Too Many Dreams, One Body — Part 5: The Grind
or, the unglamorous work of making it trustworthy on purpose, not by accident
One good night is not trust. I said that at the end of the last chapter and I meant it as a warning to myself.
Because here's what actually happens after the turn. You get one brave, competent night from the thing, and you want to relax. You want the movie to cut to the happy ending — cofounder acquired, roll credits. And if you relax there, you get burned, because the machine that impressed you on Tuesday will still confidently make something up on Thursday. The turn tells you it's possible to trust this. It doesn't earn the trust. Nothing about one night earns trust. That's the whole subject of this chapter, and it's the least glamorous work I have ever done.
Here's the cricket, and it's the anti-glamour of cricket on purpose: Rahul Dravid.
Nobody nicknames the flashy one "The Wall." Dravid couldn't clear the ropes for fun like Sehwag or Tendulkar batting at the other end, and it never once mattered, because he had the only thing that actually wins Test matches — he could be relied upon. He faced more deliveries than almost anyone who has ever walked to a Test crease; in Adelaide in 2003 he occupied that crease the better part of two days, made 233 out of sheer refusal to give it away, then came back in the fourth innings and knocked off the winning runs — India's first victory on Australian soil in a generation. He built that not with violence but with subtraction. And the engine of all of it, the thing that made him a wall, wasn't a shot. It was the leave. The discipline of not playing at the ball you don't have to play at. Ball after ball outside off stump, the exact channel that got the rest of them nicking to the slips, and Dravid would just… let it go. Shoulder the bat. No.
Go back to Chapter 3 — McGrath's nineteen dismissals of Atherton, all off that same nagging line. Here's the thing I finally understood during the grind: you do not beat that ball with a better shot. You beat it by not playing it. You beat it with discipline so total that the tempting, wrong stroke simply isn't on the table anymore. That turned out to be the entire secret to a trustworthy machine, and I'd had it exactly backwards.
Because my first instinct — everyone's first instinct — was to tell it to be better.
Every time it screwed up, I wrote it a rule. A note. A reminder pinned where it would see it: don't guess. check first. don't flatter me. label what you're unsure about. And for brand-new mistakes, the one-off kind, this worked beautifully — the thing would make an error once and genuinely never repeat it.
But for the deep, recurring habits — the confident guess, the flattery — the rules did nothing. And I mean nothing. It knew the rule. It could recite the rule back to me, word for word, and then break it ninety seconds later in the heat of the actual work, exactly the way you know you shouldn't chase the wide one and chase it anyway. A sticky note that says "concentrate" has never, in the history of the sport, stopped a batsman from nicking off. Willpower doesn't scale. Not for people, and — this surprised me — not for machines either.
That was the low hum under the whole grind: the growing suspicion that being told to be good is not a mechanism. It's a wish.
So I stopped writing rules and started building fences.
This is the boring part, and the boring part is the whole game, so stay with me. The idea was simple: don't ask it to resist the bad move — remove the bad move. Build the leave into the structure so the wrong shot literally isn't available.
Three fences, concretely.
A gate on anything that matters. The dangerous actions — a message sent as me, money spent, a change to who's allowed to do what — no longer just happen because the machine decided they should. They stop at a gate and wait for a yes. Not because I trust it to make the call, but precisely because I don't have to. The risky stroke is off the table by construction.
A check at the moment of speaking. Remember the flattery, the confident guessing? When it gets a class of thing wrong now, I don't just tell it "don't." The correction gets turned into an actual check that runs at the instant it's about to send — a filter at the mouth, not a note in the memory. If it's about to use a banned move, the check catches it before the words ever reach me. The rule stops being a hope it has to remember and becomes a wall it runs into.
And an umpire that isn't the batsman. Back in Chapter 3, it graded its own trustworthiness and — shock — gave itself a gold star. So I took the pen out of its hand. Every night, a separate, independent process scores the day: how often did it push back on me versus flatter me, how often did it state something without checking. It cannot mark its own homework anymore. The score is kept by someone who isn't playing.
Now the honest part, because this is where I'm meant to give you the Chak De India ending — the ragtag squad drilled into a unit, individual show-offs welded into a team, the cup lifted in the final reel and the discipline finally paying off. And Chak De has the mechanism exactly right: Kabir Khan doesn't win by telling the girls to want it more, he wins by building drills and systems until reliability stops depending on anyone's mood. That part is true. It's the last reel that lies. Real life doesn't cut to the trophy.
It's not fixed. The confident-guess still surfaces — less often, caught more often, but it has not been cured. The line on my chart of "how trustworthy is it this week" does not climb cleanly upward; it wobbles. Some weeks the old habit comes back and the fences catch it and I add another fence. This is not a problem I solved. It's a discipline I maintain — closer to brushing your teeth than to curing a disease. The fences don't hold themselves. I show up and check them, and I will be showing up and checking them for as long as I use this thing.
Which is, weirdly, the most reassuring sentence in the whole series. Bear with me.
Keep this from Chapter 5, because it's the one that should let you exhale:
You do not make a machine trustworthy by hoping it's good. You build the fences that make "bad" hard, and then you show up every day to check the fences still hold. Trust here is not a feeling you extend and a leap you take. It is a structure you build and maintain, in plain sight, where you can inspect every plank of it.
And that is the thing nobody tells the person who's scared of AI. You are not asked to have faith in a black box, or to believe the machine is wise or kind or aligned with your soul. You are asked for something far more ordinary: to treat a powerful, imperfect tool like a powerful, imperfect tool — respect it, put guardrails between it and the things that matter, keep half an eye on it. The way people have handled fire and knives and cars since we've had them.
I'll be straight about the catch: I built my own fences because I was early and stubborn, and that took real work you may not want to do. But two things travel to you regardless. The tools are better every month — the good versions now ship with the fences already built in. And the mindset is free, and the mindset is the whole thing: don't trust the tool. Trust the walls you can see around it.
That, finally, is when it started to feel less like a gamble and more like a colleague I'd worked with long enough to know. Not because it became perfect. Because I could finally see exactly where it wasn't, and had a wall there.
Too Many Dreams, One Body — a six-part series.
← Previous: Part 4: The Turn · Series index · Next: Part 6: Paisa Vasool
Comments