Volume 43: The Terminal Teaches Faster Than the Tutorial Does

The same terminal window sits open at home and at work, and last week it asked for two very different kinds of trust. At home I approve without thinking. At work, the same request stopped me cold, and taught me more than months of easy reps.

🧭 Founder's Corner: Argues that claiming the future of work means choosing discomfort on purpose, one small act of courage at a time, before anyone hands you permission or perfect timing.

🧠 AI Education: Shows that the same meaning-matching idea from Vol 5 already runs quietly underneath Copilot, Claude, and NotebookLM, and where that convenient shortcut can quietly mislead you.

✅ 10-Minute Win: Walk away knowing the two moments, before an agent starts and after it finishes, where ten minutes of your attention turns a generic report into one you can actually trust.

Let's get into it.

1) New York becomes the first state to pause new large-scale AI data centers

Summary: Governor Kathy Hochul signed an executive order on July 14 pausing permitting and construction of new "hyperscale" data centers, those drawing 50 megawatts or more, for up to a year while the state studies their effects on the power grid, water, and utility bills. It is the first statewide data center moratorium in the country. 

Why it matters: The real bottleneck for AI is turning out to be electricity and community goodwill, not just chips and models. Expect where and whether AI infrastructure gets built to become a live political fight in more states, which shapes cost and availability for everyone downstream.

2) Corner Health raises $32.5M for AI that lets nurse practitioners run their own clinics

Summary: Corner Health raised $32.5 million to expand Cora, an AI system that automates the back-office work of a primary care practice, including scheduling, billing, referrals, lab orders, and prior authorizations. The company says this lets about 90 percent of its clinics run with no extra staff, freeing nurse practitioners in Arizona and Washington to see patients independently.

Why it matters: More than 100 million Americans lack a regular primary care provider, and this is a concrete case of AI removing the administrative overhead that keeps small practices from existing. The pattern travels well beyond medicine: when AI absorbs the paperwork, one skilled person can run what used to take a whole team.

3) FDA clears a first-of-its-kind AI that catches ventilator breaths a patient fights

Summary: The FDA granted first-in-class authorization to Syncron-E, software from Autonomous Healthcare that analyzes a ventilator's pressure and flow waveforms to flag "ineffective efforts," the common and easily missed cases where a patient tries to breathe but the machine does not respond. The clearance, announced July 14, creates an entirely new device category, ventilator waveform analysis software. 

Why it matters: Severe patient-ventilator mismatch is linked to worse ICU outcomes, and it is hard for a respiratory therapist juggling many patients to catch by eye. This is AI covering a narrow, high-stakes gap in human attention, and the new FDA category it opens will shape how similar bedside tools reach hospitals.

4) Half of enterprises have already had an AI agent security incident

Summary: A VentureBeat survey of 107 companies found that 54 percent have already had an AI agent security incident or a near-miss, yet only about a third give each agent its own identity, and roughly two-thirds let agents share login credentials. The split showed up in the outcomes: organizations that shared credentials were hit about 64 percent of the time, versus 41 percent for those that gave every agent its own scoped access. 

Why it matters: Companies are handing agents real access to systems faster than they are building the controls to contain them. If you are deploying agents, the lesson is specific and actionable: give each one its own scoped identity and stop sharing credentials, because that single habit tracks with who gets breached.

5) AI's biggest labs get a report card, and the best grade is a C-plus

Summary: The Future of Life Institute released its 2026 AI Safety Index on July 7, an independent panel's report card grading nine leading AI companies across dozens of safety measures. Anthropic scored highest with a C-plus, OpenAI and Google DeepMind earned a C, and xAI, DeepSeek, and Mistral failed outright. No company earned an A or B, and reviewers found several labs quietly walking back earlier safety commitments. 

Why it matters: The point is the independent scorecard, since AI companies grade their own safety everywhere else. When even the top of the class earns a C-plus, it is a nudge to check a vendor's outside reviews, not just its own safety page, before trusting it with work that matters.

The same black window sits open on my screen at home and at work. At home, I move through it fast. I set up the build, approve each change as it comes, and let the commands run because I trust the rhythm I have built inside it over months of reps.

At work, that rhythm was gone. Before I started, I had set the tool to ask my permission before it changed anything, and that setting was the only reason I caught what happened next. I was laying out the architecture for a new build, setting up folders, when the tool asked to start making changes on its own. Instead of approving it the way I would have at home, I stopped and read up on what that permission actually granted before letting a tool reach that far into a machine that was not mine.

Nothing about the tool had changed. The responsibility had.

The Lessons Ran Backwards

My employer recently opened access to the same AI coding tools I use at home, and I partnered with my engineering team to get set up. I work in a regulated industry, so the access came wrapped in rules, and the rules deserve respect. Building with company resources means governing yourself harder than curiosity ever demands. When the tools and the tokens belong to someone else, that governance is simply the right thing to do.

The version of Claude Code I had access to lives only in the terminal, and that constraint turned out to be a gift. It forced me to ask my engineering partners different questions, better questions, about how to set myself up to succeed. Those questions produced a folder structure on my machine and a project setup I could carry from build to build, one that other tools like Cursor plug into cleanly.

Then the constraint taught me something about my own garage. For years I had built cloud-first. Nearly every project I ran at home lived in Claude Projects, ChatGPT Projects, or GitHub, and my code went straight up to the cloud without any structure underneath it on my own machine. The cloud was my source of truth. The terminal forced the opposite, local files as the ground I stood on with AI layered on top, and that different foundation made me think and build in ways my own projects never had.

Terminal Only, On Purpose

One week of those lessons was enough to change how I build at home. I downloaded Claude Code in the terminal for the first time, on my own machine and on my own time. Then I made a commitment to myself, terminal only, no matter what I was building. No wrapper, no familiar buttons, no escape hatch back to the comfortable version.

I made that commitment for more than the discipline it would build. I believe the work of traditional engineers and knowledge workers is blending together, visible today in small pockets in any industry outside of tech. People who have never touched a code base will describe what they need in plain language, shape working prototypes with AI, and hand them to engineers to deploy. The skills companies hired for five years ago already look different from the skills the next five will reward.

Underneath the shift, the frameworks hold steady. Identify problems, build the solution, iterate, move fast, sell your vision, and drive measurable results. How the work gets done is evolving in real time.

Hold that belief, and the timing answers itself. The future of work is here, waiting for workers to claim it. Nobody schedules that moment for you, and no employer hands it over in an onboarding packet. Discomfort has been my tuition since my first rep, and the terminal is where the current semester is held.

I had reached into the terminal once before, for a single build, and left it there when the project shipped. This commitment asks for something that first reach did not. Every build from now on starts in the same demanding place, whether or not an easier path is one click away. Choosing discomfort on purpose is the only way I have ever grown, and I would rather train for the future than read about it.

Building for One Is the Easy Part

The commitment rearranged everything behind it. My entire Google Drive folder structure was rebuilt around how these projects will grow over time. My production system runs on written plans and gates, and multiple projects run on auto-mode with my judgment saved for the checkpoints that matter. Each build is a rep, logged the way a runner logs miles. The pattern is the point, reps against a future I can see coming.

At work the same muscles are pointed at other people. I want to build tools that make my team more creative, the kind that take busy work and repetitive tasks off their plates and unlock time for the work that actually matters. Giving my team time back matters more to me than any personal build shipping on schedule.

Every tool I build at home serves an audience of exactly one. That is the harder truth underneath the ambition. Nobody asks how access works, nobody needs training, and nobody has to trust my code but me. The moment I want a teammate to use something I built, the difficulty changes species. Deployment, access, permissions, and other people's trust turn a working tool into a shipping problem, and I am still early in learning how to solve it.

The real work now is building something a teammate can trust, and every private rep at home is practice for it.

Claim It With One Command

I do not know how the world shifts from here. I cannot tell you how quickly employers will integrate AI, which tools will win, or which skills the next hiring cycle will reward. I have stopped needing to know. I am building for the end state no matter how we get there. The reps travel, and so does the judgment they build. Whatever the future of work turns out to look like, the person who has been training for it will recognize it first.

If any part of this sounds out of reach, I want to hand you the smallest possible beginning. You do not need a project or a single line of code. You need the terminal, a program already sitting on the computer in front of you. On my Mac it is a plain dark square called Terminal. Open yours, type curl parrot.live, and press Return. A parrot made of keyboard characters will start dancing across your screen, streamed to you by a stranger on the internet for no reason except joy.

When the dancing is done, press Ctrl+C and the stream stops. That is the whole exercise, thirty seconds from start to finish, and you controlled both ends of it. You started something in the terminal, watched it run, and shut it down on your own terms. Every build I have shipped began with exactly that much courage.

Do not fear the terminal. The future of work is already running inside it, and it answers to whoever shows up. Claim it.

The Idea You Already Own

Back in Vol 5, embeddings arrived as a budgeting trick. Charges labeled "FitCo Monthly" and "Health Club Charge" landed near each other on a map of meaning, and an app could group them without matching a single word. Similar things sit close together, and closeness is meaning.

That map was not a teaching device. It was a preview. The same primitive now runs quietly underneath most of the tools you opened this week, and once you notice it in one, you cannot stop seeing it.

Same Idea, Three Products

Ask Microsoft 365 Copilot a question about your own company. Before your prompt reaches the model, your emails and documents have already been turned into vectors and stored in what Microsoft calls a semantic index. Your question becomes a vector too, and the system returns the nearest ones. Microsoft describes this in nearly the words Vol 5 used, as placing "close numbers ... in proximity to one another to represent similarity," then searching across billions of them. The map that sorted your receipts now sits under the assistant you use at work.

Claude's memory runs on the same move. Ask what you discussed last month, and Claude does not recall it the way a person would. Anthropic says those searches use retrieval-augmented generation, the search-first pattern from Vol 20, which finds the past conversations closest in meaning and reads them back. What feels like memory is retrieval across a map of your own words.

Upload a stack of PDFs to NotebookLM and the pattern repeats a third time. Each source is split into passages and turned into vectors. Your question becomes a vector, the nearest passages come back, and the citation it hands you is, in effect, the closest dot on the map. Three companies, three products, one idea beneath all of them.

What You Could Not See in October

When Vol 5 published, the map looked like a way to sort receipts, nothing more. Since then this series walked through retrieval in Vol 19 through 22, agents in Vol 24 through 27, the Copilot stack in Vol 29 through 34, and Claude's memory in Vol 40. Each rested on the same foundation without announcing it. The concept never changed; your vantage point did. That is the quiet reward of a curriculum. What you first met as a toy turns out to be the ground everything else stands on.

Where the Map Misleads You

Powerful as it is, the map is not the same as truth. It finds what sits closest in meaning, which is not always what is correct. A system can retrieve the nearest passage with full confidence and still hand you the wrong one, the exact failure Vol 20 warned about.

Three cracks are worth watching.

Start with the source itself. Blurry scans, inconsistent labels, and half-written notes produce muddled vectors, and no clever search rescues bad input. Clean data still wins, the way it did in Vol 5.

Specialized language is the next trap. A meaning map built on general text can drop a clinical abbreviation into the wrong neighborhood, reading PA as physician assistant when your document meant prior authorization. In a healthcare workflow that misplacement is not academic, because it routes the wrong record to the wrong step.

Subtlest of all is permission. Copilot retrieves only what you already have access to, which sounds safe until you notice that the map faithfully surfaces anything overshared. Microsoft is explicit that indexing does not change who can see what. The map mirrors your governance, for better or worse.

How to Work With It

Two habits follow. Ask in plain language, which is what meaning search rewards, instead of hunting for the magic keyword. Then verify what comes back, because the nearest match is a strong guess, not a guarantee. When a tool cites a source, open it. The map takes you to the right neighborhood far more often than not, but you still confirm the address.

How This Connects

A Flashback is a once-a-quarter step back to see how far an old idea has traveled. Vol 5 handed you the map. Today you saw it running inside Copilot, Claude, and NotebookLM, three tools you likely used this month without once thinking about the machinery. Next week the series turns to multimodal AI, where the same move, turning content into vectors, stops being about text and starts working on images and sound. The map is about to get much larger.

A Flashback revisiting Vol 5.

Steer the Agent Before You Launch It

In ten minutes of your own attention, you can hand an afternoon of research to an agent and get back a cited report you actually trust. The agent reads dozens of sources while you step away. Your ten minutes go to the two things that decide whether the report is any good, telling it where to look and checking what it found.

Most people never touch either lever. They type a question, click Start, and take whatever comes back. It reads fine and rests on sources they never chose, with no easy way to sort the solid claims from the guesses. Steering fixes that, and it costs almost nothing. The workflow has four moves, and the skill lives in two of them.

The Workflow

1. Brief it like an analyst, not a search box (3 minutes)

A vague question gets a vague report. Give the agent a decision to serve, not a topic to summarize. Name the situation, the constraints, and what a good answer lets you do next.

Copy/Paste Prompt: "I need a research report to make a decision. The decision I am trying to make is [WHAT YOU ARE DECIDING, e.g. which of three vendors to shortlist, whether to enter a market]. My situation and constraints are [BUDGET, TIMELINE, INDUSTRY, ANYTHING THAT RULES OPTIONS IN OR OUT]. Prioritize [THE SOURCES OR ANGLES YOU TRUST MOST]. Ignore [WHAT IS NOISE FOR YOU]. Structure the report around [THE SECTIONS YOU ACTUALLY NEED]. Flag anything you cannot verify instead of filling the gap with a guess."

Turn on Deep Research and submit.

2. Edit the plan before it runs (2 minutes)

This is the move almost nobody uses. Before the agent browses anything, it shows you a research plan, the questions it intends to chase. Do not click Start yet. The plan is editable, so fix what it missed, adding back the competitor it left out, the angle it skipped, or the source you want it to lean on. You are redirecting a ten-minute run in about thirty seconds.

Tell it in plain language, the way you would correct a colleague. "Add a section comparing pricing tiers. Prioritize primary sources over blog roundups. Drop the general background, I already know it."

Now start the research and walk away. It works for five to ten minutes on its own time, not yours.

3. Interrogate the report, do not trust it (3 minutes)

When the report lands, the instinct is to skim it and drop it into a deck. Give it two minutes first. Gemini marks which claims came from a source and which it synthesized, so the synthesized ones and the load-bearing numbers are where you look before you trust anything. Then make the agent expose its own weak spots.

Copy/Paste Prompt: "Before I use this, audit your own report. Which three claims here matter most for my decision, and how strong is the source behind each one? Where did you synthesize or infer rather than cite a source directly? What did you fail to find, or find only weak evidence for? List everything I should verify myself before I rely on it."

4. Send it back once (2 minutes)

You now know where the report is thin. Close those gaps in one more pass instead of living with them.

Copy/Paste Prompt: "Run a focused follow-up on the gaps you just named. Strengthen [THE WEAK CLAIM] with a better source, or tell me a good one does not exist. Add [THE MISSING ANGLE]. Keep everything that already holds up and give me the corrected version."

The Payoff

You walk away with a report built on sources you chose and checked against its own soft spots, aimed at the decision you actually face. The tool did the hours of reading. You did the ten minutes that made it worth reading.

The AI Concept You Just Used

This is agentic delegation, and the skill in it is bookending. An agent runs on its own once you start it, so your control sits at the two ends, the plan it follows and the output it returns. The same pattern holds the day you point an agent at your calendar, your inbox, or a stack of code. The tool changes. The two places you stay in the loop do not.

Transparency & Notes

• Gemini Deep Research is free, currently around five reports a month on the standard model, with paid tiers raising the cap. On a small free quota, the steering matters more, not less, since you do not want to spend a run on a vague question.
• The same pattern works in ChatGPT Deep Research, Perplexity, and Claude, even where the plan step looks different. Steer first, verify last, whatever the tool.
• Keep confidential or employer data out of any agent that also browses the open web. Strip the sensitive details first, or keep the question public.
• The ten minutes is your hands-on time. The agent's own run of five to ten minutes happens while you do something else, which is the point of handing it off.