At the top of my career at Verizon I was turning down promotions. I could see up around the corner what AI was going to do to every industry, so at 40 I took a voluntary separation, walked away from everything I'd built, and sat down in front of a black screen terminal to learn.
That was two years ago. Here's the bill.
101 billion tokens. 95,000 prompts, 11,000 commits, 3.4 million net lines of code, 96 projects. At list API rates that's about $135,000. What I actually paid was closer to $20,000 in subscriptions and APIs, plus about $15,000 in hardware, and I don't think most people have noticed that gap yet. None of those numbers measure skill. They measure hours. The skill came from the 35 tools I threw away, the production database I deleted by accident, and the seven features I ran at once and didn't ship a single one of them faster.
The numbers and what they measured
101 billion tokens isn't a skill number. It's an hours number. Two years, 10 to 12 hours a day, no weekends carved out, which by my own count is 8,000 to 9,000 hours minimum. The tokens are just what that time looks like once it's been turned into machine reading and writing.
95,000 prompts is the one that actually tracks learning. About 130 a day. Early on my prompts were paragraphs, every edge case spelled out, whole design briefs pasted into a box. The ones that work now are two sentences. I've noticed the more I add to these things, the more they go haywire. They need to be short and leave room for the AI to do its thing.
11,000 commits, 3.4 million net lines of code added, and I want to be honest about that one. I didn't type most of it. Two years ago I couldn't have written one line of it by hand. But I read every diff that touched money, auth or a delete. The hard part was always knowing which lines to ask for, and two years in that still hasn't gotten any easier.
96 projects. Most of them nobody will ever see.
What the money actually bought
$135,000 is the API-equivalent number, what the invoice would've been if I'd paid list rates for every token. I didn't. I paid roughly $20,000 in subscriptions and APIs across two years, plus about $15,000 in hardware. Two of those machines are Mac Studios I bought in early 2026 after a builder I follow warned that local models were about to matter. Felt aggressive at the time. Looks obvious now.
Biggest cumulative spend, if the shape helps: ChatGPT $2,935. KlingAI $2,689. Grok $2,410. Claude over $2,000. Cursor over $2,000.
$135,000 against $20,000. Subscription pricing is a really good deal right now for anybody willing to sit in the chair ten hours a day. I don't think it holds. If you're waiting for AI to get cheap before you start, you've got it backwards, it's cheap right now for what it does, and that closes as the tools get better at charging what they're worth.
The 35 tools I dropped and why
70 tools tried, 35 kept, 35 dropped. The 35 that died all died the same way. I bought them before I had a problem for them.
Something looked like the next thing, everybody was posting about it that month, and I'd sign up before I could name one task in my week it would replace. Then I'd go looking for a problem the tool could solve, which is exactly backwards, and it sat there unopened until I cancelled it.
What came out of that pile: start with the problem first. Never the tool. Find the friction, the thing you hate doing, and then solve that with any of the top tools. The tool didn't matter. I could've used Codex, Claude, or three other things. What mattered is I had a problem and I solved it surgically.
The 35 that stayed all pass one test. I can name the thing I stopped doing by hand.
The three things I would do first if I were starting today
One, speech to text, full time. I dictate everything and I never go back and read it before I hit enter. About 2,600 of those dictations turned into a voice profile an AI now writes from, which I never planned, it just happened. The real benefit is I think out loud faster than I type, so the model gets the messy version of my thinking, and the messy version is the one with the actual constraints in it.
Two, a second model reading the first one's work. Not on everything. On the risky stuff, payments, auth, deletes, anything doing money math. I logged 822 caught errors in 31 days, a 74.9% hit rate on the reviews I ran, and an earlier log across 53 runs found a real bug 88% of the time. The models lie. Not on purpose, but they do. They'll look you in the eye, tell you it's done, and it's not. When it matters I never use just one.
If you don't write code, two chat windows does the same thing. Paste the first answer into the second one and ask what's wrong with it, what it assumed, what it skipped. Don't tell the second one where the answer came from.
Three, write the criteria before you ask. Oldest habit I brought with me. At Verizon I applied for 14 internal roles and got 14 offers, and that wasn't me being good at interviews. I decided what the job had to prove before I walked in the room, then made the room answer it. Prompts work the same way. Say what done looks like and how it gets checked, then ask. Most of the bad AI output I see, nobody told it where the bar was.
The one thing I would not repeat
Working seven features at once.
I did it because I was waiting on the coding. A model would be running, I'd get restless, spot another issue, not want to forget it, so I'd open another session and start that one too. Seven going at once on the bad days. I didn't ship any of them faster. What I actually built was seven half-finished things and no way to tell which change caused which problem.
Same mistake, more expensive version. In November 2025 I deleted my entire production Supabase project. I was trying to delete a different one. I rebuilt it. Now I've got backups I've actually restored from, a confirm step I can't click past, and a hook that blocks destructive commands before they run. All three exist because of one bad click while my attention was in four places.
What I'd pay for again
So $135,000 on paper, about $35,000 in cash. Honestly the part I'd pay for again is neither one. It's the 35 tools I threw away and knowing why each one died. Every number above is an hours receipt.
Friction first, tool second. Criteria before the ask. A second model reading anything that can cost you money. It took me two years to land on that order. I still catch myself running seven things at once.
