Spending $10,000 on Artificial Intelligence Models: Lessons From 50 Attempts

In ten months I finished more than fifty artificial intelligence experiments and sent over 8,000 prompts. At the start the mood was electric: with the arrival of a new Claude model, people claimed a single prompt could do any job and programmers would become needless. I switched to the Max 20X plan and started producing like crazy. This YouTube release gives a frank account of those days: which attempt worked, which one turned out to be junk, and where the money went. The first lesson stands at the door: as the model grows stronger, appetite grows, and unfinished work piles up in folders.

Looking at my development folder, I counted 58 separate drafts. Work that once needed programmers, designers and managers can now be produced by one person sitting before a model until morning. That is where the real question starts: has artificial intelligence made programmers needless, or has it become a strong lever in the hands of people who know the craft? My answer leans to the latter. Every prompt burns fresh tokens, and every token means money; each draft started without a plan waits on the shelf as a half-built file. Speed is stunning, but without direction speed only grows the mess.

My first big build was a site showing motor insurance values, fed by price lists published by TSB, the Turkish insurers association. I downloaded 11 to 12 years of records covering the value of vehicles sold in the country, a lovely dataset. Millions of people search these values each month because notaries ask for them during a sale. I built the site with a model and generated hundreds of thousands of pages split by brand, version and year. The result was a flop: Google refused to index it. In its eyes this was bulk content with no gain for the user. Today the third version is live with nearly 100,000 pages indexed, but the first version went down in history as junk.

I did not stop. Next came a compact listing site tracking car price moves from the same dataset, with fewer pages, yet the outcome stayed bleak. For the third move I looked abroad; with the translation power of the model I decided to build an English site. I found a 147-year record of names Americans gave their babies. The agency called SSA publishes it openly for everyone: state by state, year by year, name by name. I added a cute ghost theme, meanings, games and dictionary notes, and the new site was born. Google stayed cold again: nearly 150,000 pages left outside the index. Three sites standing there with a few visitors a day.

Today I see more clearly why they failed. The spam policy of Google is plain: producing bulk pages without adding value for users counts as worthless, no matter how it was produced. My sites were exactly that: records were there, labor was there, yet most pages copied each other. On top came my perfectionism: never calling anything finished, always tinkering, overdoing the engineering. I cannot say the learning was free, but the lesson is plain: building a site with artificial intelligence is the easiest part. The hard part is bringing visitors, truly helping people, and turning it into earnings.

On the other side stands the automation I built for our company. Entering product records into Google Merchant Center used to be heavy manual work taking days. With an artificial intelligence setup I built a flow that processes 3.5 to 4 million records within minutes, sends them over, and refreshes prices. Products open to ads stayed current and the team could breathe. This shows the gap between a trial and production: target plain, measure plain, gain plain. Here the model is no decor; it works like a shift. Next to three junk sites, this job was the prize for using the same tool in the right room.

Now the money. On the Claude side I use Max 20X at $200 a month. TechCrunch wrote when this plan came out that it offers up to 20 times the usage, and my bills confirm it: in ten months subscription payments alone neared $2,000. Hosting and Cloudflare costs add $400 to $500 a month. Added up, we set aside nearly $10,000 for artificial intelligence this year. A serious sum for a small company, yet we see the return in our work. My last word: ignore the tales of getting rich with one prompt on social feeds. Bring artificial intelligence into your job, your kid's school, your daily life, but never hand it the wheel.