Stefans Keiss @memphyssk

There is one interview with Rick Rubin, the 60 Minutes one, where he gets asked whether he plays instruments, barely, and whether he knows how to work a soundboard, no, no technical ability at all. So what is he being paid for then? His answer was that the confidence he has in his taste and his ability to express what he feels has proven helpful for artists. And this is the guy behind Red Hot Chili Peppers and Linkin Park. With AI it is the same story, imo, because the tooling and the prompt library and the speed at which you ship are buyable within a week by anybody now, and the only real lever left is your confidence in your own broad, developed judgment, in your taste. Where did your own taste come from, and would you be able to teach it to somebody?

@dhh yup. tests and review agents catch bad code fine. what they never catch is a pipeline that is still built for the product you had three months ago

@jasonlk well, the fear usually sits with people whose value sat in execution only. the ones who keep asking what else is possible here do not need that talk

@LeoBuilds_ hand written code is the new vinyl

imo half of "I embraced AI" is just autocomplete with swagger. shift begins where you stop approving every line.

DHH @dhh

a week ago

The only programmers who are really in trouble with the competition from artificial intelligence are the ones who insist the world today isn't all that different from the one we lived in last year. Embrace the paradigm shift or perish. (This has always been true.)

285 416 6K 214K 628

The longer you postpone learning a new technology, the more complicated it looks to you, which is exactly backwards, because most technologies get simpler over time while your fear of them keeps growing. I keep watching people around me hit this wall with AI, and it reminds me of the App Store around 2011, which was much harder than it is today, because you could not buy traffic inside the store, you could not test two versions of your page against each other, and analytics meant bolting on somebody else's SDK and hoping for the best, so you shipped an app and it either flew or it did not while you had almost no control over which of those happened. Apple only opened search ads in 2016 and only let you run a real A B test on your product page in 2021, so the founders who arrived later actually had the easier ride, not the harder one. That is roughly where AI sits right now, because people can see there is something sweet in there and the tooling has already appeared, and yet they keep standing in the doorway thinking, well, now i have to sit down and press buttons and figure things out, and honestly that thought is the only real barrier left. What i take from this is that adoption will be slower than the tooling suggests, because the barrier was never technical and people do not change how they work on a schedule. Which also means service companies are not going away here, somebody has to sit in that chair and press the buttons for those who will not do it themselves, and that job is going to exist for years. What did you postpone this year that turned out to be easier than you expected?

@realRyanNamba yes, that's the whole trick. tools will become simpler, but business doesn't buy a tool, it buys the fact that someone took its specific problem and translated it into a working thing.

Everyone keeps saying the hard part of AI adoption is the technology, and i spent a couple of thousand dollars on an experiment to check whether that is true. A founder i have known for years runs the accounting firm, and he came to me wanting an AI transformation and asked me to recommend an AI engineer, so i told him there is a tool where you write what you need and it does the rest, and he answered that they need somebody who will do the writing, because they do not understand any of this themselves. Fine, i thought, if you will not type into the machine yourself then we will do it the other way around. I hired a junior analyst for a couple of thousand dollars and told him he is the one pressing the buttons now, and i made the setup as unfavourable as i could on purpose, because i wanted to see where the thing breaks, so he had a very shallow idea of what AI even is. About a month later he shipped them a product they are testing right now, and nobody in that whole chain had to become technical. What this experiment actually says is that the problem with AI right now is not in the tools at all, it is the mental impossibility, for most people, of spending the time to work it out even a little, and that impossibility grows out of two things, a big fear first, and the idea that it is terribly complicated second. There is a gap hiding underneath that, because the longer you postpone sitting down with a new technology, the more complicated it feels, you assume the distance keeps growing while you wait, when in reality most technologies only get simpler with time, since founders keep turning the raw thing into something user-friendly. So the tools were never the expensive part, the fear was, and it costs far more than any tool ever will. So what do you think happens first, mass adoption finally picking up speed, or service companies living long and happily on everybody who will not press the buttons themselves?

A journalist asked me last week why people are so disillusioned with corporate careers, and she clearly expected me to blame AI for taking the jobs, while the honest answer starts decades earlier. The deal made complete sense in the sixties, when the world moved slowly, so you stayed loyal, you worked long, you brought value, and in return you got a decent pay-off and a pension and a union that protected you and the reasonable certainty that nobody fires you out of nowhere. The problem is that this whole period turned out to be an anomaly, because predictability never really existed before it and it does not exist after it, two years ago most of what is happening in the world right now was not there at all. Joining one of the big AI labs today is a semi-entrepreneurial move rather than a corporate one, because you take real risk in exchange for real upside, and everything sitting in the middle stopped making any sense, since the middle pays you an average salary and offers predictability in return while having none of it left to give. The outliers who took the risk and won are also the ones accelerating the whole economy, that unpredictability pushes inflation and money printing, and the printed money lands disproportionately with those same outliers, so life keeps getting more expensive for the people who stayed in the middle while their income sits exactly where it was. So the next time somebody turns down your offer because they need something stable, it is worth working out together what their current job actually still guarantees them.

Predictability was always the anomaly. It did not exist before the sixties and it does not exist now, and the corporate deal was always that anomaly sold back to you for money. An AI lab looks corporate from the outside, but you are paid in a bet there, real risk for real upside, while an ordinary job still charges you a premium for safety it ran out of.

@naval well the problem here is wider than EA, as soon as you hang a label on something good and start publicly flashing your involvement, this turns into a race for PR, and the path itself gets devalued

@yongfook well look, these three have a common thing -- there you won't roll back the error, data is lost and that's it, you can't bring it back. add payments and auth to this list, same logic - fucked up billing and real money is gone

interesting, at what percent the word leads stops meaning "the person approves" and starts meaning "the person finds out post factum"

Mark Kretschmann @mark_k

3 weeks ago

Claude is now helping build its own successor! Anthropic says Claude already leads 26% of its model R&D, up from just 1% in March, and participates in more than 90% of its research work. Around 30,000 AI agents are simultaneously active on its internal platform. It isn't fully

30 7 178 9K 14

the gatekeeper here is not new, just before the publisher rejected the manuscript once a year, and the algorithm rejects your post every day, and this daily dose of rejection hooks faster than any publisher.

Kieran Drew @ItsKieranDrew

3 weeks ago

The paradox of the Internet is that it allows people to get paid for their art. But instead of creating art, most people get stuck creating for an algorithm.

38 6 86 8K 9

Seat-based pricing was never fair to anybody, and agents doing the actual work are about to make that impossible to keep hiding. Take any product tool with two users sitting inside it, a product manager who gets a hundred percent of the value out of it and a developer who gets maybe twenty, and you bill both of them exactly the same. That gap is where the margin of classic SaaS actually comes from, it has been quietly financing this industry for years, and everybody in it knows this. Some verticals started moving to usage-based pricing back in 2018 and 2019, slowly and without much noise, and AI made that shift far faster, so the conversation has already moved past usage and landed on outcomes. HR tech is the clearest pioneer here, where you pay neither per seat nor per resume processed but for a hired person, and that number is defensible in a way that seats never were. IMO, this will spread across the whole industry, and not because founders suddenly became principled, but because the gap becomes impossible to defend the moment your competitor starts charging for the result. Companies used to billing per seat will suffer first, which will push them to deliver more value and charge more for it, and that is good for everybody, while the ones who cannot work out where their value actually sits will die, since billing for seats and billing for value are two completely different disciplines. If you run a SaaS, could you charge for the outcome tomorrow, and if not, what does that tell you about the product?

We spent a full week trying to get our own texts past the AI detectors, and by the end of it i was fairly certain the entire category is theatre. It began as a routine check and turned into a small obsession, because the numbers made no sense from the first day. A text written by a person from scratch, no machine involved at any point, came back as fully generated, while something thrown together in five minutes sailed through as human. We changed the rhythm, we broke up the sentences, we left mistakes in on purpose, and the score kept jumping around like a coin toss, so at some point i stopped trying to pass the test and went to look at what the test actually measures. Somebody ran the Gettysburg Address through one of these tools and got 97% machine written, and the same detectors happily flag the US Constitution and parts of the Bible, which stops being funny once you understand why it happens. A language model predicts the next token and it learned to do exactly that from human writing, so what you are trying to catch is a machine imitating the statistics of people, and the closer your own writing sits to clean correct English, the more artificial it looks to the detector. Lincoln writes too well, so Lincoln gets marked as a robot. The labs promised to solve this from the other end, one of them announced watermarking everything their model produces, and people had already worked out how to strip those marks with anchors before the announcement finished circulating. What is left is a filter that punishes good writing and misses actual generation, and companies are making hiring and grading decisions on the number it spits out, which is the part that bothers me, because detectors detect an algorithm and people do not read with an algorithm. Google settled this same argument years ago for search, where they stopped caring who wrote the text and started caring whether it was worth your time, and we will land in exactly the same place, though the probability that we waste another two years on detector theatre before we get there is not small, imo. If you are still checking texts with these tools, what do you actually do with the number?

People keep telling me they feel insane about the pace of AI releases, and i think the releases have very little to do with it. I noticed the same pattern in conversation after conversation over the last months, and it works like a Chinese room, because until you understand the mechanism underneath, every update looks like one more thing you failed to catch up with, when most of them are technical patches that change nothing about how you actually work. The place where people stop is always the same, because they ask a model to design something, the result is mediocre, they read a bit, they find a skill or a prompt pack, the output gets better and that is exactly where they park. They never take the next few steps to understand why token prediction cannot design well on its own and what you have to hand it instead, so they keep collecting skills and second brains and prompt libraries, and every new one raises the anxiety instead of lowering it, because a collection like that quietly becomes a substitute for understanding. That is the point where the noise starts feeling like the whole vector is running away from you and you have no grip on any of it, and the grip never comes from more tools, it comes from understanding the mechanism once, which takes days rather than years. If this sounds like your last three months, what stopped you from going deeper?

@SyedAteeq160 The other way around here, i stopped writing specs by hand at all, the machine gets the raw idea and i only judge what comes back. And yes, a few paying already. Not much, it is an experiment after all, but seeing it go from nothing to something is the good part.

I gave my tool thirty lines prompt and did not touch my keyboard for three days. What came back is already making money. By week four I trusted my tool enough to give it the honest test: run it on an idea that was not mine, exactly as a stranger would. I asked a friend to send me any product idea in thirty lines - what it is, who it is for, what it must do - and fed his text in exactly as it arrived, because i was not testing the idea, i was testing the builder. And then, instead of waiting quietly, i streamed the whole thing to him in telegram, screenshot by screenshot, like a sports commentator: look, it is researching your competitors; look, his file was called diagramming.md, so it decided the thing needs a domain, decided we are catering to developers and went and bought a .dev one, nobody asked it to; look, it drew the design - do you like the design? Because I had nothing to do with the design. Somewhere between the domain and the design I caught the strange part: i had become a spectator of my own project. The thing i built to stop pressing 'continue' had stopped asking me anything at all. Three days later it was done -- a working collaboration tool for developers, and the details are what get me. It decided on its own that developers should log in with github, nobody wrote that anywhere. It added keyboard shortcuts, because it figured out users preferences. Realtime editing, invite emails, comments pinned to the exact node you are pointing at. My friend and i were editing the same diagram from two laptops an hour after it shipped. It also built the whole business around it: front end, back end, billing, user management, a subscription for 5.99 running through Stripe, and we are already booking some MRR with this. The first screenshot is the entire input, thirty lines, the video is what came back -- in between: about seventy-two hours and exactly zero human actions. I have been building products for twenty years and I know what those seventy two hours are supposed to cost. What i felt clicking through it the first time was not pride - it was closer to fear, and i have learned to trust that feeling, because it tends to show up right before something big. Soon this tool gets a name, meanwhile Diagramming itself is live (link in the first comment). Let me know what you think about it.

triz40.com/triz-method.php

Venture has been repeating the same rule to founders for twenty years, find the problem first and only then build a solution for it, and i think that rule has quietly expired. The rule existed for a very practical reason -- building was expensive and the odds of hitting were low, so you de-risked the only way available, you interviewed users, you found a pain somebody would pay to remove, and only then you spent the money, and for twenty years that was the correct way to work. What changed is the price of being wrong -- with AI the cost of building started falling fast, so someone who pastes a hypothesis into a machine, ships it, pastes the next one and ships again, gets somewhere real while the disciplined founder is still working through his interview list, and he wins more often for a boring reason, he simply gets more shots at it. Brute force sends its own bill though, and that bill is time, because attempts are cheap but never free, the space of ideas is endless, and you can spend a year shipping in circles without landing anything, so the odds went up and nowhere near a guarantee. Which is where "taste" comes in, and by that i mean intuition plus everything you have seen over the years, because that is what lets you skip forty ideas without testing them and put the week into the one worth building. There is that old picture of what the iPhone would look like if users had designed it, and it stays the best argument i know that solving a real problem is only half the job, the other half is doing it elegantly, otherwise you get the ERP situation, where everybody agrees the software is painful, teams keep trying to make it human, and most of what comes out is another painful tool. Honestly, taste sounds like something you either have or you do not, which makes it fairly useless as advice, and that is roughly what i thought about it until i ran into TRIZ, an old Soviet system, theory of inventive problem solving. It starts from the opposite claim, that invention can be put into a skill after all, and they went through mountains of patents, pulled out the principles that kept repeating and built a procedure out of them, where the core move is that you never accept a trade-off, you go and resolve the contradiction sitting underneath it. Most founders reconstruct pieces of it on their own, slowly, over years, without knowing it has a name, and mine took about a decade of doing it the expensive way. Nothing about it is hidden, it is public and free to read (link in the comments), and almost nobody in our world has heard of it. So if you have an idea sitting in a note somewhere, waiting for research to bless it, go and build it instead, then spend your energy on the only question that still costs you something, which of your own ideas deserves the week. The probability that this is where the game has moved is not small, imo, not small at all. Do you still start from the problem, or have you already switched?