The slot machine metaphor mirrors evolution perfectly.
Evolution generates thousands of variations — most fail, few survive. AI does the same. The human bottleneck was never creativity, it was volume. One person can't produce 100 serious drafts simultaneously.
Our key skill becomes judgment: not generating — selecting.
Ruben, A month ago my manager shared one of your LinkedIn posts and made the case that we all need to become AI-native. I figured if I just spent three months happily messing around with Claude, I'd end up reasonably good at this, and I talked myself into believing I already was. Then a couple of days ago I started actually reading your newsletter, properly, and it finally hit me that there's a much easier, more direct path I'd been walking right past.
Thank you for all of this. I feel like I'm at the front edge of an AI addiction, and I'd like to keep it in check and put it to work as the excellent assistant it can be, which means I'll be coming back to you more often. Thank you.
From a new fan in Korea, a country that was once home to you, too.
Thank you. A great article! I appreciated the parallel with grief. This article helped my thinking about the AI workflow and the flaws I have faced in my own understanding of AI.
The one aspect I feel AI use has most helped my output is when I have iterated a successful prompt, includes the elements you describe, and being able to run the process again multiple times.
Providing the limitations and clear examples -linguistically, the format etc has been a big help.
But seeing AI more as a 'slot machine' means I'll definitely be asking for more versions, and requesting clarifying questions from now on to help me 'play the odds' better.
It's exactly what it needs to be understood in order to get results from an AI agent: in order to accomplish a task, it needs either human feedback, or measurable error feedback.
If you want to delegate something, the only way is to find a way to measure the error of the result, because chances are that the error is high.
If you find it, you can use any LLM (even cheaper ones are ok) and ask the agent to try different results until it finds a good one. Retry-and-error is the path.
I feel like I'm at the point in my journey where I've spent scores of hours working with AI on tasks large and small. I've hit the jackpot a few times and it is truly impressive when you do. I'm using a similar about me and style guide approach I learned here and that has been a game changer. I've figured out many ways not to do things (read failed fast) and am making progress.
The problem is I find myself working longer and harder trying to get AI to help me and the decision fatigue sets in and I'm left wondering if I'll ever see the productivity gains I was hoping for or ever truly finish and release the agents, skills or other projects I'm working on to benefit myself and my team.
Going from cool idea to MVP, not (terribly) hard. Going from MVP to production deployment, so. much. harder. Figuring our when AI is good enough versus taking weeks to get a little better takes a lot of practice and experience to get right. I'm still not there yet.
to me the real way of thinking is where and how am i wanting to manage my systems and workflows in 6 months? AI is moving so fast i sometimes fear what I'm mastering today will be irrelevant tomorrow when another update comes out.....
that fear is really you trying to keep control - nobody keeps up with every tool, including me :)
i stopped trying. one tool per job, and i let the rest go. when the tool gets better, free upgrade. when it gets replaced, i move and bring my folder with me.
the answer is boring: write the task, generate a lot, call out the mistakes, pick the best, finish it yourself.
To deal with the issue that Cowork doesn't always get my files, I created this small start file that I'm hitting on the beginning of each conversation.
The stage that breaks people is the silent one between 3 and 4. You stop blaming the model but haven't yet accepted you're the variable that needs to change. That stretch where you keep trying the same prompt slightly differently is the longest stall. The AI gambler frame names the exit: stop solving for the model, start stacking attempts.
For me, the game isn’t gambling until the output looks right, it’s designing systems where the variance is constrained, and the hard work is problem framing and governance, not just selection. This is crucial especially in my context of gtm for ambitious founders and companies operating in high compliance regulated environments. Happy to connect and talk more with everyone, love this debate! And if you need more brains in your team, I'll relocate yesterday https://www.linkedin.com/in/kallemakinen
People get stuck at AI because they haven't grieved the control they thought they'd have
that's a much harder thing to name, and this piece names it
once you accept you're gambling
the need for control disappears. you pull more levers
The slot machine metaphor mirrors evolution perfectly.
Evolution generates thousands of variations — most fail, few survive. AI does the same. The human bottleneck was never creativity, it was volume. One person can't produce 100 serious drafts simultaneously.
Our key skill becomes judgment: not generating — selecting.
Greetings from Bremen, Germany.
greetings, Filip :)
you get better at selecting after seeing enough bad ones
slot machine pays out if you know what winning looks like
Ruben, A month ago my manager shared one of your LinkedIn posts and made the case that we all need to become AI-native. I figured if I just spent three months happily messing around with Claude, I'd end up reasonably good at this, and I talked myself into believing I already was. Then a couple of days ago I started actually reading your newsletter, properly, and it finally hit me that there's a much easier, more direct path I'd been walking right past.
Thank you for all of this. I feel like I'm at the front edge of an AI addiction, and I'd like to keep it in check and put it to work as the excellent assistant it can be, which means I'll be coming back to you more often. Thank you.
From a new fan in Korea, a country that was once home to you, too.
three months of messing around built your instinct to recognize what good looks like
keep watching the addiction. stay the editor :)
Hard agree, that 2023 Hallucination part
story never gets old :)
the model that burned you in 2023 barely exists anymore
All good points indeed. For those curious to explore and understand more the logic and value of using AI to generate many alternative options look into the "wise winnowing" method. There's a grreat video by Jakob Nielsen that explains it beautifully. https://robingood.substack.com/p/expert-curators-crystallized-intelligence-wise-winnowing-method
thanks for sharing, Robin :)
Thank you. A great article! I appreciated the parallel with grief. This article helped my thinking about the AI workflow and the flaws I have faced in my own understanding of AI.
The one aspect I feel AI use has most helped my output is when I have iterated a successful prompt, includes the elements you describe, and being able to run the process again multiple times.
Providing the limitations and clear examples -linguistically, the format etc has been a big help.
But seeing AI more as a 'slot machine' means I'll definitely be asking for more versions, and requesting clarifying questions from now on to help me 'play the odds' better.
Thank you
thanks for reading :)
not enough people use clarifying questions
pull the lever more than you think you need to. the first output is rarely the one you keep
I like the slot machine metaphor.
It's exactly what it needs to be understood in order to get results from an AI agent: in order to accomplish a task, it needs either human feedback, or measurable error feedback.
If you want to delegate something, the only way is to find a way to measure the error of the result, because chances are that the error is high.
If you find it, you can use any LLM (even cheaper ones are ok) and ask the agent to try different results until it finds a good one. Retry-and-error is the path.
tricky part is skipping the definition of wrong before starting so you just stare at output and shrug
once you've got a measurable error, even a bad model gets there eventually :)
The AI gambler idea makes sense. My best outputs usually show up after I've already said “one last try” three times.
the slot machine doesn't care how many times you've pulled, neither should you :)
"Because I trained Claude Cowork on knowing who I am & my taste"
That is more honest than most AI tutorials I have read this year.
You have to know what you want before you can teach it.
do the work of knowing yourself well enough to write it down
you bring the want, Claude brings the volume
I feel like I'm at the point in my journey where I've spent scores of hours working with AI on tasks large and small. I've hit the jackpot a few times and it is truly impressive when you do. I'm using a similar about me and style guide approach I learned here and that has been a game changer. I've figured out many ways not to do things (read failed fast) and am making progress.
The problem is I find myself working longer and harder trying to get AI to help me and the decision fatigue sets in and I'm left wondering if I'll ever see the productivity gains I was hoping for or ever truly finish and release the agents, skills or other projects I'm working on to benefit myself and my team.
Going from cool idea to MVP, not (terribly) hard. Going from MVP to production deployment, so. much. harder. Figuring our when AI is good enough versus taking weeks to get a little better takes a lot of practice and experience to get right. I'm still not there yet.
when we don't have a clear enough picture of what "done" looks like, we keep generating, keep iterating, and the time just piles up
what does shipped look like? and what does good enough mean for you specifically?
AI will never answer that for us, its our job
🎰 🤔 What an encouraging way of looking at it! Thank you, Ruben!
you’re welcome, keep generating!!
The example was very interesting.
super cool - which one? :)
to me the real way of thinking is where and how am i wanting to manage my systems and workflows in 6 months? AI is moving so fast i sometimes fear what I'm mastering today will be irrelevant tomorrow when another update comes out.....
that fear is really you trying to keep control - nobody keeps up with every tool, including me :)
i stopped trying. one tool per job, and i let the rest go. when the tool gets better, free upgrade. when it gets replaced, i move and bring my folder with me.
the answer is boring: write the task, generate a lot, call out the mistakes, pick the best, finish it yourself.
To deal with the issue that Cowork doesn't always get my files, I created this small start file that I'm hitting on the beginning of each conversation.
/start
https://ibb.co/RpkqPv3b
So this way, I'm always on the track.
cool, i have a separate folder that i point to Cowork. works just as good :)
The stage that breaks people is the silent one between 3 and 4. You stop blaming the model but haven't yet accepted you're the variable that needs to change. That stretch where you keep trying the same prompt slightly differently is the longest stall. The AI gambler frame names the exit: stop solving for the model, start stacking attempts.
the real deal is to stop hoping and start treating bad outputs as the cost of doing business
volume is not embarrassing when its the point
For me, the game isn’t gambling until the output looks right, it’s designing systems where the variance is constrained, and the hard work is problem framing and governance, not just selection. This is crucial especially in my context of gtm for ambitious founders and companies operating in high compliance regulated environments. Happy to connect and talk more with everyone, love this debate! And if you need more brains in your team, I'll relocate yesterday https://www.linkedin.com/in/kallemakinen
thanks for sharing!!