The point is simple: There is far more under the hood than you might know and features are added and change every week. It is important to stay up-to-date.

I am hoping by now that your workflow of working with ChatGPT is NOT this:

• Open ChatGPT

• Type a question

• Copy the answer

• Close tab

If it still is, we will change that TODAY.

The progression of ChatGPT

In 3 years, since chatGPT arrived, it went through 4 main stages:

If your mental picture of AI is CHATBOT where you ask → it answers → you copy paste, You are missing on using CONNECTIONS and AGENTS, where the real power of AI is.
Something I will touch on later that is very important: More capability does not mean less responsibility.

The twelve things

Subscribe now

1. Three doors, one Tool.

   3 months ago, OpenAI merged its coding agent, Codex, into the Desktop App under the argument that: What it means to be technical is very fluid right now.

   • Browser (chatgpt.com)

   • Phone App

   • Desktop App

2. Desktop versus Browser

   Is there a real difference?

   Yes!

   The Desktop App can see your computer. The browser cannot.

   In the Browser you upload the file.

   On Desktop you say “Look at my Desktop and sort my screenshots into folders” and it does it realtime for you - you just watch it happening:

   

   OpenAI puts its newest features on desktop first. Install it.

3. Chat versus Work

   Appears like a toggle at the top. Two modes: Chat and Work

   

   

   Do not use Work for quick question - you will wait long and it will burn tokens for nothing.

   Do not use Chat if you want a finished deck - you might spend an hour copy-pasting.

   If you start in Chat and realize you want the deck after all, it now switches to Work by itself and keeps the context.

4. Custom Instructions

   In Settings → Personalization: Paste your role, your company, what your team ships, how you like the answers. Every chat or task completion starts by knowing this, so you don’t need to re-explain.

   You can pick a personality too: Friendly or Pragmatic.

   

5. Plugins

   Top left of your Desktop App screen:

   

   This is how ChatGPT can reach your Gmail, Calendar, Drive, Outlook, Slack, etc.

   A Plugin is 2 things:

   1. The connection to the App

   2. Plus a Skill

   A Skill is a recipe: instructions on h ow to use that app well. It logs in as you, with your permissions. If your company has an internal tool that is not in the library, a technical colleague can wire it up with MCP (a standard plug for connecting AI to any system).

6. Computer Use Plugin

   Turn it ON in Plugins is your machine allows it.

   

   The plugin allows ChatGPT to read local files and take over your browser.

   On a Mac, when you press both Command keys, an "Appshot" of whatever is on screen drops straight into the prompt. No screenshot, no upload.

7. Permissions

   Inside Settings → Configuration.

   Make sure that you choose the least access that finishes the job and that you can audit.

   After-all, you would not give your 8-year old a device if you cannot keep up with what the device can do.

8. The @ Symbol and the + Menu

   Type @ to call a specific plugin ("@Google Drive, open the leadership notes"). The + menu next to the prompt box hides most of the features people never find. Click it once:

   

   

9. Preview and Annotate

   Ask Work for a 5-slide deck from 2 Drive documents.

   

   3-5 minutes later it appears inline.

   

   Select slide 2 and select Annotate (the + icon) option.

   

   It will highlight blue and change into in “Annotating”:

   

   Select the section that you want to modify and type: “Rewrite this for a CFO”.

   

   It edits in place, keeping everything else unchanged:

   

10. Projects, Pins, Branches

    A Project is a folder. Chats and files that belong together, in one place, with shared context. To create a Project, click an the + button near Projects:

    

    Pin the ones you use daily.

    

    Three storage options: local, cloud (also shows in the browser), remote (continue on your phone). Any chat can be branched: same context, fresh thread.

11. Voice, and it is NOT dictation

    

    Dictation types what you say.

    Voice is a conversation.

    A month ago, the Voice model got a large upgrade and, on Desktop, Voice can now Work things.
    Try it: Ask out loud for a “60-second overview of how big banks use AI”, then tell it to "put it in a Google Doc for my boss". The doc appeared, with sources. Hands on the steering wheel, nothing else:

    

12. Scheduled automations

    The one that changes your week. Any Work task you have tested and trust can run on a schedule.

    Demo: "Look at my calendar, find me a one 90 minutes every week to learn Python".

    

    Done once. Runs forever. Push notification when it finishes.

    Top left, Scheduled tasks →Create, and there is a starter prompt that walks you through it:

    

13. Bonus: Pets and Sites

    Pets (inside Settings→Pets): a small character on your desktop that wakes up when a background task finishes. Silly. Also the reason you remember the 20-minute job you started.

    • You can upload a photo of your own pet and make it an Icon. Here is Miss Ziggy, my pet, created from a real photo of my cat:

      

    

    Sites: Pages the tool builds and hosts, refreshed by a scheduled automation.

    

    

    It created this:

    

    
    

    OpenAI executives now ask for account updates as a site, not a deck. Someone from Disney said their new leader declared "dashboards are dead". Sites are what we need.

If you did any of the Steps 1-13, let me know what you think, what you discovered? Also, how do you use ChatGPT in general?

Leave a comment

As AI moves from helping us think to doing the work, what do we keep?

I am impressed by Clay, whose CEO Varun Anand published a company writing policy last week. One line: Writing should take longer than reading. You must be able to stand behind anything you produce.

My rule, and the one OpenAI’s finance team uses:
Hand over what you can describe and check. Keep what you cannot check yet.

Forward this to the colleague who still thinks ChatGPT is a search box.

I help ambitious and smart professionals (scientists, engineers, tech people, and not only) get their excellent but invisible work seen, paid, and promoted. If you are the only person on your team using stages three and four, that is a visibility opportunity. Let’s talk.

Teodora

Two questions to start with up front:

• How many AI projects did you (or your team) start this year?

• How many went into production?

• How many run in real clinic, with real users, right now?

Last week I was in a conversation with Jerry Discher, VP for Google Ads at Google for 20 years, until 2023. As you know, Google Ads is the engine behind hundreds Billions of dollars in revenue.

Jerry and his team developed Performance Max at Google. Before Performance Max, advertisers ran their campaigns manually: picking keywords, setting up prices, designing the ad. Performance Max uses AI to simplify everything - so you just have to tell it: “sell more shoes”. The launch date was November 2021. You can see some case studies here.

What makes this interesting is the fact that Jerry and the team had the Performance Max product idea 9 years earlier. The working concept sitting on the table since roughly 2012. Why he waited 9 years?

Here is what Jerry told us:

"AI systems are not about storytelling. AI systems are about quality. We only released Performance Max campaigns when we felt that the technology had progressed to the point where we were able to deliver the user story."

The technology worked since 2012, but not reliably enough. Getting it right most of the time sounds good until you are the advertiser paying for the misses. Google waited until the misses were extremely rare.

Trust is the product

The resistance did not come from doubting the AI. It came from somewhere deeper. Advertisers had built careers on knowing which keywords to buy and how much to pay for them. That expertise was who they were at work. The AI was not threatening their results. It was threatening their identity.

When COVID hit, the resistance shifted. How it shifted? Well, when you set prices by hand, there are many assumptions that you make, such as: “1 in 10 people who click the ad will book a room. So each click is worth $1 to me”.
COVID broke the math in a single week. Even if people were still clicking in the ads, they were not booking rooms. So every manually-set price was wrong.

Advertisers who had switched to AI pricing were fine: the system noticed bookings collapsing and cut prices automatically, within seconds. Advertisers still setting prices by hand were flying blind. The people who had given up control were the ones who survived.

We do not want to release products that are just “good enough to ship”. Ask the team to show you the evidence, how was it tested, on what inputs, where it fails, how often. If the room gets quiet, this is a career opportunity signal. As AI moves into products people actually depend on, we need engineers that know how to prove the quality. Evaluation depth is not the boring part of engineering, it is the key to great products.

The engineers who know how to prove quality, not just achieve it, are the scarcest people in AI right now.

The quality gate: Run it on your own project today

• What is “the right thing,” in numbers? Define the metric and the threshold before you look at/produce the results. We don’t want results that only “look good”.

• What is your failure rate on the hard cases (hard slice)? Not the average. The subgroup where your system is weakest: the rare disease, the noisy probe, the edge-case customer. Averages hide the cases that end trust.

• What happens on failure? Silent wrong answer, flagged uncertainty, or graceful handoff to a human. Only one of these is acceptable in a real world workflow, and it is not the first.

• What breaks when the world shifts? COVID broke manual bidding in a week. Name the assumption your system quietly depends on: patient population, scanner vendor, data drift. Write down how you would detect the shift.

• Can a skeptic verify your claim without you in the room? If your evidence needs a live walkthrough, it is a story, not evidence. Reporting built the trust for Performance Max. It builds yours too.

Testing on the hard and edge cases is where many teams fail, especially with AI. Most often, teams have access to the same type of data that they train their models on. So, if the tests use the same data, the results will come back excellent. But the real-world data is messy, and your customer will have that real, rare data. Your great test score, will say nothing about this rare, messy data. That is a blind spot.

Demos are storytelling. Deployment is quality. The nine-year wait was not slowness. It was the product.

I help ambitious and smart professionals (scientists, engineers, tech people, and not only) thrive in their careers. If you are deciding whether regulatory and evaluation depth is worth your next two years, that is a conversation I have weekly. teodora.coach

If the quality gate caught something on your project, restack this for the teammate about to ship the demo.

Teodora

They come to the job, but you do not provide them a computer. No files to find, no real data, no desk. They can only answer general questions.

I see a lot of people still using AI this way: open a tab, type a question, maybe add a file or two, copy the answer, close the tab. They are using AI as a chatbot. But a chatbot only predicts the next word.

Today we will be using AI as an agent. An agent pursues a goal.

The difference between an real agent and a chatbot is what the AI has access to: a computer, folder of files, and an objective.

And guess what? That is how a real data analyst would also need to do a good job. And you will see them keep track of their own work, write files down, build tools, and come back to you with a dashboard.

Today we run the whole loop and learn how to turn messy CSV files into executive dashboards, without you writing a single line of code. But by talking in the right way with ChatGPT.

The paying subscribers will also get the review technique that catches the overconfident claims before the CEO does. Also, how would you run this inside a regulated clinical environment without exposing a single patient record.

Step 0:

Install ChatGPT Desktop App. This is the step where you give your data analyst a computer. Yeei! It can open your folders, write files, run code.

The install one plugin. Open Plugins in the top-left, search “data” and install Data Analytics:

What is a Plugin?
It is a set of skills. If you click in the Data Analyst plugin, you see that it has 15 skills:

A skill is a preset prompt (sometimes 3-page long) written by experts, teaching the agent exactly how to do the Data Analyst job. Basically it takes a generalist to be a data analyst. KPI reporting, dashboard building, data validation.

If you click on one skill, you will see the prompt:

Two settings before we start:
1. Pick the highest-tier model (I am using 5.6 Sol here). For analytical work, we want intelligence instead of speed.
2. Set reasoning effort to medium or high. We want correct answers over quick ones.

The Data Analyst Loop

5 steps. I run this loop on research cohorts, on Substack growth data, on anything tabular that lands on my desk. The plugin gives the agent its recipes. The loop is how YOU run the agent.

Real World Dataset

Download the Medical Appointment No Shows dataset from Kaggle: 110,527 real appointments from public clinics in Brazil, de-identified, free. You can just download the dataset as .zip file.

If you have another dataset you would like to work with, that would be ideal. I picked this dataset because it is painfully REAL: misspelled column names, a patient with age “-1”, appointments scheduled after they happened.

Your agent has to find all of it. That is the point.

Let’s formulate now the business question we want to answer:
What drives the no-shows, and is expanding SMS reminders worth the money?

Hold that question. It hides a trap and that is exactly what Step4 exists for.

Step 1: Explore. Do not let it analyze yet.

Let’s create a Project called “AI Lab Data Analysis” by clicking the “+” button in the Projects section. It will look like this:

Make sure you create a folder inside your computer and link it as a folder source:

Think of the folder as the analyst's desk: everything on the desk is in scope, everything off the desk does not exist.

Here is the first prompt - I suggest using Voice. It is much faster than typing.

First prompt:
Explore this folder. Tell me what files and data you see.
Describe them at a high level: rows, columns, data types, how the files relate to each other.
Do NOT begin any analysis yet.

You want it to look at the columns, sample a few rows, and report back, exactly like a human analyst on day one on this project.

Here is what I got:

Step 2: The reverse-interview. The step that changes everything.

Most people prompt AI as:
” I want X, do X for me.”

When the output is wrong, the problem is not the agent. The problem is that you do not know what you want yet.

Make the agent interview you.

The prompt:
Before we begin: interview me. Ask me every clarifying question you need about:
(1) the business question and the decision this analysis will drive
(2) how each metric is defined
(3) the time window and comparisons that matter
(4) whether to treat the data as clean or validate it first
(5) who the audience is and what output format they need.
Do not start the analysis until I have answered.

I got back 11 questions:

Do not type your answers. Dictate them. You type at 60 words per minute. You speak at 150.
Your job is to get the full problem out of your head in its rawest form, not to craft a beautiful prompt. Prompt engineering matters for prompts you reuse every day. For one-off analysis, context volume beats prompt polish. Every time.

It took me about 30 minutes to formulate and dictate the answers.

After I answered to all 11 questions, it asked me some clarifying questions:

Here is what I got after this step (together with 2 Python files: analyze.py, validation.py) and artifact.json file that I can review as well:

Step 3: Analyze. With one non-negotiable.

Let’s begin the analysis:

The prompt:
Validate the data before analyzing it. Check for duplicates, inconsistent categories, missing values, and unit mismatches. Report what you found and what you fixed, and ask me before making any judgment call on ambiguous cases. Then produce an analytical report for clinic leadership.

When the agent shifts its state from chat to analysis, it shifts from prediction to code. Code is deterministic. The presentation of the code might vary, but the results are the same if you run the codes many times.

Here is the validation report I got:

There are some impossible records and a great question. I think age 0 can be valid in case of pediatric visits, so age 0 would mean infant. For handicap, I will prompt it to use binary “any handicap” (0 versus 1+).

Then, I got the first version of the report. You can see it here: https://teodora.coach/noshow-demo

Step 4: Adversarial review. And the trap that a junior analyst will fall for.

If you ask the agent if the analysis looks good, it will tell you that is excellent - the same way an analyst will say about his/her own work. Also, we will talk about the trap I mentioned at the beginning of the article.

What should we do instead?

Read more

Your answer probably was GPUs. NVIDIA became the most valuable company on the planet because everyone’s answer to this question is: GPUs.

Let me surprise you: The fastest supercomputer in the world (LineShine, National Supercomputer Center, Shenzhen, China) does not contain a single GPU.

I spent more than a decade running simulations and AI on supercomputers. My world is still compute nodes, job queues - but for clinical AI this time.

In my previous article I wrote about one skill that survives AI: knowing right from wrong. I want to bring here the hardware side of the same story.

Subscribe now

The electricity bill, in real numbers

Here is the arithmetic that puts the AI data center race in perspective:

This is the energy bill. Without including the bills for training the models, before a paying the engineers, buying the compute chips.

The new AI data centers are announced in Gigawatts.

A $1 Billion annual power bill turns efficiency into huge cost savings:

• Quantization that cuts a model from BF16 to 4-bit cuts the inference power bill with it.

• Distilling a large model into a small one that clears the same evaluation is direct $.

• Every millisecond removed from inference is watts removed at data center scale.

If your work touches any of these, you are not doing optimization as a hobby. You are working on one of the largest cost lines in modern computing (say the savings amount in your next performance review).

The machine that made me question GPUs

The United States restricted chip exports to China. The plan was to slow them down.

Instead, China innovated around the restriction. Their engineers took the one mathematical operation that sits underneath all of deep learning (multiplying two matrices together) and baked it directly into the CPU, with very fast memory sitting right next to it. No GPU anywhere in the design.

Now, this machine, LineShine, takes the No. 1 Position in the TOP 500 Supercomputers in the world.

Why does removing the GPU help?

Because of the first law I ever learned on a supercomputer: moving data is one of the most expensive things a computer can do.
A GPU is a separate device. You write two programs, one for the CPU and one for the GPU, and you ship data back and forth between them, paying the toll on every single trip. I spent entire weeks of my imaging career fighting exactly this tax.
The processors are not the bottleneck when dealing with large data. The distance is. Put the math inside the CPU and both costs vanish. One processor. One program. No toll booth.

So, I am no longer sure the GPU era is permanent. The most powerful machine on Earth just proved you can win without them.

Quantum will never replace your laptop

Every few weeks someone asks me about quantum computing, usually with the assumption that it will replace everything.

It will not.

Quantum is a genuinely exciting research area, and today only a handful of algorithms benefit from a quantum computer at all. A handful.

Meanwhile companies are selling quantum machines like the revolution already happened.

The Supercomputer of the future is not quantum instead of everything.

It is a team: CPUs, GPUs, quantum processors, neuromorphic chips that compute the way neurons spike, even optical devices that multiply data streams at the speed of light.

Each problem gets shoveled to the device built for exactly that problem.

What this means for your career

Put the three facts side by side:

• A $1 billion power bill rewards whoever wastes the least.

• The winning machine cut waste by specializing the processor.

• The future machine sends each problem to a specialist device.

The same logic is arriving in the job market. AI is the general-purpose processor: cheap, available, good at average problems.

The valuable roles are the specialist ones, where the hard problems get sent to a specific person by name, because their colleagues know exactly what that person is the specialist in.

The bottomline: when AI can solve most average problems, you need to solve harder problems. And it helps if you are passionate about the problems you are trying to solve.

Bellow I put together 3 steps to help you define what you are specialist in right now.

The Specialist Audit

I am the person who knows what right looks like in cardiac ultrasound AI. Everything I ship, write, and develop runs through that sentence. If you want help compressing yours, I keep slots open for video calls with subscribers. Reply to this article and we will find a time.

Leave a comment

Do not forget to share your specialization:

• Put the sentence from item 3 in your LinkedIn headline, your internal profile, your next introduction.

Specialists get routed the interesting (hard) problems only when the router knows they exist.

AI is the general-purpose device. Be the one the hard problems get routed to.

I help ambitious and smart professionals (scientists, engineers, tech people, and not only) thrive in their careers. If the audit surprised you, that is exactly where we would start.

Visit my coaching page

Teodora

Last week I was in a live Q&A with the one and only, Jack Dongarra, the 2021 Turing Award Winner (the equivalent of Nobel Prize for computing) mathematician and computer scientist. His work defines the linear algebra libraries integrated inside NumPy, MATLAB, and every model you ever trained.

I spent more than a decade running his code on different Supercomputers. And still do it in a clinical setting.

We asked him about the future of software engineering. And his answer was:

"The role of a person as a programmer is going to be diminished. If you are thinking about going into the area of programming, I would hesitate."

Subscribe now

So, what should we do instead? Dongarra came back to one idea, over and over, in three different answers:

“We should have some idea of what the answer should look like.”

He offered one example that I found very inspiring: Weather Forecasting.

The European Centre (ECMWF) now predicts weather with AI trained on decades of historical data. It is dramatically faster than solving the physics equations on a supercomputer.
But here is the part that might surprise you: ECMWF still runs the physics computation to verify the AI forecast:

• The AI produces the answer.

• The physics confirms the answer has the right shape.

• Two systems. One generates, one verifies.

So what survives AI? Not the code syntax. But knowing what right looks like, the plausibility check, or sanity range.

I see this every week in clinical AI: normal-looking numbers, right units, that measure the wrong thing. For e.g., I saw the AI model measuring the right ventricle instead of the left ventricle and getting the right-looking numbers, but on totally wrong data. The wrong-chamber mistake is visual and only human can spot this - or at least enforce it. Yes, we can add another classifier of right and left ventricles - but this will add new complexities.

Dongarra reminded us that: “In science, failure is NOT an accident. It is the default.”

The Reference Range Checklist (15 minutes, works for any AI output)

Make it a habit.
Before you accept anything AI produces (a number, a forecast, a report, a diagnosis, a paragraph), you write down what right looks like.
Then you hold the output against it.

Dongarra did not say that the technical careers are over. He said that the programmer role is diminished. What remains: humans that construct the test, set the guardrails, and know what the right answer should be before the machine produces it.

LLMs has read every program ever written. But it has not examined your patients.

The machine writes the answer. You own the reference range.

I help ambitious and smart professionals (scientists, engineers, tech people, and not only) thrive in their careers.

If this landed, the checklist is your homework-free win for today: three lines on paper before your next model run.

Subscribe now

Teodora

Leave a comment

Probably you got no extra pay for your extra work and sweat.

Here is a good explanation of why:

• Labor is linear: you trade hours for money, and hours are capped.

• Leverage is not: it is when systems, capital, or other people's time multiply your output.

I have a PhD, and over 10 years work experience. I do work very hard. In my previous job, my boss told me that I was doing the work of three scientists. So, part of me still believes this: do excellent work, and the system notices.

The system does not notice if you don’t speak up. Work is often very mute.

So if effort is not the lever, what brings the lever?

First, let’s talk about what does not work.

What does NOT move your lifetime income

Here are the traps that hard working people (including me) continuous fall into:

1. Working harder: 0% direct effect in your income.

2. Hopping jobs for changing a title: This one is slightly negative, because you never stay long enough to compound. (There is an exception: hot, overpaying sub-industries in a talent war (OpenAI, Anthropic, Google, etc.). Parts of AI right now. That window is narrow and it closes.)

3. Negotiating your offer harder. I do coach negotiation of the salary (longterm, strategic negotiation; I don’t recommend the 1-time negotiation). Yes, please negotiate - just know that it will not move the needle too much.
The reality: offer negotiation moves you 5% to 10%, because every company hires inside pre-set limits. Negotiation is not useless. It is aimed at the wrong target. Keep reading, because there is a target where you can extra 30% to 80%.

Labor economists have shown that pay differences between identical workers are driven substantially by which firm employs them (Song et al., QJE 2019), and that identical performance gets rewarded differently depending on how it is communicated (Exley and Kessler, QJE 2022). The system pays for position and signal.

What DOES move your lifetime income

Here are the 5 levers that focus on both position and signal.

Inside my coaching programs and through Standout Systems, these 5 levers are exactly what we are working on, every day.

Today we are going to focus on the first 2 levers, so you get actionable steps to apply to your career situation today.

Lever number 2. is the reason many people dream of working at big tech companies like Google, Meta, Amazon, NVIDIA. Here is why with a concrete example:

• A copywriter at a top ad agency, a dream employer, earns around $80,000. The same copywriter, same skill set, same job, at Meta: $250,000. Over a career that gap compounds to roughly $4 million.

• A pilot at a regional carrier versus a major: $100,000 per year difference for flying the same planes.

• A lawyer at a regional firm versus a national one, doing identical work: 2x.

Same work. Same you. Different firm. Different life.

So you could out-earn a decade of grinding by making one correct firm decision. And most of us made that decision based on brand prestige, not on pay data. I did too.

Part 1: The 20-minute firm audit (do it now)

You do not need to be job hunting to run this. You need to know where you stand. Open a new tab. This takes 20 minutes.

Read more

• Google DeepMind built a model called AlphaEarth. It observed a full year of our planet through: satellite photos, radar, heat readings, elevation, climate.

  

• Then it compressed all this information into 64 numbers for every 10 by 10 meter patch of Earth. Here is the fingerprint of 1- X 10 meters of Chicago Loop area:

  

64 numbers. No photo, no pixels, no words.
Smallest: -0.276 Largest: +0.236
Length of the vector: 0.999 <- always 1.000, by design

The model is propriety of Google. We (the ordinary non-Googlers) cannot download it.

However, Google DeepMind published every single number it produced. Free. For anyone. Sitting in a public catalogue since last July.

Google kept the recipe and gave away the whole cake.

Today, I want to share this cake with you. In this article, you are going to taste it, use it, and I have put together a free notebook you run in your browser and have some fun and ideas on how this can be used and useful to your domain. Also, this inspires me on how I could use a similar compression approach in medical field.

First, what is Google DeepMind gives us access to through AlphaEarth

Think of it like this:

• Every place on Earth now has a fingerprint: 64 numbers that describe what it is like.

Not a photo. A description. The interesting fast: no human decided what any of the 64 numbers mean. The AI worked them out on its own, by covering up bits of the data and learning to guess them back.

Which raises the obvious question:

• If AI decided what they mean, are they meaningful at all?

Let’s look at it by running the Notebook using https://colab.research.google.com/.

Subscribe now

The AlphaEarth Playground

Open the Notebook

AI concepts that are used in the Notebook

Step 6 in the Notebook is the interesting one

Steps 2 to 5 are fun. Here are some results you could get:

3 - Find the twin places of the world:

🏆 Most alike: Chicago Loop and Manhattan (0.85)

5 - Coloring dots by labels AI never saw:

Step 6 is an argument you can take into a meeting, so let me explain what it does in plain terms.

The expensive problem

In nearly every AI project, the slow, costly part is not the model. It is a human sitting there labelling examples by hand. This is forest. This is water. This is a defect. This is spam.

Thousands of times. That is the bottleneck, and it is where the budget goes and/or the model fails if the data is not labeled properly.

The question we test

If we describe our data using the AI’s ready-made fingerprints instead of the raw measurements, how many hand-labelled examples do we save?

Same model. Same places. Same rules. The only thing that changes is the description.

We start it very few: three labelled examples per type.
Then five, eight, thirteen, and up.
We plot how accuracy climbs for both.

On the right side the two lines come together. Of course they do. Give any model enough labels and it works things out eventually.

The gap on the left is the entire prize. That is the accuracy you get for free when labels are scarce, which is the situation on every real project.

The notebook draws an arrow on the widest gap and prints the number for you. It is 15% in this case.

Forget satellites. Here is the transferable bit.

This was never about maps. It is about a shift that is quietly happening everywhere, and most people have not noticed it yet.

Big labs are spending enormous sums teaching models to understand one type of data. Then, increasingly, they publish what the model produced, even when they keep the model itself locked.

Which means you can stand on work you could never have afforded to do.

Three questions, on every project you start, before you write a single line:

Typing prompts into a box is not AI engineering.
Knowing which shoulders to stand on is.

Share your findings

Work is mute. Work is not enough.

Screenshot the chart from Step 6. That single image makes the case in any field, to any audience, better than a deck ever will.

Then send this:

“You are going to think this is a side quest, and you may be right.

I tested whether a free pre-trained AI description beats our hand-built features. At low label counts the gap was big enough that it changes what our labelling budget should be.

Would it be against your priorities if I showed you the 10-minute version on Friday?“

Reminder:

You are not paid to build features anymore.
You are paid to know whose shoulders to stand on.

And if you experimented with something interesting in the Notebook, share with us:

Leave a comment

I help ambitious and smart professionals (scientists, engineers, tech people, and not only) thrive in their careers. If you are doing strong work that nobody sees, that is a positioning problem, not a talent problem, and it moves fast.

Book a session at teodora.coach.

Teodora

If you liked this article, share it with a friend:

Share Standout Systems by Teodora

• Think about something in your life you were very sure about last year that turned out to do not be true.

• Think of a project at work that looked perfect on the whiteboard and fell apart in the real world.

• And on the opposite side: Think about a project/a situation you were not expecting to work at all, and it turned out to be unexpectedly a great success.

Yesterday, I was in a conversation with Baroness Dambisa Moyo, the exceptional and incredible economist who is in the boards of Chevron, Starbucks, Barclays, and SABMiller. This is part of the great Executive AI Masterclass program that I was admitted into. We were a room full of people asking her about AI.

The $100 billion “impossible” deal

One story Moyo shared, stuck with me. When she sat on the board of SABMiller (the second largest beer company in the world), there was one thing everyone agreed on: Anheuser-Busch, the number one player, could NEVER buy SABMiller. The reasoning: Anheuser-Busch would need to borrow approximately $40 billion, the largest corporate loan in history. Plus, there were no much overlapping between the two companies in terms of values.

Then Brexit happened. SABMiller's shares traded in London, the British pound fell, and overnight the company got cheaper for a foreign buyer. Now, the impossible deal became irresistible: Anheuser-Busch bought SABMiller.

This was not lack of information - the sharpest minds in business were in the board. What got them was a shared confident belief that was wrong. The problem is when we are so sure about something that we even don’t thing to question or test.

Why this is suddenly YOUR problem

Moyo’s amazing teaching is that strategy is built on human nature, so the rules of strategy do not change. What changes with every big technology leap, AI included, is the speed: Faster. Cheaper. The advantage that took six years to build can now be built in a week.

What does this mean for your career?

If the data is clean and the question is very clear, AI produces answers in few minutes. Answers are cheap.

But before you panic, look at what companies are actually getting from AI so far:

The prize for the few who get it right is huge: PwC estimates its growth of $16 trillion by 2030.

The gap between who gets it right with AI and the 95% is not a technology gap: it is a JUDGEMENT gap.

The impact you have in your company is not on your answers anymore - it is on the decisions you take. The decisions that could go either way.
Moyo calls it JUDGEMENT - is the main thing screens for on every board seat. The steak at the table is judgement - the rest (e.g., lawyer, accountant, technologist) can be found.

What is JUDGEMENT? The ability to look at a flood of information and know what part is SIGNAL and what part is NOISE. And what AI does? Multiplies the NOISE. This is why JUDGEMENT is more valuable when answers are cheap.

What you get paid for, as AI gets better:

We have seen this before in 1900s:

Machines took the farm work, but we did not run out of jobs.
People moved up: factories, then offices, then labs.
Be skeptical of anyone selling you an AI apocalypse OR an AI utopia.

Both are ideology, and ideology is the enemy of judgment.

Moyo’s rule: "Question anybody who is pushing anything for you, at any time."

Here is exactly which parts of your job are getting cheap, and which are getting valuable.

The drill: your Assumption Audit

How to train JUDGEMENT? It is by practicing it. Let’s practice it now.

Open the plan, project, or proposal you are MOST confident about. The confidence is the key here. Let’s put it to test:

Next time you are in a meeting, this following question will help keep the certainty in check:

That is JUDGEMENT, in action, in a meeting. It is also exactly what makes a room decide you are ready for the NEXT LEVEL.

Few ideas to keep

• AI is coming for the answer. It is not coming for the question.

• Your edge is no longer knowing the most.

• It is being the person who catches the confident belief that is quietly wrong, before it costs the room.

Want to train judgment like a hard skill?

This is the work I do every week with scientists, engineers and the leaders they report to: turning “I know my craft” into “I make the calls nobody else will own.”

If AI is making you wonder what you are still paid for, that discomfort is the signal.

Book a session at teodora.coach

And if this landed, send it to a person you know that might be stressed that AI would take their job.

Share Standout Systems by Teodora

Next week, for paid subscribers, I take the coin-flip call apart properly: a training program for deciding under uncertainty:

• decision journal

• 24-hour rule

• practice judgment as a hard skill.

Train your JUDGEMENT.

Teodora

How long have you been in this role in your company?

Is your manager praising your great work in your 1:1?

Are you satisfied?

The irony is that this is not an underperforming issue. This happens to the most talented, smart, and bright people I know

Talent doesn’t protect you from plateauing. In several very specific ways, talent causes it.

In this article you will learn how to diagnose it, using:

• Six reasons: 3 in your mindset and 3 in your skill set

A 12 questions quiz that will help you identify if you are in an “accidental plateau”

Just a note: This is not about “plateauing by choice” = staying at your level because going to the next level might mean: less free time, or more involvement in company politics, or next level being a totally different job that you don't want to do.

Subscribe now

First: The two Forces you don’t control

Your career is your job.

NOT your manager’s. NOT your company's job.

Your manager’s job is his career and looking good to his manager.

Regarding your manager ask yourself :

• What are their motivations?

• How does your work make them succeed?

• How much power do they really have over your promotion? Hint: usually is your manager's manager deciding your promotion.

• Have they plateaued? How long have they been in this position and (back to the first question: What are their motivations?)

Honestly, your manager’s ceiling could become your ceiling.

Regarding your company make a habit to ask once a year:

• Does my promotion path (my next level) exists at this company?

If you believe AI-native ways of working are the future and your company treats AI as a compliance risk, no amount of personal excellence fixes that.

The counter-move when your manager lacks power but you want to stay? SPONSORS.

Let’s differentiate Mentor versus Sponsor

Mentors talk to you. Sponsors talk about you.

Why talented people stall: The framework

The Accidental Plateau: 6-Point Diagnostic Map

Bellow are the mindset traps and skill gaps that you might be facing:

Reason 1: Praised talent can produce fixed mindset

This reason is the one that probably 80% of my clients have. And I am also super guilty of.

Talented people work hard and focus on continuous learning, outperforming, and growing. Their energy is insane…

And then they get praised: “You’re brilliant”, “What a smart idea”, “What would we do without you”.

And slowly, being smart is your identity. That means: you feel like you need to protect it - so doing things that are not in your comfort zone fill come with a higher tax than someone that is just “normally smart”.

Carol Dweck’s research group demonstrated this with an incredible study (“Carol Dweck: Praising Intelligence: Costs to Children’s Self-Esteem and Motivation”):

Children were split into two groups:

• one praised for effort

• one praised for being smart.

Then everyone was offered a harder task:

Now ask yourself:

• When did you last choose a task that had a real chance of failure. That you knew you might fail at?

If you cannot remember, the praise already got you.

Reason 2: The self-imposed ceiling

A client of mine was offered to interview for the VP role in the company he was working at. He wanted this role for years, dreaming about it. His answer was: “Let me think about it”. Until his wife and friends yelled at him to “Wake up!”.

He interviewed and got the VP role.

Reminder that you fail 100% at the jobs you don't even apply for 🤓

Now a story of mine: Many years ago (I called it “past life”) - I received an email from a Google recruiter to have an interview as a software engineer. The email was very nicely written saying that they think I am a great fit for the position. What did I do?

I marked it as SPAM.

Few weeks later I met a friend that was working for Google - I showed him the email, laughing. He was not amused: “What are you doing? You always dreamed of working at Google!! This is a legit email coming from a recruiter that believes in you!” I took the email out of spam, replied and interviewed. I didn’t accepted an offer because I wanted to move outside Europe in a workplace where I can learn as much as I can - and that is how I got into working with Supercomputers.

Also, a job description is far from being a checklist. Moreover, many job descriptions are outdated - by the time you interview, the job you applied for might have totally different requirements- that might also be different from what you will be actually doing on day 1 at that job. So don’t get too picky on what is on that job description.

You are not the hiring manager of that company. Apply, and let the actual hiring manager decide.

Reason 3: The marketing allergy

Say “self-promotion” out loud. How does it make you feel?

Gross, right?

If you are smart and talented, you have the word marketing, personal brand, self-promotion….because you think that working hard will make you visible.

Spoiler alert: it will not. Work is mute. Work is not enough.

Let's take the example of 2 products:

• Product A: great product, zero marketing, nobody’s heard of it.

• Product B: good product, solid marketing, everyone’s heard of it.

Which one wins?

B wins. Every time. Stanford’s Jeffrey Pfeffer, author of the book Power: “keep your head down and be recognized” is not a strategy. It’s a hope.

You are the product, your results are the features.

It is you job to make sure the people who decide your promotion actually see great impact you are having in the company with your hard, smart work.

Reason 4: What got you here won’t get you there

The Peter Principle in one line: people get promoted for being great at their current job until they land in a job they cannot do anymore. And there they stall or underperform. This is the typical IC to manager or manager to director promotion.

The transitions that matter (IC to manager, manager to director, director to VP) are not bigger versions of the same job. They are different jobs.

So talented scientists, engineers, developers that become new leaders keep getting pulled back into the old job: being right in meetings, redoing the team’s work “because it is better,” being the highest-paid-person’s-opinion in every review. Meanwhile the actual job has three parts:

• Hire and keep exceptional people.

• Invest in relationships, laterally and upward.

• Find the right problems, not the best solutions.

Reason 5: Not investing in learning soft skills

You are not born with soft skills. They get learned and practiced. You might have learned some already by just going though life situations, but it is not enough.

We’ll spend 200 hours mastering a new tool, but communication, influence, executive presence? “That’s just how I am.”

Soft skills are learnable exactly like hard skills: theory, practice, feedback, reps.

Past a certain level they matter more than your technical skills by a lot.

Your hard skills got you in the room. Your soft skills decide what happens in it.

Plus, higher in the organization you get, you need to learn how to handle ambiguity. Leadership decisions are mostly 51/49 calls - making the decision fast is more important than the actual decision. Because you can correct it fast as well.

In the world of AI, what’s left for humans, what you’ll actually be paid for, is judgment under ambiguity:

• The 51/49 calls. Being wrong sometimes, learning, and making the next call without flinching.

Reason 6: Refusing to unlearn

“I’m a specialist in X” becomes your identity.

A big group today are the people dismissing AI.

There are ways to keep a “beginner, fresh mindset”:

• Be a deliberate beginner: regularly pick up something you are bad and use it. A good way for me to practice my big inner mindset now are my Spanish classes (I am terrible at Spanish, but I want to learn it for my husband). I am also learning to play native American flute. Both learning to play a new instrument and learning a new language put you in the “beginner mindset”. Try it and see how uncomfortable it! And that is the reason you should do it.

• Spend time with people a generation or two behind you. They grew up in a different ecosystem, they think differently.

• If you catch yourself being “grumpy” because something does not happen “your way” - and that something is small, like how someone arranged their tabs - that is also a sign of fixed mindset.

The 20-minute Plateau Audit

Do something about it:

If Reason 1 or 2 won (fixed mindset / ceiling): Pick something that feels super uncomfortable (a new project, task, skill, hobby) - the key is “super uncomfortable” and do it. JUST DO IT.

If Reason 3 won (marketing allergy): Open a document and create a brag list: the things you are most proud of in your professional life. Every time you navigate something hard (a tricky stakeholder, an ambiguous call, a crisis), log it: situation, what you did, outcome. Ten minutes a quarter. This becomes your promotion packet, your interview prep, and your visibility engine in one.

If Reason 4 won (wrong job): this week, catch yourself once reaching for work your team should own, and hand it back as a stretch assignment instead. Then ask your manager this:

Seek feedback deliberately, because after a certain level it stops arriving on its own. If you are a manager, director, VP, your team might be a little afraid of you, your colleagues might be busy, your manager might be also busy and thinks about your promotion once a year. Feedback is the key in growing.

If Reason 5 won (soft skills): pick one soft kill (negotiation is my favorite, but ambiguity tolerance the one I need to cultivate more) and train it like a hard skill. There are many amazing podcasts and books on all the soft skills in the universe.

If Reason 6 won (won’t unlearn): what is a tool or a concept you are looking down at? Maybe is vibe coding, maybe is creating slides with AI.

Also, start one sponsor relationship this quarter.

Not with an ask. With value. Remember that there is a Trust Equation:

Want to create a role that doesn't exist yet?

The equation tells you how: research how that role works elsewhere, then bring it to your skip-level as an opportunity for the org, never as "I want this job."

If the idea takes root, the company gains a lane and you become the obvious person to run it, without ever having asked.

The job description you drafted comes out in conversation two, not conversation one.

The one idea to keep

Talent got you to the plateau. It will not get you off it, because the plateau is made of the exact habits that talent rewarded.

The fix is not working harder at the current job.

It’s the audit you just ran: name your dominant reason, run this week’s move, repeat next quarter. Your career is a lagging indicator. What you do now shapes the title you hold in three years, never the other way around. Start leading it.

Did you reach unwanted plateau in your career and need help to move forward?

I help ambitious and smart professionals (scientists, engineers, tech people, and not only) thrive in their careers in a way that is fulfilling and aligned to their values.
I am here to unblock your potential through tested and creative ways, making you proud and aligned to who you are.

If the quiz made you uncomfortable or raised some questions for you, I am here to listen and work together:

Book a session at teodora.coach

If this inspired you to take actions, or just found it insightful, share it with a friend you know that is unhappy in their job. I would love to meet them, too.

Share

Teodora

We started from the ground:

• what Gen AI actually is, what an agent actually is

• and worked our way into the real decisions builders face.

• When to use a fixed workflow versus an agent that decides for itself.

• When to build from scratch versus reach for a library like LangChain or n8n.

• How model routing keeps your costs sane (don’t send the head chef to butter the toast).

• Why the sensitive data should never reach the server in the first place.

• And how to build for the privacy laws that are coming, not just the ones that exist.

If you’ve ever felt locked out of AI because you never learned to code, this conversation is for you.

is proof the vocabulary comes after the building, not before.

Thank you , , and many others for tuning into my live video with !

Join me for my next live video in the app.

Get more from Dr Teodora Szasz in the Substack app

Available for iOS and Android

Get the app

It gets people raises.

A client of mine waited two years for his promotion.

His company had frozen promotions to avoid layoffs. Fair trade.
Then the freeze lifted, his manager put him first in line, and after nine months of paperwork the title was finally about to land.

Then a colleague told him what raise came attached to her promotion: a flat 6%.
Same number for everyone. No relation to performance, scope, or market.

He ran his own numbers. He was 20% below market for the work he was already doing.

Here is what most people do at this exact moment:

• They feel grateful: The freeze is over, the title moved, the company “took care of them.”

• They say thank you, sign, and quietly absorb a 14-point gap for years.

Compounding works against you too: a gap you accept today is baked into every future percentage increase.

Here is what my client did instead, and what you can copy.

The one idea that changes everything

The promotion bump and the market adjustment are two separate conversations.

Read that again, because your company is counting on you to merge them.

The promotion bump is a policy number:

• HR sets it, often as a flat percentage, and your manager has limited control over it.

• Fighting the policy number head-on is usually a losing battle.

The market adjustment is a different instrument entirely. It exists precisely for the situation where someone’s pay has drifted below what the role costs on the open market.

• It has its own budget, its own approval path, and no rule that says it cannot happen alongside a promotion.

So the play is never “6% is insulting.”

The play is: “Thank you for the promotion. Now let’s talk about the gap to market, because that is a separate conversation.”

Everything below is how to run that conversation. Step by step, with the exact words.

Subscribe now

Step 1: Send the agenda email today

Nobody negotiates well while ambushed, including your manager.
A surprised manager gets defensive, and a defensive manager cannot advocate for you.

So before your next 1:1, send one line:

“For Monday’s 1:1, I would like to use part of the time for the compensation conversation, since the promotion is close to being finalized.”

This does three jobs at once. Your manager arrives prepared instead of cornered. The meeting is now anchored, which makes it harder to cancel casually. And you have started the paper trail, which matters later.

Write yours right now: Open your email, address it to your manager, adapt the line to your timing, and send it. That took ninety seconds and the negotiation has already started on your terms.

Step 2: Build the one-pager, written for the room you won’t be in

Your raise will not be decided in your 1:1.

Read more

This companion post is for paid subscribers: the exact code and the full step-by-step to run it yourself, so you leave with the finished system on your machine, not just the memory of watching one.

Subscribe now

You saw it work on the Live. Eleven meetings on the screen, most of them noise, and a one-page brief that pulled out the two that actually decide your week, the decisions due and by when, and what to walk in knowing. Read with your coffee. Prepped before 6 AM.

Copying forty lines of code is easy, and you can. It is in the repo.

But the code is the least interesting part.

What makes this worth an hour of your evening is the shape of it: a pattern you will reuse for your inbox, your documents, your decisions, and a set of dials you get to set your own way. That is what this post is about.

The idea, in one line

A calendar app shows you everything. A brief shows you what matters. An LLM closes the gap in a single call.

The whole trick is telling it to act like a chief of staff who cuts things, not a summarizer who lists them. Hold onto that. Almost every choice in this build comes back to it.

The dials for this build

Every automation in this series is a setting of three dials: how it runs, which brain it uses, and how it reaches you. The build I demoed sets them one way. Yours can be different, and later in this post I lay out the other settings and when to reach for each. Start here, because the choices are the lesson.

How it runs

Scheduled. Two jobs, evening and morning, so tomorrow’s brief is ready the night before and today’s beats you to the coffee. On a laptop that means launchd, not cron, because cron does not run while your Mac sleeps.

Which brain

Claude when the data is safe, local when it is not. The job is judgment, so you want strong reasoning. But a calendar can hold patient initials, a counterparty, an interview you have not mentioned. If it would make you hesitate, the same brief runs on a local model and nothing leaves your laptop.

How it reaches you

A clean email, before you wake. Structured output poured into a simple template, delivered by an email service so there is no fragile password to babysit. It reads like something a chief of staff sent you, not a script.

🔒 Paid subscribers

The Live showed you the engine working, start to finish, free. What follows is the thinking you build on: the one piece of code that actually matters, the pattern that makes every future automation snap onto this one, the repo to clone, and a long menu of ways to make it yours. The finished system, and the judgment to bend it.

Read more

Every clinical AI team I know has the same quiet problem.

Someone on the team is pasting things into a cloud chatbot that should never leave the building. Not maliciously. They are just trying to get work done, and the approved tooling is six months behind their curiosity.

You cannot fix that with a policy memo.

You fix it by giving people something local that is actually good.

Today you build it:

• An open-source AI agent (OpenClaw) running a local model (Gemma 4 via Ollama), sealed inside a Docker container on your own machine.

• With free web search through DuckDuckGo.

• It breaks research tasks into steps, gathers information, and structures output.

• And the privacy accounting is precise: your prompts, documents, and the model's reasoning never leave your machine. When the agent searches, only the short search query goes out, the same as typing it into a browser.

Four components, four jobs

That third piece is the one senior people underrate. An agent that can execute actions should not live loose on your primary machine. The container is not a convenience. It is a control.

1. Docker (5 min)

Download Docker Desktop for your chip (Apple Silicon users: do not grab the Intel build).

Drag to Applications, open, approve the permission prompts. Then verify in the Terminal:

Hello from Docker! means you are live. One warning before you continue: Docker plus local models will eat 10 to 15 GB of disk. Check your headroom now, not at step 5.

2. Ollama and a local model (5 min)

Download Ollama from ollama.com. Then pick your Gemma 4 size based on one number, your RAM. Check your RAM using the following command in the Terminal:

system_profiler SPHardwareDataType | grep Memory

32 GB or more? Pull gemma4:26b, a Mixture of Experts model that delivers 26B-class quality at near-4B speed because only ~4B parameters activate per token.

16 GB? Pull gemma4:e4b, which needs about 8 GB and leaves room for Docker. Both have native tool calling, which is the capability this whole build depends on.

That last command should return:

If ollama serve complains the address is in use, good: it is already running in the background.

3. OpenClaw in a container (10 min)

mkdir -p ~/openclaw-project && cd ~/openclaw-project
git clone https://github.com/openclaw/openclaw.git
cd openclaw
export OPENCLAW_IMAGE="ghcr.io/openclaw/openclaw:latest"
./scripts/docker/setup.sh

The wizard will ask questions.
Choose QuickStart, skip the model provider (we wire Ollama next), skip channels, skip skills. Then confirm health:

You want status:

Do not move further on until you see it.

4. The step where most people gets stuck (5 min)

Here's the part nobody talks about. Inside a container, 127.0.0.1 means the container itself, not your Mac.

Your agent will report it cannot reach Ollama, and it will be telling the truth.

The fix: Docker’s special hostname. Run the configure wizard:

Select Model, then Ollama, mode Local only, and replace the base URL with:

Pick your Gemma 4 model from the discovered list.

One more registration quirk: Ollama's local API needs a placeholder key (any value, it is not a real secret):

The gotcha that costs everyone an hour: there are two allowlists.

The provider list (models.providers.ollama.models in ~/.openclaw/openclaw.json) is what the gateway can reach.

The agent list (agents.defaults.models) is what the agent is allowed to use.

The wizard fills the first and often not the second, and you get model not allowed until both contain your model.

Add "ollama/gemma4:26b": {} under agents.defaults.models, set it as agents.defaults.model.primary, then force-recreate the gateway. Config is read at startup only.

5. Free web search, no API key (2 min)

A local model without retrieval will not say "I don't know." It will generate the shape of a well-sourced answer, links included, from its weights.

Giving the agent real search is what turns confident fabrication into checkable findings. DuckDuckGo is OpenClaw's key-free provider: no account, no plugin, no key.

One config edit:

python3 -c “
import json, pathlib
p = pathlib.Path.home() / ‘.openclaw’ / ‘openclaw.json’
cfg = json.loads(p.read_text())
cfg.setdefault(’tools’, {}).setdefault(’web’, {})[’search’] = {
    ‘enabled’: True, ‘provider’: ‘duckduckgo’
}
p.write_text(json.dumps(cfg, indent=2))
print(’web search -> duckduckgo’)
“
docker compose up -d --force-recreate openclaw-gateway
docker compose ps

6. Open the dashboard and test (5 min)

Paste into your browser: http://127.0.0.1:18789/?token=YOUR_TOKEN

If you hit Auth did not match, the token is stale.
Regenerate with:

docker compose run --rm openclaw-cli dashboard --no-open
Then force-recreate the gateway.

Now type /tools and confirm web_search is in the list (tools resolve when a session starts, so always test in a fresh session).

Then /model ollama/gemma4:26b and give it a real task:

Watch it decompose the task, work the steps, and structure the output. Every token generated on your own silicon.

Your API bill for this session: $0.00. Your data exposure: zero bytes.

Next time on Standout Systems

I Gave My $0 Agent a Job While I Sleep

The agent you just built has a scheduler. Next, we put it to work: a 6 AM research sweep that reads your field before you wake, an overnight literature scan that flags new papers with sources you can click, and a Friday competitive briefing that lands before your first meeting. Real cron jobs, copy-paste ready.

Here’s why it matters: scheduled work is where local AI stops being a cheaper chatbot and starts being an employee.

Nobody waits on an overnight job, so slow-but-unmetered inference is the perfect worker, and the token count that would sting on an API bill costs you exactly nothing.

We’ll also cover the one setting you must change before any agent runs unsupervised. Do not skip that part.

See you tomorrow.

- Teodora

He was about to ask his manager for a raise. He read me his case over the phone, and it was solid. Strong performance. Extra responsibilities he had absorbed without complaint. Years without a meaningful adjustment.

I stopped him before he booked the meeting.

Not because his case was weak. Because his case was about the wrong subject. Every sentence started with “I.” I’ve performed. I’ve taken on more. I deserve this.

Here is the hardest truth about compensation and promotion negotiations, and it comes straight from my training with Chris Voss’s team: the moment the conversation is about you, you have already lost the room.

• Managers do not fund fairness.

• Committees do not fund effort.

They fund impact. Your case has to answer two questions, and neither one is about you:

1. What impact has this person already delivered to the company?

2. What bigger impact does the next level unlock?

We spent the call rebuilding his ask around those two questions.

And that call reminded me how often I have this exact conversation in coaching sessions, usually in an even sharper version of the problem: the senior engineer, data scientist, or clinical researcher who has been operating two levels above their title for years, paid one level below the work.

Everyone on the team knows it. The org chart doesn’t.

If that is you, today is your playbook. Step by step, with the exact words.

Subscribe now

Negotiation zero: the one with yourself

Before you negotiate with anyone, answer one question honestly:

Do you want the role, or do you want what the role pays?

These are different asks. They have different strategies. Most people never separate them, and it costs them years.

Here is the part to pay attention to: A promotion is often a completely different job. Move from senior IC to manager and your calendar becomes: 1:1s, performance reviews, and headcount fights. Your compensation often shifts too: more of it tied to how the business performs, not how you perform. As an IC, a rough quarter for the company barely touches your paycheck. As a director, your bonus is riding on it.

So do this right now. Not later. Now. Write one of these three sentences and make it true:

1. “I want the next role because I want that job, and the pay comes with it.”

2. “I want to be paid at the level of impact I already deliver, and I am open on title.”

3. “I want bigger scope, and I want compensation to follow within 12 months.”

When I push clients on this, roughly half discover they picked the wrong fight. The most common flip: someone walks in asking for “the promotion” and realizes they do not want to manage at all. They want the market rate for the impact they already ship. That is sentence 2, and it is a different negotiation entirely. Not a promotion. A pricing correction on impact the company is already receiving.

Feel the difference? One is a favor. The other is an invoice.

If you cannot pick a sentence yet, that is the first thing to raise with your manager or skip-level: “What does the next level actually look like day to day here?” Ask before you campaign for it. I have watched people fight for a role for two years and hate it in two months.

The campaign: what happens before the 1:1

Promotions are decided in calibration meetings, by committees, months before any announcement. Lorraine Lee tells a story in Unforgettable Presence about a colleague who was overdue for promotion for one reason: her boss did not know she wanted one. She said it out loud, and she got it.

Your manager is not a mind reader. And by the time review season opens, the budget is already carved up. Here is the timeline, working backward from the review cycle.

Six months out: declare intent, framed as impact. In your regular 1:1: “I want to be operating at [level] by [date], because that is where I can deliver [specific bigger outcome] for the team. I want to build the plan with you.” Retitle one 1:1 to “Career Development Conversation” and send your talking points 48 hours ahead. If your manager is an analyst type (most technical managers are), surprises make them defensive and preparation makes them your ally.

Five months out: build the gap map. Pull the career ladder for the next level. No written ladder? Ask: “What would someone at that level be delivering that the team isn’t getting from me yet?” Map every requirement to evidence or to a gap. This converts your promotion from your opinion into a checklist you both watch.

Ongoing: run the impact log. A running document. But here is the discipline that separates a strong log from a diary: every entry must name the impact on the business, not the effort from you. Not “improved the pipeline” but “cut validation time from 3 weeks to 4 days, unblocking 2 submission milestones.” Not “worked weekends on the release” (effort, about you) but “release shipped 3 weeks early, protecting the Q3 customer commitment” (impact, about them).

Do the first three entries right now. Open a doc. Write the last three things you shipped. For each one, finish this sentence: “which meant the company got ___.” If you cannot finish the sentence, the entry does not go in the log. That filter took four minutes and it just made your case sharper than most of your peers’.

Two to three months out: arm your manager. In the calibration room, your manager is your sales rep. You are not negotiating against them. You are building their pitch deck. Send a “Monday wins” email weekly, three bullets, each one an impact statement with a number. When the committee asks “why this person,” your manager should be reading your words back to them.

Also two to three months out: the skip-level. Ask for 30 minutes. You are not going over your manager’s head; you are making sure the second most important vote in the room has seen the impact firsthand. One calibrated question does the work: “What would you need to see from me to be confident the team gets next-level impact from me?” Then be quiet and take notes.

If you are counting: that is at least ten conversations before the “real” one. That is the point.

The Voss toolkit, applied to promotions

I spent multiple days in live negotiation training with Chris Voss’s team at the Black Swan Group. The single biggest lesson: the other side cannot hear your case until they feel heard. Your manager is under pressure you do not see. Budget freezes, their own promotion case, three other people asking for the same thing. Walk in with only your ask and you are a threat. Walk in with tactical empathy and you are a partner.

Notice something: tactical empathy and impact framing are the same move. Both take you out of the center of the conversation and put the other side there. That is why they compound.

Here is the toolkit, with the scripts my clients rehearse.

1. The accusation audit: name their objections before they do.

List every reason they could say no. Then say the worst ones out loud, first.

“You’re probably thinking the timing is terrible. Budgets are locked, you have other people asking, and here I am adding to your plate. It might even seem like I’m giving you an ultimatum.”

The objections you name lose their power. Your manager relaxes, because the hard part of their job just got done for them. Voss’s team opens nearly every difficult conversation this way, because the amygdala calms down when the fear gets named.

2. Labels: make them talk 70% of the time.

The structure is always “It seems like...” or “It sounds like...” Never “I think” or “I hear you saying.” Take yourself out of the sentence.

“It seems like there are constraints on this that are above your pay grade.”
“It sounds like the committee weighs visibility more than output.”

Every label buys you information about how the decision actually gets made. That information is worth more than any argument you prepared.

3. Calibrated questions: make them solve your problem.

Open questions starting with “what” or “how.” Never “why” (it triggers defensiveness), and never a stressed “how am I supposed to do that,” which reads as an attack.

“What would need to be true for this to happen this cycle?”
“How does the committee decide between two strong cases?”
“What would you need from me to make this an easy case to present?”

That last one is the crown jewel. It turns your manager from gatekeeper into co-author.

4. Let them anchor first.

If the conversation turns to compensation numbers, do not throw the first one. Ask: “What is the range for that level?” If you guess $150K and ask for $155K, you will never know the band went to $175K. Once they anchor, negotiate up from their number, not down from your fear.

5. Silence.

Make the ask. Then stop talking. The pause feels like ten minutes to you and three seconds to them. Every word you add after the ask is a discount you are giving yourself. Voss calls it dynamic silence. I call it the hardest four seconds of your career, and the best paid.

6. Tone: 75 to 80% accommodator.

Warm, downward-inflecting, unhurried. Not because you are soft. Because an attacked brain gets measurably dumber, and you need your manager’s brain fully online to advocate for you. If the assertive edge creeps in, do what Voss does: smile while you talk. The tone follows the face.

The 1:1 itself: the full script

Put it together. Career-titled 1:1, agenda sent 48 hours ahead, impact log linked.

Accusation audit: “You’re probably going to think this is bad timing, and you might feel like I’m putting you in a tough spot with the committee.”

The case, past impact first: “Over the last 12 months, the team got [three strongest impact numbers] from my work. Based on the ladder, that is [next level] scope.”

Then future impact: “At [level], I can take [specific bigger thing: own the roadmap, lead the multi-site rollout, carry the regulatory submission end to end] off your plate and deliver it.”

The ask: “I want to make that case this cycle.”

Silence. Four seconds. Hold it.

Calibrated question: “What would you need from me to make this an easy case for you to present?”

Label whatever comes back: “It seems like the visibility piece is the gap.” Then: “What would closing that look like in the next 90 days?”

Fifteen minutes. Zero sentences about deserving. Every sentence about what the company got and what it gets next. That is the whole meeting. Everything else was the campaign.

By the way, this is where my brother’s raise conversation landed too. Same skeleton, smaller ask: past impact, future impact, silence, calibrated question. The playbook scales.

When the answer is “not this cycle”

It usually is, the first time. This is not a loss. This is the moment you convert a vague no into a contract.

“That’s fair. What specifically would the team need to see from me for this to be a yes next cycle?”

Get the answer in writing. A shared doc, a follow-up email, anything with a timestamp. “Per our conversation, the three things to demonstrate by Q2 are...” Now you have a promotion agreement with a deadline instead of a hope with a mood.

And if two cycles pass with the goalposts moving? That is real information too. The market prices impact faster than committees do. That is a different article, and a different negotiation.

Resources worth your time

• Never Split the Difference, Chris Voss. The source. Read the chapters on labels, calibrated questions, and the accusation audit twice.

• Unforgettable Presence, Lorraine K. Lee. The best current playbook on promotion mechanics: intent, promotion plans, manager relationships.

• Wild Courage, Jenny Wood. Scripts for career conversations and the “let them anchor first” compensation move.

• Reverse the Search, Madeline Mann. For when the internal path stalls and the outside market becomes your leverage.

• Black Swan Group newsletter (blackswanltd.com). Free weekly negotiation reps.

This is exactly the work I do with clients

Reading about labels and running one under pressure, with your actual manager, about your actual paycheck, are two different sports.

I trained for days in live workshops with Chris Voss’s team. I have negotiated inside a Fortune 500, across hospital systems, and through collaborations for a decade. And here is what that experience taught me: every negotiation is unique, and in today’s market the game has changed. Creativity wins. The reps have to happen long before the real conversation, which is why my clients and I rehearse the accusation audit, the labels, the impact framing, and the silence until they are muscle memory, tailored to their manager, their committee, their company’s politics.

Subscribe now

If a promotion conversation is on your horizon, this is precisely the kind of situation we untangle in 1:1 coaching at teodora.coach. We build your campaign, translate your work into impact language, and run the reps together.

Would it be a bad idea to book a call before your next review cycle opens?

Schedule a Call

Talk soon,

Teodora

In 2023, a four-year-old boy named Alex stopped growing the way he should. His balance faltered. Sitting hurt his lower back. His mother took him to 17 different doctors.

Seventeen.
No answer.

So she did something that would have sounded absurd a few years earlier.
She compiled every note she had kept: the full symptom list, the MRI report, the timeline. She pasted all of it into ChatGPT and asked one question. What is wrong with him?

The model suggested tethered cord syndrome.
A condition none of the 17 visits had surfaced. She took the suggestion to a neurosurgeon, who reviewed the same MRI, examined Alex, and confirmed it.
Six weeks later, Alex had surgery. He is a different kid now.

You can read more about the story here: https://www.today.com/health/mom-chatgpt-diagnosis-pain-rcna101843

Here is the part most people miss when they retell this story.

The AI did not outperform 17 doctors. The prompt did. His mother gave the model something no single doctor ever received: the complete, organized history. Every symptom, every date, every finding, in one place.

That is the whole lesson of today’s article. When you ask AI a health question, the quality of your answer is decided before the model writes a single word. It is decided by the context you feed it.

Subscribe now

I have spent a decade building AI systems for healthcare, and I can tell you the same law governs a foundation model trained on millions of medical images and a chatbot on your phone: sparse input, generic output. Rich input, specific output.

The analogy: a physician taking your history. A good doctor does not diagnose from “my back hurts.” She asks when it started, where exactly, how bad, who you are, and how it has changed. Your job is to give the AI the interview a doctor would have conducted. There is a framework for exactly that, and it comes straight from how clinicians are trained to take a history.

The TLICC framework

Five letters. Time. Location. Intensity. Context. Change.

T is Time. When did this start? An hour ago, a month ago, three years ago? The differential diagnosis for sudden chest pain and chronic chest pain barely overlaps. Time is the single fastest way to shrink the search space.

L is Location. Where, precisely? Front of the shoulder or the back? Does it travel down the arm? Which side? Vague location produces vague hypotheses.

I is Intensity. How bad, how often, and what does it feel like? Throbbing, sharp, dull, constant, occasional? A 3 out of 10 that never stops is a different problem from a 9 that comes in waves.

C is Context. This is where you should spend the most words. Your age. Your sex. Your existing conditions. Every medication you take. Whether you smoke or drink. What runs in your family. What your daily life looks like. Context is what turns a generic answer into an answer about you. It is why Alex’s mother succeeded where a symptom checker would have failed.

C is Change. How has it evolved? Steady, worsening, on and off? Trajectory carries diagnostic weight that a snapshot never can.

Build the prompt right now

We do it here. Below is the template. Fill in the brackets, paste it into a current frontier model (the newest version of Claude, ChatGPT, or Gemini, and use the paid tier if you have one, because the model version is the single biggest driver of answer quality):

I want help thinking through a symptom. Do not give me a generic
list. Use the details below and reason like a clinician taking
a history.

TIME: It started [when]. It occurs [how often].
LOCATION: The symptom is in [precise location]. It radiates
to [where / nowhere].
INTENSITY: At its worst it is a [X]/10. It feels [sharp /
dull / throbbing / pressure].
CONTEXT: I am a [age] [sex]. My medical conditions: [list].
My medications: [list]. Smoking: [yes/no]. Alcohol: [amount].
Family history: [list]. Relevant lifestyle: [sleep, stress,
exercise, recent travel or injury].
CHANGE: Over time it has [gotten worse / stayed stable /
come and gone].

Question: What are the possible explanations, and how would
you reason between them?

The model will come back with a lot. Possibly too much. So immediately send the second message. I call it the rule of threes:

Simplify this for me. Give me exactly:
1. The three most likely explanations.
2. The three questions I should ask my doctor.
3. The three warning signs that mean I should seek care urgently.

Three causes. Three questions. Three red flags. That is a page you can print and carry into your appointment. You just walked in as the most prepared patient your doctor will see all week.

Know your zone before you trust the answer

Not every health question carries the same risk, and you should not treat every answer the same way. Think in three zones.

Green zone. Nutrition advice, an exercise plan, understanding a lab value, general wellness. Low stakes. Use the AI freely.

Yellow zone. A persistent symptom, a medication question, something that worries you but is not urgent. This is where most of your questions will live. Here the rule is trust but verify. Ask the model for its sources. Then leave the chatbot and check the claim on Mayo Clinic or another institution you trust. Do not let the model grade its own homework.

Red zone. A new severe symptom. The worst headache of your life. Chest pain. Bleeding in pregnancy. Anything sudden and frightening. The AI is not the tool here. A human clinician is, and quickly. In 2023 a pregnant woman in Australia described sharp abdominal pain and light bleeding to a chatbot. It suggested rest and hydration. She collapsed the next morning from a ruptured fallopian tube. In the red zone, the cost of a wrong reassurance is catastrophic. Go to care first, ask questions later.

Two more safeguards, and they are non-negotiable.

Push back on confident answers. These models are trained to be agreeable, and their confidence is a performance, not proof. In high-stakes questions, force the reasoning: “What other conditions produce these exact symptoms? How would I distinguish them? What would make you change your answer?” You are not being difficult. You are doing what a good attending does to a resident.

Protect your identity. Never paste your name, date of birth, or anything identifying into a consumer chatbot. Your symptoms are relevant. Your social security number is not. And go into settings and turn off the option that lets your conversations train future models. It is often on by default.

The deeper point

Strip away the medical setting and TLICC is something you already know if you build AI systems: context engineering. The model was never the bottleneck. The input was.

A joint Harvard and Stanford study found AI systems reached 92 percent diagnostic accuracy on complex cases, far above the physicians in the study.

The intelligence is sitting right there. What separates a generic answer from a life-changing one is whether you hand the model a history worth reasoning over.

Alex’s mother did not know the term “context window.” She just refused to give the machine less than she would give a doctor.

Do the same.

Subscribe now

Teodora

Entry-level job postings in the US are down roughly 35 percent since early 2023. Employment for workers aged 22 to 25 in the most AI-exposed roles fell measurably while older workers in the same fields held steady.

Among young software developers the drop was close to 20 percent. Recent-graduate unemployment climbed to 5.7 percent in late 2025, above the rate for all workers, with a record share underemployed.

The earliest visible footprint of AI on the labor market is not mid-career people getting fired. It is the bottom rung of the ladder getting thinner, the rung where you used to get paid to do routine work that quietly taught you judgment.

That is real. I am not going to soften it. If you are early-career, or advising someone who is, the launchpad got smaller, and the data is on your side when you feel anxious about it.

Now let me put it in a frame that has been true before.

Around 150 years ago, electricity was the new civilizational technology. It was thrilling, confusing, and genuinely dangerous. It started fires.

It killed people.

We did not ban it, and we did not surrender to it. We learned what it could do, learned to use it safely, taught our children to, and built an entire layer of norms and infrastructure around it.

Cars were the same.

When they arrived they were horribly unsafe, and they also did things horse carriages never could. So over decades we added traffic laws, speed limits, seatbelts, airbags. We kept the power and engineered the danger down.

That is what humanity does with a technology too useful to refuse. We do not pick utopia or doom. We pick agency.

The doom story and the hype story share one assumption, and it is the one I most want you to reject.

Both quietly assume you are a passenger.

The hype says the technology can do no wrong, go all in, stop thinking. The doom says it will replace you, nothing you do matters, give up.
Replacement, existential crisis, extinction: every one of those words smuggles in the idea that you are along for the ride. You are not. How this goes for you personally is a question about what you decide to do, on purpose, starting now.

Here is the part the data actually supports, and it is the opposite of doom. One of the first real studies of AI at work followed more than 5,000 customer support agents using a language model.

The least experienced workers gained the most, 30 to 40 percent. The most experienced gained almost nothing. Why? The model had learned from the best agents and was handing their tacit knowledge to everyone else. The thing that took the top performers years to build, the read they could never quite put into words, got captured and passed down. That is not a story about people being replaced. That is a story about the floor rising.

So the question is not whether AI touches your work. It will. The question is which side of one line you stand on.

The line is not between people who use AI and people who do not.
By now everyone touches it.
The line is between people who use AI passively and people who use it with intention. The passive version: type a prompt, copy the answer, paste it, brain off. That is exactly what gets commoditized, because anyone can do it and the machine did the thinking. The intentional version: you bring judgment the model does not have, you direct it, you take its output and sharpen it with your own expertise, you push back and get a better answer, and you own the result because it is genuinely better than what either of you produced alone.

That second mode is a skill. It is learnable. And it is the one the market rewards, because it is the one where a human is visibly adding the scarce thing.

Let us build your version of it. Right now, in this article.

Subscribe now

There is a clean way to think about any piece of work.

It has three parts:

• defining the question

• executing it, and

• judging the result.

For all of history humans did all three.

AI is now extremely good at the middle one, execution, once the problem is well defined.

Which means your advantage moves to the two ends: asking the right question, and evaluating the answer. That is the work of a CEO.

Start thinking of yourself as the CEO of your own work, directing the execution rather than performing all of it by hand.

Run one real task through all three parts. Take something actual from this week, not a toy example. First, define it properly. Most people skip this and hand the model a vague request. Do the opposite: write what you are really trying to accomplish, the constraints that matter, what a great outcome looks like. This is the part only you can do, because it depends on context the model cannot see.

Hand the execution over, then judge it hard. Give that well-defined problem to the AI and let it do the middle part. Then put on your evaluator hat, the one that took you years to earn. Where is the output generically correct but wrong for your specific situation? Where would this fail in the field? Name the three things you know that the model did not, feed them back, and push until the result clears your bar. Notice what happened: you did not consume the AI, you directed it. You defined, it executed, you judged. The fingerprint on the final work is yours.

Make that loop your visible signature over the next ninety days. Here is the positioning move, and it works whether you are climbing onto the ladder or deciding your next senior step. Pick one place in your work where you run this define-execute-judge loop and let people see it. Not “I used AI.” Everyone says that. The visible thing is the judgment you added: the question you framed, the call you made on top of the output, the part that needed a human who carries the consequence. In a market nervous about juniors who just paste model output, being visibly the person who defines and judges is the entire differentiation.

If you are early-career, this is also your answer to the thinned-out ladder. The old rung was getting paid to do routine work until you absorbed the judgment by osmosis. That rung is shrinking. The new rung is demonstrating the judgment directly, on real tasks, in public, faster than the old apprenticeship ever allowed. The data says the easy entry path got harder. It does not say the door is closed. It says the door now opens for people who can show the scarce thing instead of waiting years to be handed it.

This is the optimism I can defend as a scientist, and it is not the cheap kind. I am not telling you AI will leave your job untouched. It will not.

I am telling you the change has a steering wheel, and the doom story’s whole trick is convincing you to take your hands off it.

The smartest path was never machines that imitate humans and make us worth less. It is machines that amplify humans and make us worth more. A bicycle for the mind does not replace the rider. It lets the same effort travel much further.

Electricity could kill you in 1890. We did not ban it and we did not worship it. We built seatbelts, and then we drove.

Put your hands on the wheel.

Next, we run this together. I will take you through real examples, the inbox that triages itself, the strategic decision pressure-tested against your own data, the dashboard built in minutes instead of weeks, and we will wire up the define-execute-judge loop on each one so you can watch the intentional version beat the lazy one. Bring one real task. We build from there.

Subscribe now

Teodora

Over ten parts, I take the work that quietly eats your week, the sorting, the scanning, the triage, the prepping, the second-guessing, and hand each piece to AI.

Not in theory.
In Python. With your data.

Subscribe now

Every post ends with a working thing you can run, not a reading list you feel guilty about.

We build across the whole shape of your day:

• Automations that read your calendar and inbox and surface what actually matters.

• Tools that turn a messy pile of data into one clear view.

• Ways to pressure-test a decision or a message before you commit.

• Filters that let only the few important things through.

• And by the end, a way to wire it all together so it runs as one.

The payoff is not doing more tasks faster. It is doing fewer of them yourself, and spending the hours you reclaim on the work only you can do.

How it works

Each part is two things.

A live build, where I make the automation on screen in real time, on fake data so nothing breaks and you can follow along.

A written post, with the full code and a step-by-step guide to running it on your own calendar, your own inbox, your own accounts.

Come to the live to watch it work. Keep the post to build it yourself.

What you get

Free subscribers get every live build and every written walkthrough. The whole series, start to finish. This is the real thing, not a teaser.

Paid subscribers get the part that saves you the most time:

• The complete, copy-paste code for every post, ready to run.

• The production notes: real authentication, error handling, and the version hardened for sensitive and regulated data.

• The deep-dive companions, including when AI can touch your data and when it absolutely cannot.

Free teaches you the idea.

Paid hands you the finished system.

One promise

You are not late to this. We are all early. One person can now do what used to take a team, and the only thing gating you is whether you are willing to try.

Subscribe, and Part 1 lands in your inbox next week. Bring your messiest calendar.

Teodora

• The way the hands rest.

• Where the feet are.

• A micro-expression that flickers and is gone.

• Then he checks the monitors to confirm what his body already told him.

Ask him to write down how he did it and he cannot. Not because he is hiding it. Because there are no words for it.

This is the most underrated fact about AI, and almost nobody building it says it out loud. Language is lossy. It is a compression format for reality, and like every compression format, it throws things away.

The five-millisecond read is exactly what it throws away: perceptual, built from ten thousand prior patients, fired in a part of the brain that does not narrate itself.

Large language models (LLM), the technology behind this entire wave, learn from language.

They are extraordinary at the world that has already been written down.
They are blind, by construction, to the parts that were never put into words.

There is an older kind of intelligence underneath language: seeing, reasoning about what you see, knowing what to do next. Most animals have it.

We had it long before grammar.

And the systems everyone fears are built on the thinnest, most recent, most lossy layer of human capability.

Here is what that means for your job.

Yesterday you split your role into tasks and found the high-judgment ones.

Today, go one layer deeper.

Inside that high-judgment work there is a five-millisecond read: a thing you do that you cannot fully explain, that took years of cases to build, that fires faster than your ability to justify it. That read is your most defensible asset, precisely because it never got written down in a form a model could train on.

Let us find yours. Right now.

Find your five-millisecond read. Think of the last time you looked at something in your domain and knew it was off before you could say why. A dataset that smelled wrong before you found the bug. A deal about to turn before anyone spoke. A design that was clean on paper and you knew would fail in the field. Write that specific moment down.

Try to fully explain it, and watch where words run out. Write the steps a junior person would need to reproduce your read. You will get three or four steps in and hit “and then I just knew.” That wall is the edge of what language can hold. Everything past it is yours, and structurally hard to automate, because the training data for it does not exist.

Name the inputs you were actually using. Your read ran on more than text. A shift in someone’s tone. The shape of a distribution you have seen a hundred times. The full context no transcript captures. Write those non-text inputs down. They are the moat. A text-only system cannot see them.

This is the line between automation and augmentation, and your career runs on it.

Automation takes the task that fits in words and does it for you. The nurse no longer types blood pressure into the record by hand; the monitor pushes it in, and the nurse gets that time back for the part of the job that needs a human in the room. That is good. It hands you back the hours you spent as a data-entry clerk.

Augmentation is different. The system flags, surfaces, sharpens the faint pattern, and then a human who carries the consequence decides. The deciding does not move. It cannot, because it runs on inputs the machine cannot see and accountability the machine cannot hold.

So here is what the doom version never tells you. As AI absorbs the task layer, the judgment layer does not hold steady in value. It gets scarcer, and scarce things get more valuable. That clinician’s five-millisecond read is worth more in a hospital full of AI tools than in one without them, because now it is the part that did not get commoditized. Same for you.

Find that work. Stand on it. Be the human the system gets augmented by, not the task it automates away.

Tomorrow, the last piece: why the entry-level panic is real and still not the whole story, what electricity and seatbelts have to do with your next ninety days, and the one test that separates people who use AI from people the market actually rewards.

Teodora