrows the strollers that pair with it, and the crib we land on shapes the furniture that can sit around it. The workbook keeps those connections on the page, so a choice we lock in one category narrows the honest options in another.

Our crib search proved the point. We walked in wanting one specific furniture brand, and the conversation surfaced what we actually required, the safety features and the material quality underneath the label. Once we knew that, the search opened up, and the option we chose cost less with every requirement still met. Those requirements now sit in the record, quietly steering furniture questions we have not even asked yet. Every decision we finish makes the next one easier, because the reasoning gets written down before it can dissolve into chat history.

Even now, the system has seams we live with. Moving a finished decision from the workbook to the public registry is manual work, and a decision made in conversation can get lost if nobody writes it back. None of it replaces the professionals either, the pediatrician's guidance or the certified check on the car seat installation. Every final call stays with the two humans in the project. What the system took off our plate was everything around those calls. It gave us time back by handling the tedious part, and time back was the whole point.

Your Friction Point Is the Spark

You have a friction point of your own, and it probably has nothing to do with a baby. Maybe it is a move you keep postponing because the logistics feel bottomless. Maybe it is an aging parent's care spread across phone calls and paperwork, a household budget living in four apps and two memories, or a school decision arriving faster than your research is. Whatever yours is, you already flinch a little when it comes up. That flinch is worth more than every list of AI use cases you will ever scroll past.

Take that problem to the AI you already use, whichever one it is, and describe it in plain language. The constraints, the people involved, and the small decisions hiding inside the big one all belong in that first conversation. Skip the deadline and the project plan, because the exploring is the build. A few exchanges in, the first piece of your system will show up on its own. From there you are working toward something just for you, a build that solves your specific, personal problem and nobody else's.

Underneath our build sits a pattern that travels. Shared context, structured records, current research, a few decision rules, and your own judgment sitting on top of it all. None of those five pieces requires a technical background, and the judgment is the only one that cannot be delegated, which means you already own the hardest part. That is the whole distance from user to builder, and the walk is shorter than it looks.

Our baby is due later this year, and I built the system for the months in between, so I can support my wife and enjoy as much of the pregnancy phase as possible. What it really bought was presence. Somewhere in your life, a problem is waiting to hand you the same thing. Go find it. No code required.

The Assumption

For most professionals, AI arrived as a chat box. You typed a question, a paragraph came back, and the lesson stuck. This is a tool you talk to in words. Because it takes in text and returns text, using it well seems to mean writing your problem down. That belief matches how the tools were introduced and how most people still use them, and for a long stretch it held up.

Where It Breaks Down

Dana manages operations for a regional care network. A twelve-page vendor agreement sits in her inbox as a scanned PDF, and she needs the escalation terms summarized before a two o'clock call. Working from her assumption, she starts retyping the key clauses into a chat window so the AI will have text to read. A newer colleague asks why she is retyping anything and drags the scanned PDF into his own assistant instead. Before Dana finishes her second paragraph, a clean summary comes back. The scanned page had no copyable text, yet the model summarized it directly.

What Is Actually Happening

Back in Vol 6, we established that a model does not read your words the way you do. It breaks your text into tokens, the small pieces it actually works with. Then came the step most people miss. Those tokens no longer have to start as text.

In practice, a modern model can turn a photograph, a scanned page, an audio clip, or a frame of video into tokens too. A vision component, for example, slices an image into a grid of small patches and turns each patch into a token it can process alongside language. Once Dana's scanned agreement becomes tokens, the model handles it with the same machinery it would use for any paragraph you typed. Whether the input began as an image or as copied text stops mattering the moment it is converted.

The word multimodal describes exactly this capacity. A mode is simply a type of input, such as text or images. Put those together, and a multimodal model is one that takes in more than one type and works across them in a single request. Think of tokens as a shared internal currency. Whatever the model takes in, whether words, a photo, or a sound, is exchanged into that one currency first and handled the same way from there. That is why you can hand it a picture of a page and get back a clean summary, or send a document, a chart, and a question all at once.

The Revised Mental Model

A model does not read text. It processes tokens, and text is only one of the inputs that become them. That is the rule worth carrying forward.

With that single shift, your habits change in concrete ways. You no longer have to transcribe before you ask. Hand the model a scanned form, a screenshot of a dashboard, or a photo of a whiteboard directly, instead of retyping it. Your sense of usable input widens too, since a diagram or a faxed page becomes fair material rather than something to convert first. Requests can span formats as well, so a single ask can combine an uploaded image with your typed question.

What to Watch For

• Accepting an image is not the same as reading it perfectly. Dense tables, small print, and handwriting are where a model slips most, so verify anything that matters.
• Input quality shapes output quality. A clean, well-lit scan gives the model far more to work with than a skewed or blurry photo.
• Not every tool or tier is multimodal. Some accept only text, some take images but reply in text, and a few handle audio or video. Confirm what yours does before relying on it.
• Sensitive data stays sensitive in any format. A screenshot or scan can carry protected information just as pasted text can, so only upload material your organization has cleared for these tools.
• Multimodal on a product page can mean very different things. Treat the label as a reason to check specifics, not a promise that every format works well.

How This Connects

This section rests on Vol 6, where we established that a model works in tokens rather than words. Multimodal extends that idea, since those tokens can now come from images, audio, and documents, not from text alone. Next week, Part 2 goes inside the image to trace how a photo becomes tokens, and where a model's vision holds up and where it quietly fails. Part 3 then follows one professional through a week of real tasks, so you can see where multimodal input changes what is possible.

Part 1 of 3 in the Multimodal AI series.

Stop Fighting PowerPoint Shapes

Some ideas land better as a picture than as a paragraph. You can see the whole process in your head, every step and decision branch. Written out as three dense paragraphs, it still leaves a colleague asking you to walk them through it. Turning that text into a clean diagram no longer takes design skill or an afternoon.

Napkin.ai reads written text and draws the diagram for you. You provide the words, and it builds the shapes. It takes over the slow part, the nudging of boxes and arrows until they line up, so you can stay focused on the content.

The trick that makes it work well is two steps instead of one. First you ask an AI model to turn your rough notes into a clean structure. Then you hand that structure to Napkin and let it draw. The model does the organizing, Napkin does the rendering, and you make the final call on which version is right.

The Workflow

1. Structure Your Notes With AI (3 Minutes) Open Claude or ChatGPT and paste your rough description of the process, as messy as it is. Ask for a clean outline. Strong structure is what produces a strong diagram, so this step earns its three minutes.

Copy/Paste Prompt: "Here are my rough notes on a process I want to diagram. [PASTE YOUR NOTES]. Rewrite them as a titled, step by step outline. Start with a short title, then number the steps in order. Wherever the process branches on a decision, write the decision as a question and show both the yes path and the no path underneath it. Keep each line short. Give me only the outline, no commentary."

2. Paste Into Napkin and Generate (3 Minutes) Go to app.napkin.ai and sign up free, no card needed. Choose Visuals, then "By pasting my text." Drop in your outline, select the text you want to picture, and Napkin generates diagram options in seconds.

3. Pick the Layout That Keeps Your Logic (2 Minutes) Napkin often opens with a mind map, which looks clean but can leave your decision branches floating without their outcomes. Open the AI Suggestions panel and choose a flowchart layout instead, the kind where every yes and no leads somewhere real. Give it one read and confirm each branch connects.

4. Refine and Export (2 Minutes) Adjust the style to taste, then export. The free plan gives you unlimited PNG and PDF. Pick the version that reads clearly at a glance, save it, and drop it into a doc or deck.

The Payoff

In ten minutes you walk away with a clean diagram of your process, no design background required. The bigger win is the move behind it, structure first and generate second. It works again the next time something wants to be visual, a decision tree, an onboarding flow, an org chart you would rather generate than redraw. When an idea reads better as a picture, you now have a way to make one.

The AI Concept You Just Used

Diagram generation is its own AI capability, separate from image generation. An image model paints pixels. A diagram model reads meaning and arranges structure. Once you can feel that difference, you match each job to the tool built for it instead of forcing one model to do everything, and that instinct is a quiet marker of real AI fluency.

Transparency & Notes

• Napkin's free plan gives you 500 AI credits per week, at roughly one credit per word you select to generate. PNG and PDF export are free. SVG and PowerPoint export, plus removing the small Napkin branding, come with the Plus plan at $9 a month. Creating and editing work on desktop only. One tip from testing, the sketch style exports as a flat image with no searchable text, while the cleaner styles keep the text selectable.
• The layout choice matters most for any process with decisions in it. A flowchart layout keeps each branch tied to its outcome, so pick that one and give the logic a quick read before you export.
• Keep it general. Leave confidential, contractual, or patient details out of both the AI model and Napkin. Structure the shape of the process and leave the private specifics aside.
• Napkin rewards clear structure. If a diagram comes out thin, a cleaner outline in step one almost always fixes it faster than switching tools.