Learning AI Is Hard. Really.

I’ve been using AI for about three years now. ChatGPT from the early days. Perplexity when it got properly useful. A tiny bit of Gemini. A little poke at DeepSeek. But predominantly Claude Code via Anthropic for the last good chunk. Not playing with it. Not writing haikus about it. Using it. Every day. Real work, real shipping, real pain.

Here is the uncomfortable bit. The more I use it, the more I realise how little I actually know.

And there is a knock-on effect. I have developed a very short fuse for AI clickbait.

Three years in, still tip of the iceberg

The last nine months have been intense. Twelve to fifteen hour days. Weekends included. Roughly 2.5 years of a normal working life compressed into nine months of sprint (which I should probably talk to a doctor about). Add the two-plus years of normal-pace AI use before that, plus the equivalent of about five years of ordinary working hours if you do the maths, and you would think I had a handle on it.

I don’t.

What I actually know feels tiny. Tip-of-the-iceberg tiny. And the iceberg under the water? Growing so fast that even doing fifteen-hour days I cannot keep up. New model every month. New feature every week. Claude 4.6, then 4.7, then whatever is next by the time this post goes out. Anthropic ships faster than I can read their changelogs. So does OpenAI. So does Google.

The more you know, the more you know you don’t know. That sentence is older than me. Turns out it is still the truest thing anyone has ever said about learning anything really really hard.

Clickbait, clickbait, everywhere

Meanwhile, my feeds. “I built a full AI agent in two hours.” “Shipped a website in a single click.” “The one prompt that will change your business forever.” “Trading bot in a weekend.” “AI SaaS in one sitting.”

I click. I watch. I close it inside 60 seconds. Every single time, it makes me cringe.

It is the same shape as the YouTube trading gurus. Remember them? The ones selling the one indicator that will change your life forever? The one strategy that beats the market? The one course that turns you into a millionaire by Christmas?

Yeah. That easy. Not.

Why none of it works

Here is the bit the clickbait conveniently skips. An LLM in its raw state is a probability machine. Big chain of probabilistic guesses sitting on top of the same 1s and 0s I learnt about at college. I am not an LLM expert. Nowhere close. But I remember enough to know that LLMs are trained to predict the next word or dataset in a sequence. The word “predict”… Probability is the Measure. Prediction is the Output. Yes?

Which means the prompt that gave them their shiny demo? Run it yourself tomorrow and you will get something different. Slightly different. Or wildly different. Often both on the same day. The output is seriously, genuinely random.

You follow the video step by step. You get a thing that looks nothing like the thing they built. You rewind. You retry.

What made the final cut is the take that worked, after however many retakes and a tidy edit on top (perfectly normal video production, they just don’t tell you that part). What you saw is the highlight reel, not the process.

My 60-second filter

Here is the test I apply now. When I watch an AI tutorial, blog, or video, I am listening for three words. Guardrails. Harness. Workflow.

Does the creator talk about checking the output? Running tests? Catching the model when it hallucinates? Wrapping the LLM in something that makes its behaviour reproducible? Or are they just pasting raw prompts into ChatGPT, Claude, or whichever LLM is flavour of the month, and high-fiving the first half-decent result?

If it is the second one, I close the tab. That is not teaching you how to use AI. That is teaching you how NOT to use AI (just another person happy to show you the wrong way for views).

That is why having a harness matters. (Mine is SST3-AI-Harness. Use it, fork it, or roll your own. The point is you have one.) I built it because I was sick of the AI quietly wandering off task while I blinked. A harness doesn’t kill the randomness (nothing does, that is how LLMs work). What it does is pull every output closer to what you actually asked for and catch the drift early, run after run. Closer every time. That is all you need.

More on that here: Why Do We Need an AI Harness?

The real ones never say it’s easy

Here is the other tell. Notice who is actually selling the shortcut and who is not.

The people I trust on AI (the ones doing the real work, day in, day out) never boast about how easy it is. They do not. They cannot. They know how hard it is, because they have lived it: the same prompt working on Monday and producing rubbish on Friday, the same conversation with the same LLM going sideways on a Tuesday for no reason anyone can pin down, the long evenings spent undoing what the model confidently got wrong.

They stay quiet about the difficulty, or they tell you it is hard if you ask.

If someone is telling you it is easy? That is the tell sign. The charlatans boast. The people who have actually done the work do not. Watch out for the loud ones. The easier they make it sound, the less they have actually built.

No shortcut. Never was

There is no shortcut in life. There is no free lunch. Not in trading. Not in dance. Not in business. Not in AI.

You need discipline. Consistency. Grit. It will be frustrating. You need to put in a lot of time and effort to learn the trade the hard way, the real way. Back to basics. Understand the tool, not the hype. Break it. Fix it. Break it again. Read the source. Read the changelog. Do the reps when nobody is watching. Build something tiny and real, rather than something huge and imaginary. Test test test, explore explore explore, and you’ll discover a lot!

AI is a force multiplier, not a crutch. If you know what you are doing, it multiplies your work. If you do not, it multiplies your mistakes faster than you can spot them (and will do it with a smile). Take the twenty years of graft away, and all I would have is a very confident junior engineer that lies to me with extreme eloquence.

I’m a realist

I don’t like to bullshit. I don’t like to sugarcoat. I prefer to speak bluntly.

That doesn’t mean I speak rudely or disrespectfully. I try to be gentle and political because it works sometimes, but not always. A reality check is sometimes needed (tough love you know). So I also use bluntness (again, it’s just a tool, just one of my favourites lol).

I also believe, deeply, that everyone is capable of learning whatever they want, given the time and the opportunity. Or the willingness to create that opportunity when one does not show up at the door (spoiler: it won’t).

The only way out is tough it out and try to have fun and find enjoyment along the way. There is no hack. There is just the work. You put in the hours, and the hours compound. Three years. Nine months. Fifteen-hour days. Still tip of the iceberg. Still loving every minute of it. Mostly… just don’t ask me in early mornings.

If you are at the start of learning AI, or any other really hard thing: welcome to the grind. Settle in. Trust the process. Do the reps. Ignore the two-hour merchants.

See you on the other side of the iceberg. In about ten years.