Success with AI is Probabilistic and Not Deterministic
Why the best AI users think like poker players, not chess masters
I watched my friend founder blow $50,000 on AI tools last month.
Not on API costs. Not on the computer. On the wrong mental model.
He treated AI like a vending machine. Insert prompt, receive output. If the output was bad, the machine was broken.
He never pressed the button twice.
By the end of Q2, his competitor, a solo developer with a $500 monthly AI spend, had launched three products that were successful. The founder had launched nothing.
Same tools. Same models. Same access.
Different understanding of how probability actually works.
The Vending Machine Illusion
Here’s what most people believe about AI:
Good prompt + Good model = Good output
It’s clean. It’s logical. It’s completely wrong.
What actually happens:
Good prompt + Good model = One sample from a probability distribution
That sample might be brilliant. It might be garbage. It might be somewhere in the vast middle ground of “acceptable but unremarkable.”
You don’t know until you look.
And here’s the part everyone overlooks: looking once tells you almost nothing about what that distribution actually contains.
Think about it like this. Imagine you have a slot machine that occasionally pays out $10,000, regularly pays out $100, and usually pays nothing. If you pull the lever once and get nothing, what have you learned?
Almost nothing.
You certainly haven’t learned that the machine doesn’t pay.
Yet that’s exactly how most people use AI. One generation. One output. Move on.
They’re playing a slot machine that actually has good odds, and they’re walking away after a single pull.
The Mathematics of Iteration
Let me show you the math that changed how I use AI.
Let’s say there’s a 20% chance any given generation produces something genuinely useful. Not just acceptable. Useful. The kind of output that actually moves your work forward.
If you generate once, there is a 20% chance of success.
If you generate 5 times, there is a 67% chance of at least one success.
If you generate 10 times, the 89% chance of at least one success.
If you generate 20 times, there is a 99% chance of at least one success.
The formula is simple: 1 minus (0.8 to the power of n), where n is the number of your attempts.
But here’s what the math obscures:
Generation 17 might be the best output the model could ever produce for your specific prompt.
You’ll never know if you stopped at generation 3.
I call this the Instances Lottery, and it’s the most overlooked principle in AI productivity.
As I wrote in my piece on “The Stochastic Mindset“:
“Sometimes you need to embrace uncertainty, not fight it. Generate options, not perfect plans.”
But embracing uncertainty isn’t passive. It’s actively playing the probability game until the odds tilt in your favor.
Why Deterministic Thinkers Fail
The people struggling most with AI right now are the ones with the most deterministic mindsets.
Engineers who believe in repeatable processes. Lawyers are trained to find the single correct answer. Executives who want a guaranteed ROI before committing.
They’re applying factory thinking to something that doesn’t work like a factory.
AI is not a production line. It’s more like jazz.
You play. You improvise. Sometimes it clicks. Sometimes it doesn’t. You keep playing until something beautiful emerges from the noise.
The deterministic thinker sees ten mediocre outputs and concludes the tool is broken.
The probabilistic thinker sees ten mediocre outputs and thinks: “I haven’t sampled enough of the distribution yet.”
One gives up. One keeps rolling.
Guess who wins?
The Hidden Cost of Stopping Early
The opportunity cost of not iterating.
Every time you accept a mediocre AI output because you’re tired of regenerating, you’re leaving value on the table.
Not a little value. Massive value.
I tested this last month with LinguaLive, the language learning app I’m building with Gemini’s Live API. I needed a particular interaction flow. The kind of thing that makes or breaks user retention.
First generation: Clunky. Functional but forgettable.
I could have shipped it. Most people would have.
Instead, I regenerated 47 times.
Generation 47 wasn’t just better. It was a completely different concept I never would have imagined. It solved problems I didn’t even know existed in my original framing.
The time cost? Maybe 30 extra minutes.
The value difference? Probably months of product-market fit acceleration.
That’s the hidden math of probabilistic AI use.
The best outputs aren’t marginally better than average outputs. They’re categorically different. And they only appear after you’ve exhausted the boring middle of the distribution.
What the Poker Players Know
The best AI users I know don’t come from tech. They come from poker.
Poker players understand something most people don’t: you can make the right decision and still lose. You can make the wrong decision and still win.
What matters isn’t any single hand. What matters is expected value over many hands.
They don’t ask: “Did that work?”
They ask: “Is this a positive expected value strategy across hundreds of attempts?”
Apply this to AI:
The question isn’t whether this prompt worked this time.
The question is whether this approach, with sufficient iteration, consistently produces valuable outputs.
One failed generation isn’t data. It’s noise.
Ten failed generations might be data.
Ten failed generations with three brilliant ones mixed in? That’s a strategy.
The Temperature Dial Nobody Uses Correctly
Let me clarify a feature that many people misunderstand: temperature.
Temperature controls randomness. Low temperature (0.1-0.3) makes outputs more predictable. High temperature (0.7-1.0) makes them more creative but also more chaotic.
Here’s how most people use it: they don’t. They leave it at default and hope for the best.
Here’s how probabilistic thinkers use it:
For convergent tasks (code, analysis, factual writing): Low temperature, moderate iteration. You’re looking for the most probable correct answer.
For divergent tasks (ideation, creative work, strategy): High temperature, heavy iteration. You’re deliberately sampling from the long tail of the distribution.
The magic happens when you realize these are different games with different optimal strategies.
Low temperature + heavy iteration = diminishing returns fast (the outputs are too similar).
High temperature + heavy iteration = increasing returns (each output explores genuinely different territory).
Most people run high-stakes creative work at low temperatures with minimal iteration.
They’re doing the opposite of what the math suggests.
The 100x Framework
Here’s the framework I now use for any significant AI task:
Step 1: Define “success”
Before you generate anything, know what you’re looking for. Not vaguely. Specifically. “A metaphor that makes distributed computing feel intuitive to non-technical readers.” Precision in your success criteria makes rapid evaluation possible.
Step 2: First pass: 10 generations, high temperature
Don’t evaluate deeply. Skim for anything surprising. You’re not looking for polished. You’re looking for unexpected directions.
Step 3: Identify the interesting outliers
Usually, 2-3 out of 10 will have something. A phrase. An angle. A structure. Pull those out.
Step 4: Second pass: 10 generations per outlier, medium temperature
Now you’re exploring specific veins. Take the best elements and regenerate around them.
Step 5: Third pass: 5 generations, low temperature
Refinement. You’ve found the gold; now you’re polishing it.
Total generations: ~45. Time: Maybe 20-30 minutes for something that would take hours to develop from scratch.
Most people do step 1 (poorly) and stop.
They’re mining with a spoon when they have a drill.
The Emotional Resistance
I know what you’re thinking: “I don’t have time to generate 50 versions of everything.”
You’re wrong.
You don’t have time NOT to generate 50 versions of essential things.
Consider the alternative: You spend 4 hours manually crafting something mediocre when you could have spent 30 minutes iterating to something brilliant.
The resistance isn’t about time. It’s emotional.
It feels like “cheating” to regenerate endlessly.
It feels like you “should” be able to get it right the first time.
It feels like iteration is admitting you (or the AI) failed.
That’s your deterministic conditioning talking. The factory mindset that says good output requires skill and effort, not probability and volume.
Drop it.
Iteration isn’t failure. Iteration is the strategy.
Instances
Every single interaction with AI is a dice roll. A cosmic lottery where the same prompt can get you gold or garbage, genius or gibberish.
In Instances, the house edge is actually in your favor. You have to keep playing.
The Compound Effect
What most people miss about probabilistic success.
It compounds.
Every time you iterate to a brilliant output instead of settling for mediocre, you’ve raised the baseline for your next piece of work. The brilliant output becomes the reference point. The training data is for your own taste.
Over time, your bar for “acceptable” rises. Your ability to recognize brilliance in a sea of mediocrity sharpens. Your iteration becomes more efficient because you know what you’re looking for faster.
Meanwhile, the person who never iterates stays stuck. Their reference point is mediocre. Their taste never develops. They don’t even know what brilliant looks like because they’ve never sampled deep enough in the distribution to find it.
This is the real AI skill gap.
Not prompt engineering. No technical knowledge. No access.
The willingness to iterate until probability becomes certainty.
The Leaderboard Fallacy
Every week, I see new “AI tool comparisons” where someone tests GPT-4 vs Claude vs Gemini with a single prompt and declares a winner.
This misses the point entirely.
The variance within any single model across multiple generations often exceeds the variance between models on single generations.
Translation: Claude’s 10th attempt at your task might beat GPT-4’s 1st attempt, even if GPT-4’s 1st attempt beats Claude’s 1st attempt.
Model selection matters less than iteration strategy.
I’m not saying models are identical. They’re not. Each has genuine strengths. But obsessing over which model to use while ignoring how many times to use it is like optimizing which dice to roll while only rolling once.
Building Probabilistic Systems
If you’re building products on AI (like I am with LinguaLive and NoCancerAI), the probabilistic nature isn’t a bug you work around. It’s a feature you design for.
User-facing AI products need:
Multiple generation paths (not just retry, but alternative approaches)
Quality scoring to surface the best options automatically
Graceful handling of the inevitable bad outputs
User control over how much to iterate
Internal AI workflows need:
Batch generation by default
Automated first-pass filtering
Human review concentrated on the most promising candidates
Metrics on iteration depth vs. output quality
Most AI products are built by deterministic thinkers who treat variance as a problem to eliminate.
The best AI products will be built by probabilistic thinkers who treat variance as a resource to exploit.
The Instances Lottery Winners
Let me tell you about three people I know who’ve internalized this:
Founder A runs a content agency. When ChatGPT launched, she built a simple workflow: every piece of content gets 25 variations generated, then ranked by a lightweight scoring prompt, then the top 3 go to human review. Her team’s output quality went up while their production time went down. She’s not hiring prompt engineers. She’s hiring people who can evaluate quickly.
Engineer B is building an AI coding assistant. He realized that 10 parallel solution attempts with different approaches beats 1 carefully reasoned attempt every time. His product doesn’t show users a single answer. It shows a ranked list of possible solutions with confidence scores. Let the user iterate.
Creator C makes AI-generated music. He told me he generates 800-1,000 vocal stems for every usable 30-second clip he publishes. Not a typo. 800-1,000. His hit rate is around 0.1%. But that 0.1% is genuinely good. Good enough that listeners can’t tell it’s AI. Good enough for real streaming numbers.
None of them are more talented than average.
All of them understand probability better than average.
The Uncomfortable Truth
The uncomfortable truth nobody wants to admit:
You probably aren’t underperforming on AI skills. You’re underperforming on iteration volume.
The prompts matter less than you think.
The models matter less than you think.
The number of times you’re willing to roll the dice matters more than almost anything else.
This is both frustrating and liberating.
Frustrating because there’s no silver bullet. No magic prompt that always works. No technique eliminates the need for iteration.
Liberating because the path forward is simple:
Generate more. Evaluate faster. Select ruthlessly. Ship what survives.
That’s it.
You don’t need to be smarter. You don’t need better tools. You don’t need more training.
You need to embrace that success with AI is probabilistic, not deterministic. And then play the probability game like you actually believe it.
The Assignment
This week, pick one crucial task. Something that matters. Something you’d usually approach carefully and try to get right the first time.
Now change the approach:
Generate 50 variations, generate those instances.
Spend no more than 10 seconds evaluating each
Pull the top 5
Generate 20 more variations based on what made those 5 interesting
Pick the winner
Time yourself. The entire process should take no more than 30-45 minutes.
Then compare: is your 50th generation better than your 1st would have been if you’d stopped there?
I already know the answer.
The instances lottery is real. The only question is: are you buying enough tickets?
Post Credit Scene
The content this week.
“The Bear” (Hulu)
If you haven’t watched this yet, stop reading and go. A chef inherits his brother’s chaotic Chicago sandwich shop and tries to turn it into something great. Season 3 is out. What makes it brilliant: every episode is about iteration under pressure. Carmy fails constantly, adjusts, fails differently, adjusts again. It’s basically the probabilistic mindset applied to beef. Also, the one-shot episode “Fishes” might be the best hour of television this decade.
“Stranger Things” Season 5
It’s finally live. I wanted to love it. I don’t. The nostalgia engine ran out of fuel somewhere around episode 3. If you’re a completionist, go ahead. If you’re not, your time is better spent elsewhere. Sometimes even beloved franchises hit their probability ceiling.
“Fooled by Randomness” by Nassim Nicholas Taleb. The essential text on why humans catastrophically misunderstand probability. Written before AI, but reads like a prediction of exactly these challenges. Chapter 3 on sample size is worth the entire book.
“Severance” Season 2 (Apple TV+)
Just wrapped. Without spoilers: the finale delivers. The whole show is about split realities and parallel versions of self, which maps eerily well onto how we interact with AI instances. Your “innie” and “outnie” are basically different temperature settings of the same model.
“Thinking in Bets” by Annie Duke — Former pro poker player on decision-making under uncertainty. Short, practical, no fluff. The framework for separating outcome quality from decision quality is exactly what AI users need to internalize.
“Anora” (2024)
Sean Baker’s Palme d’Or winner. A stripper marries the son of a Russian oligarch, chaos ensues. Nothing to do with AI. Everything to do with how life rewards people who keep rolling dice when others would quit. Also just genuinely fun.
Thanks for reading.
Vlad