AI, From a Human Point of View

  • The Human Touch That Made ChatGPT Work

    When ChatGPT appeared in November 2022, the surprise was not simply that the model was powerful.

    Powerful language models already existed.

    The real change was that, suddenly, ordinary people could talk to one.

    That sounds obvious now. At the time, it was the difference between having an impressive engine in a laboratory and handing someone the keys.

    One of the techniques behind that shift was RLHF — Reinforcement Learning from Human Feedback. OpenAI had already used it with InstructGPT, and ChatGPT was trained with a related process designed to make the model follow instructions and behave more usefully in conversation. OpenAI

    The idea is easier to understand than the acronym suggests.

    First, people showed the model examples of useful responses.

    Then they compared different answers and ranked them.

    Those preferences were used to train a separate reward model — essentially a system that learned which responses humans tended to prefer.

    Finally, the main model was fine-tuned to produce more of those preferred behaviours. OpenAI

    There is something almost ironic about it.

    After years of trying to make machines more intelligent, one of the important steps was teaching them something much more ordinary:

    how to behave when somebody asks a question.

    Not perfectly, of course.

    RLHF did not suddenly make AI truthful, harmless or infallible. OpenAI itself has long acknowledged that models trained this way can still make mistakes, fail to follow instructions, produce biased outputs or give convincing answers that are simply wrong. OpenAI

    But it changed the experience.

    And sometimes the experience is what turns technology into a product.

    Before ChatGPT, using a large language model could feel like interacting with the machinery behind the curtain.

    ChatGPT moved the curtain.

    You typed.

    It answered.

    You disagreed.

    It tried again.

    You could continue the conversation without knowing what a transformer was, how many parameters were involved or what happened between pressing Enter and seeing the next sentence.

    That accessibility mattered enormously.

    So when people ask what made ChatGPT different, I would not reduce the answer to model size.

    Part of the story was technical progress.

    Part of it was the interface.

    And part of it was human feedback quietly teaching the machine that being intelligent is not especially useful if nobody can work with you.

    Which, now that I think about it, is probably true of humans too.

  • Nano Banana 2.1: The Update That Matters After the First Image

    There is a moment with AI image models that rarely appears in launch videos.

    It is not the first image.

    The first image is easy. You write a prompt, press a button, and sometimes something beautiful appears.

    The real test starts five minutes later.

    Change the background. Keep the same character. Fix the text. Move the camera. Try again.

    That is where many models begin to forget what they were doing.

    Nano Banana 2.1 caught my attention because Google seems to be focusing on exactly that part.

    It keeps the speed and efficiency of Nano Banana 2, but improves visual quality, prompt-following, multi-turn consistency, text rendering and infographic layouts. It also fixes some artefacts in very wide formats.

    Nothing especially dramatic.

    Which is probably why it matters.

    Nano Banana 2 Lite still looks like the speed-and-scale option. Nano Banana Pro remains the more demanding choice for maximum control and complex design work.

    Nano Banana 2.1 sits somewhere in the middle, and that middle ground is often where useful tools live.

    Fast enough for everyday work.

    Better when consistency matters.

    Less heavy than Pro when you do not need studio-level control.

    For me, the most interesting part is not whether it can generate one impressive image.

    It is whether it can keep the same idea alive across several edits.

    A character.

    A product.

    A visual identity.

    A sequence.

    One good image can be luck.

    Five consistent images are a workflow.

    That is the test I want to run.

    Because in practical AI image work, the first result gets the attention.

    What happens next decides whether the tool stays.

  • The First Draft is Supposed to Be Bad

    The first thing AI gives you will often be wrong.

    Not catastrophically wrong.

    More annoying than that.

    Almost right.

    A little too polished. Slightly generic. Missing the point you thought was obvious. Using exactly the sentence you would never use yourself.

    That is not necessarily a failure.

    It is a first draft.

    We tend to forget that because AI delivers it in seconds, neatly formatted and wearing the confidence of something finished.

    Speed is persuasive.

    A page that appears instantly looks more complete than a page you spent an hour writing.

    It usually isn’t.

    Nobody looks at a sketch and complains that it is not yet a painting.

    Nobody expects the first pancake to represent the entire career of the cook.

    The first one tells you the pan is too hot, the batter is too thick, and perhaps breakfast needs another five minutes.

    Then you adjust.

    AI works surprisingly well the same way.

    The Draft Is Not the Deliverable

    The first output is not something you accept or reject.

    It is something you react to.

    That distinction changes everything.

    A blank page is polite but useless. It gives you nothing to argue with.

    A mediocre draft is much more generous.

    It gives you the wrong tone to correct.

    The missing idea to add.

    The sentence that is nearly right.

    The paragraph that makes you realize what you actually wanted to say.

    That is where your judgment starts doing its job.

    And judgment is still the expensive part.

    AI can get words onto the page quickly. It can suggest structure, offer alternatives, shorten something, expand it, simplify it, or give you five versions before your coffee has cooled.

    Useful.

    But it does not know which version sounds like you unless you teach it.

    It does not know which detail matters because you lived it.

    It does not know when a technically better sentence has somehow become emotionally worse.

    That part still belongs to you.

    The people who get the most from AI rarely expect perfection on the first try.

    They expect material.

    Then they cut.

    They redirect.

    They change the angle.

    They keep one good line and throw away the rest without holding a small funeral for the prompt.

    The machine did not finish the work.

    It simply made sure the work had somewhere to begin.

    So yes, the first draft is supposed to be bad.

    Not always.

    But often enough that we should stop being surprised by it.

    The first pancake was never the problem.

    The problem would be serving it to the guests.

  • I Don’t Know, and Neither Does Anyone Else

    People ask me where all this is going.

    Will machines become smarter than us? Will AI take my job? Should I be worried?

    They usually expect a confident answer, probably because I spend a lot of time working with these tools.

    The honest answer is less impressive.

    I don’t know.

    And neither does anyone else, at least not with the certainty some headlines, keynotes, and press releases would like us to believe.

    There are thoughtful, serious people who think we are only a few years away from something extraordinary. There are equally thoughtful, serious people who think the current wave of AI will eventually run into limits we are underestimating.

    Both camps have been wrong before.

    That does not make them foolish. It simply means the future has a bad habit of refusing to follow the script.

    A few things I would have dismissed not long ago now work surprisingly well.

    A few things I was convinced were just around the corner are still standing somewhere around that same corner, apparently in no hurry to arrive.

    My own record at predicting AI is probably about as good as everyone else’s.

    Not great.

    So I try to treat the really big questions the way I treat a weather forecast three weeks ahead.

    Interesting.

    Worth checking.

    Probably not enough reason to cancel dinner.

    What interests me more is the next twelve months.

    That horizon feels slightly less theatrical.

    I can watch which tools people around me are actually using. I can see what has changed at work over the last quarter. I can notice which tasks suddenly became easier, which services became cheaper, and which “revolutionary” feature quietly disappeared because nobody cared.

    That is usually where the real story is.

    The future does not always arrive with a dramatic soundtrack.

    Quite often, it sneaks in through small updates, lower prices, better interfaces, faster workflows, and habits we adopt without even noticing.

    And those boring little changes matter.

    Because that is where decisions get made.

    If someone gives you a very confident forecast about AI, in either direction, be a little suspicious.

    Not of them.

    Of the confidence.

  • Agree on a Word With Your Family

    There’s a scam worth talking about because it exploits something stronger than technology: panic.

    You get a call from someone you love. They sound frightened. Maybe they say they’ve had an accident, been arrested, lost their wallet, or need money immediately.

    The voice sounds right.

    That is the dangerous part.

    Voice-cloning tools can now imitate someone from a short audio sample, including material taken from videos posted online. The FTC has specifically warned about scammers using cloned voices in fake family-emergency calls. Consumer Advice

    The good news is that defending yourself does not require another app, another subscription, or a degree in cybersecurity.

    You need a word.

    Agree on a private word with your family. Something random enough that nobody could guess it from social media, birthdays, pet names, schools, football teams, or the usual trail of clues we scatter around the internet without thinking too much about it.

    Make sure everyone knows it.

    Parents.

    Children.

    Grandparents too.

    Especially grandparents.

    Then agree on one simple rule: if someone calls claiming there is an emergency and asks for money, they have to give the word.

    If they can’t, stop the conversation.

    And call the person back using a number you already know.

    That last part matters. The FTC recommends independently contacting the person supposedly in trouble rather than trusting the incoming call, even when the voice sounds convincing. Consumer Advice

    We tried this in my family.

    At first, everyone laughed.

    It did sound slightly ridiculous, like we were organizing a low-budget spy operation over dinner.

    Then someone asked me to repeat the word so they could write it down.

    That was the moment I knew the idea had landed.

    There are two other things I would remember.

    The first is the combination of urgency, secrecy, and money.

    Scammers like pressure because pressure shortens the distance between fear and action. The FTC warns that fake-emergency scams often rely on urgency, secrecy, emotional pressure, and demands for immediate payment. Consumer Advice

    So when several of those ingredients arrive together, slow everything down.

    Five minutes can be surprisingly powerful.

    The second is even simpler: verification beats instinct.

    You do not need to decide whether the voice sounds artificial.

    You do not need to detect tiny glitches.

    You do not need to win an argument with a deepfake.

    You only need to check.

    Call back.

    Message someone else in the family.

    Ask the question that interrupts the script.

    AI may be making scams more sophisticated.

    Our defenses do not always have to be.

    Sometimes the best security system is still a family dinner and one ridiculous secret word.

  • Everyone Wants to Be Your Default

    Listen to the way the big AI companies talk and you could easily believe the race is mainly about intelligence.

    Whose model is smarter.

    Whose benchmark score is higher.

    Whose assistant can reason better, write better, code better, summarize faster, understand more.

    Maybe.

    But I’m not sure that is the whole contest.

    Think about how you search the web.

    You do not wake up every morning and conduct a careful comparison of search engines before deciding where to type your question.

    You use the one that is already there.

    In the browser.

    On the phone.

    Inside the device you bought.

    Wherever your thumb lands first.

    Convenience has always been an underrated form of power.

    And I suspect something similar is happening with AI.

    The company that manages to place its assistant inside your phone, your email, your browser, your office software, your operating system and the tools you already use every day may not need to convince you that it is objectively the best.

    It may only need to become familiar enough that replacing it feels like work.

    That changes the nature of the competition.

    A slightly better model can lose to a slightly more convenient one.

    A brilliant assistant can remain invisible if nobody opens it.

    And an average one can become part of your routine simply because someone placed the button exactly where you were already looking.

    Which explains some of the apparent chaos.

    Why AI assistants keep appearing inside products that were doing perfectly well without them.

    Why companies offer generous free tiers.

    Why new features arrive with the regularity of supermarket promotions.

    Why every platform suddenly seems determined to become the place where you write, search, organize, create, summarize, plan and possibly decide what to have for dinner.

    It is not only a race for intelligence.

    It is a race for position.

    For habit.

    For the little piece of digital territory where the next question begins.

    You could call it a land grab.

    Just with nicer icons.

    None of this is automatically bad.

    Integration can be genuinely useful. The best tool is often the one that removes a few steps from something you already do.

    But there is a practical detail worth noticing.

    Defaults are powerful precisely because they stop feeling like decisions.

    Once something is built into the software you already use, it becomes very easy to accept it without ever asking whether it is the tool you would have chosen independently.

    So every now and then, it may be worth looking at the AI that has quietly moved into your digital life and asking a very simple question:

    Did I choose this?

    Or was it simply there?

    I’m not suggesting there is anything sinister about that.

    It is how products have won for a very long time.

    The interesting part is noticing when convenience becomes preference before we have actually made the choice.

  • Before You Pay for a Fourth Subscription

    At some point this year, you probably signed up for a free trial of something with “AI” in the name and forgot to cancel it.

    I’ve done that more times than I’d like to admit.

    The problem is not that these tools are bad. Quite the opposite. Most of them are impressive for the first hour.

    The trouble usually starts three weeks later.

    So now I use a very simple test before paying for another subscription:

    What did this tool actually save me from doing by hand last week?

    If I can name something specific, it stays.

    A report.

    A translation.

    A batch of images I needed to clean up.

    A transcript I would otherwise have typed manually.

    A repetitive task that disappeared from my day.

    That is enough.

    But if my answer is something like “well, it’s useful to have”, I already know where this is going.

    Cancel.

    It sounds almost too obvious, but it filters out a surprising number of tools.

    Many AI products solve problems beautifully.

    Sometimes they even solve problems I never had.

    And they often ask me to change the way I work just to justify their existence.

    That is usually a bad sign.

    The tools that survive on my list tend to be much less glamorous.

    A transcription app.

    A writing assistant that catches mistakes when I’m working in a second language.

    A small automation that saves me ten minutes every day.

    Nothing especially exciting.

    Certainly nothing I would bring up at dinner.

    But useful tools have a habit of becoming invisible. They quietly earn their place by removing friction.

    Two habits help me avoid subscription archaeology.

    First, I pay monthly until a tool has survived at least three months of real use.

    Second, I keep a simple note on my phone with one line explaining what each subscription is for.

    If I cannot explain its purpose in one sentence, I probably do not need it.

    AI subscriptions are starting to behave a little like kitchen drawers.

    Everything seems useful when you put it in.

    Then one day you open the drawer and realize you have four corkscrews.

  • Half a Cabbage and a Tuesday

    I had half a cabbage, two eggs, a can of chickpeas, and absolutely no plan.

    So I typed exactly that into a chatbot.

    Mostly as a joke.

    It came back with a skillet recipe, suggested putting a fried egg on top, and told me to finish everything with a spoonful of vinegar.

    I tried it.

    It was good.

    Not restaurant good.

    Tuesday good.

    Which, honestly, is the standard that matters.

    I mention this because conversations about AI tend to drift quickly toward jobs, creativity, productivity, surveillance, or the possible collapse of civilization before dinner.

    Meanwhile, the biggest change in my kitchen is much less dramatic.

    Nobody asks, “What are we eating?” in that tired voice anymore.

    Now we ask a phone.

    And sometimes the phone has a decent idea.

    That small shift fascinates me more than some of the grand predictions.

    AI is becoming useful in ordinary places.

    Not because it is replacing anything spectacular.

    Because it saves five minutes of thinking when you are hungry and nobody wants to decide.

    It still gets things wrong.

    Once, it suggested a cooking time for chicken that I did not trust. I checked it against a recipe I already knew and used the longer time.

    That became a useful rule for me:

    For anything involving food safety, money, or health, I use AI for ideas, not for the final word.

    For the smaller things, though, it has quietly earned a place.

    What can I cook with this?

    How should I word this message to my landlord?

    What can we do on a rainy Sunday?

    How do I use up three slightly tired carrots before they become part of the composting ecosystem?

    Nothing revolutionary.

    No keynote required.

    No dramatic music.

    Just a tool becoming useful in small, almost boring ways.

    And perhaps that is how a lot of technology really enters our lives.

    Not with a grand announcement.

    With half a cabbage and a Tuesday.

  • The Part of Your Job That Was Never Typing

    A friend of mine in marketing told me that she can now produce a first draft of almost anything in about forty seconds.

    Briefs.

    Emails.

    The slightly uncomfortable message to a client who is late paying.

    For about a week, she was delighted.

    Then she noticed something strange.

    She was spending roughly the same amount of time on her work as before.

    The forty seconds had simply moved.

    Before, the slow part was typing.

    Now the slow part was deciding what she actually wanted to say.

    And once she noticed that, she realized something mildly uncomfortable: she had been treating writing as the work, when a large part of the real work had always happened before the first sentence.

    That distinction matters.

    When a task takes an hour, we tend to assume the hour is the value.

    AI makes that illusion easier to see.

    It can make the mechanical part of many jobs much faster. Drafting. Summarizing. Formatting. Rewriting. Organizing information.

    Useful, yes.

    But speed has a way of exposing whatever is left.

    Judgment.

    Taste.

    Context.

    Knowing when to push a client and when to leave things alone.

    Knowing when the obvious answer is probably the wrong one.

    Knowing what matters enough to keep and what can disappear without anyone noticing.

    I suspect this will become familiar in many jobs.

    The repetitive part gets cheaper.

    The decision-making becomes more visible.

    That is slightly worrying if most of your value came from doing something quickly.

    It is much better news if your value came from knowing what was worth doing in the first place.

    Here is a small test I find useful.

    The next time you use AI for something at work, notice where you slow down.

    If you are waiting for the tool, that is one thing.

    If you are staring at the screen because you do not know what to ask for, what to keep, or what you actually mean, that is something else.

    That pause is interesting.

    It may be the part of your job that was there all along.

    You just could not see it because typing was standing in front of it.