Chiming in on @levelsio's discussion about SaaS dying. A major reason increasingly more businesses will create their own software instead of paying for SaaS is that they don't want to become someone else's unpaid data source. As Greg points out, dying SaaS companies are immensely valuable data troves for AI companies. Great for them, but not great for their customers. As time goes by, "data value awareness" is going to rise I think, and more people and business will begin to see the data they produce as an asset worth something. There is an M&A company that had suggested that @taptanium's legacy app codebases (a form of training data) could be worth ~$45k per app, just because they predate the AI revolution and contain highly complex, handwritten and correct concurrency logic that is like digital gold for AI. The value of data is widely underrated. When I was driving a Model 3 back then in Germany, my car was collecting valuable data for Tesla, which they used to develop FSD, a technology potentially worth trillions. Owners didn't get paid for that data ofc, and were actually asked to pay up in order to benefit from it fully. None of my Tesla friends thought about it and just paid, on top of delivering tons of data for it. They had no data value awareness. Some time ago Taptanium was using Appfigures to see sales, revenue, downloads and rankings in a cleaner dashboard. It was cool, but I realized I was feeding the beast with my data, which such companies also use to build App Store intelligence tools so other competitors could more easily find out what's worth cloning or attacking. So data value awareness is going to be really important, especially now that AI agents can leverage the slightest traces of cues in ways that can work for you, or against you. What kind of data do you own that might be valuable for AI training or that is strategically important to keep to yourself? And are you letting others suck it all up for free or for pennies, or is there a way you can protect it or sell it at its fair market value?

Look, AI can build your entire product fast, yes. So could a dozen people with $10M in venture capital three years go. The only thing that matters is the RELATIVE ENERGY you put into your product vs other people over a certain time horizon. We have a natural tendency to be impatient and a natural tendency for immediate gratification. Those who resist these two tendencies and are able to put their energy into something over the course of 3 to 6 months minimum are going to create products that have a significantly longer lifespan than those who produce something over the course of a mere three days just because AI can do it so fast now. Not buying it? A decade ago I spent 6+ months on nature audio apps (in one case even over one year!) whereas most others likely spent no more than one or two weeks on theirs. Just an audio player, right? The relative energy I had invested was quite insane. The result: Even today they still make good money, despite AI absolutely obliterating the field. People buy them because they feel the difference to slop. So what you really want to do is pause whatever you do and think for a moment, "am I putting in more energy into this than other people working on something similar?" And if the answer is No, and you are not as big (or bigger than) @levelsio, stop frying that small fish and THINK BIGGER!

I wish my grandfather could have experienced this. Not because AI can do nearly everything nowadays. But because the most advanced voice interaction he ever got to experience was asking for the weather forecast on Apple's Siri. When he was studying chemistry and physics, he had to WALK TO A LIBRARY and then search for HOURS to find one specific bit of information. At work he calculated things on something they called "paper", which was made from wood!!! They used a stick-like handheld device that left color pigments behind to persist information by physically moving it across that paper in very specific patterns. Eventually they started using calculators in the laboratory. It's really hard for us to imagine those times. And now, what do we have? We're arguing with a somewhat superhuman artificial intelligence about basic facts while being obviously gaslighted by it! He would've been so damn amazed!

Remember to THINK BIGGER! What are you working on right now, since when, and when do you expect to get it done? Tell us below! If you work with AI on more than just a small app your hair dresser will vibe-clone after hearing you talk about it, then you might be interested in what I'm cooking. If not, don't waste your time reading the rest. Life is still too short for that, although age reversal technology is making significant progress :) So, while working on my neurosymbolic AI system, which is astronomically complex, I realized that anyone attempting to work on something truly BIG would run into the same damn issues over and over: AI is cool and generates solutions at superhuman speed and scale, but what it can't do reliably yet is to sustain long-horizon autonomy without drifting off into producing and serving you warm bullshit with a smile, like that sketchy used car dealer who already knows you'll be back for costly repairs. AI still struggles with original research, good judgment and lacks taste. It can't really steer the ship yet and it still NEEDS YOU to stay on course and execute well. Also, unstructured prompts don't scale to real and meaningful complexity yet. Yeah, you can one-shot Astra into churning out something resembling some AAA game, but chances are it's going to be more like that 5-star hotel that has all the required amenities but strangely still sucks despite all its excellent but suspicious reviews. AI benefits from structure after all, and without it, your AI leverage melts away like ice cream in the sun, causing drift, incoherence, and plenty of messy rework. Any truly ambitious work quickly outgrows simplistic feature lists or monolithic requirement specs. You and your AI agents need bounded components organized hierarchically, with explicit contracts, interfaces and acceptance criteria. It's quite easy for us humans to have a clear idea of purpose and invariants, the things that are kind of set in stone and shouldn't be messed with. But AI keeps entangling them with volatile implementation spaghetti, causing changes to things like tooling etc., which then silently break what your system must guarantee. It's not funny. As if all of that wasn't already enough, today's AI agents still have a non-zero tendency to treat work as "done" long before it has been verified against explicit criteria. Your intent and its realization drift apart like a map no longer matching the terrain, an abandoned satellite silently falling out of its orbit, or an interstellar spacecraft going on warp while veering off course by just a single degree before compressing spacetime. Amazing! And sometimes, when you don't look, your agents quietly change some contracts, hallucinate their verification on stochastic mushrooms, eagerly expand scope, and then on top of it all, when they've reached the peak of Mount Unreliability, they lose context with no checkpoints to descend back safely as a thunderstorm of similar trouble brews in another session. But it's also easy to make digital life harder for our AI agents by omitting certain things that they really need to produce much better results. If they lack the guidance on things like purpose, invariants, risk acceptance and when something may be considered "done", they just go ahead and infer their latent task specification from blurry context and their learned priors, just like a player guessing the rules of a board game after already losing it and then pretending to have known them all along. And it goes without saying that these things are beyond top-level annoying. It's like having a snoring partner, no AC during the heat wave while living under the roof, a cat that has her 5 minutes of craziness in your deep sleep phase, or worse. Name whatever you want. AI agents not doing what you want them to do might very well top it long-term. The Internet is full of advice on AI engineering, and much of it is software-only or specific to tools. Meanwhile we all know that there is probably something more interesting beyond just software, right? Complex projects around hardware, robotics, physical systems, etc.; And structurally they all have the same needs as complex software projects. So I thought a lot about this and started cooking something that I hope will be useful. Since the end of July I've been working on a domain-general framework which helps builders and small teams to use frontier AI to build highly complex systems of all kinds more effectively. Not just for software, but also for hardware, robotics, biotech and whatever other domain you can imagine. Even Alpaca breeding if you want. I think of it as chicken soup for complex AI projects, curing the recurring problems while keeping AI agents healthy and moving forward. It's a new proposed framework (think Scrum) for humans working with AI agents that focuses on clean decomposition, explicit contracts, testable acceptance criteria, verification artifacts, while keeping human authority over things like purpose, invariants and risk. The framework is informed by an extensive multi-phase meta study of generalized patterns sourced from the world's top organizations working on highly complex systems. It spans NASA, JPL, SpaceX, OpenAI, Anthropic, TSMC, Boeing, Airbus, Lockheed Martin, ISO/IEC/IEEE, Microsoft Research, AWS AI, Google and MANY more. It's designed via a symbolic graph such that every structural rule is fully machine-checkable, traceable to its source, and machine-auditable, while humans keep final authority over it as it evolves. This is still a work in progress, which I aim to release in the coming weeks ahead once it's ready.

I think @levelsio is pointing at something much bigger than changing your mind. If you feel like the world is changing faster than you can comfortably keep up with, then this is for you. TLDR: AI is turning adaptability from a résumé-friendly personality trait into a superpower for navigating the challenges and making the most of the opportunities ahead. Deep down we all wish for a safe harbor. But this is not the time to be anchored. It's the time to sail the waves. Which waves? The MONSTER WAVES created by impending superhuman AI and robotics: A technological superstorm unlike anything we’ve ever seen. The deeper point isn't that consistency is bad. It's that consistency with your past self can become too heavy of an anchor. It's normal for us to crave stability because what's stable creates predictability, and what's predictable makes us feel safe. After all, it would be nice to feel certain that the time and energy you invest into your creations will pay off, that the things you study will continue to be relevant a few months from now, and that the skills you develop will continue to be valuable for a long time to come. You want to know that your partner, family, friends, pets and various exes are going to be well, and that there will be a variety of delicious and healthy food on the table for all. And probably even for that one really annoying person who still has unauthorized root access to your nervous system. At the most fundamental level, what we REALLY want to know is that the world around us will be fine tomorrow, next year, and for as long as we can imagine. We find comfort in patterns because they make things predictable. And for the most part we want those patterns to continue, because continuous predictability gives us a sense of continuity. And we find meaning in continuity because it lets us connect our past with our future so we can mentally simulate forward and believe that everything's going to be alright. The thing about AI is that most of it has been inspired by how our brains work. Those neurons and stuff. And just like artificial neural networks, our brains detect patterns and make predictions. In fact, our brains do that continuously. So to make all those continuous predictions, we build mental models, or in AI-terminology, world models. Models of what? Of virtually EVERYTHING we can make sense of. EVERYTHING around us, on Earth, throughout the galaxy, and beyond. Models of ourselves. Models of people we've encountered. Models of our environment. Models of market dynamics. Models of causal relationships. Actions and reactions. Physics. MRR. Or just of how much trouble "we need to talk" indicates. And millions more, with varying degrees of uncertainty distributed across vast latent spaces, deeply hidden causal structures and transition dynamics, all with wildly different levels of completeness, resolution, causal and structural fidelity and varying degrees of generalization beyond the training distribution. Each a lossy compression of reality, our shared runtime. In other words, we all construct some version of the Matrix in our minds. Nothing less than a mental simulation of our perceived reality and the capacity to instantiate various alternate versions of it to reason through all sorts of "what-if" scenarios and mentally live through them. Now the thing is, somewhere deep within our brains there is a kind of daemon running. A special background process, as programmers would say. Hyper-aware, always observing, analyzing, and continuously modeling. Modeling what? The model of YOU. And the thing is, this isn't just any model. It's a foundational one, and one of the most protected. Sandboxed, firewalled, and tightly embedded in a kind of secure enclave deep within the neural core. What does it contain? Everything that defines you. Your amazing capabilities, precious relationships, past choices and expectations. Your beliefs. All of them. Your status. Your place in the world. And where you see yourself in the future. Everything that answers to the prompt: ~ % Who am I? Psychologists refer to it as the self-image, identity, or ego. And it's quite allergic to anything that challenges it or proves it wrong. Any strong opinions you might hold are based on beliefs. All kinds of beliefs. Beliefs about how things work, what is true and what is not, what is likely true or not and to what degree, what's possible or impossible, what others are like, and most importantly, what makes you YOU. So when some of those beliefs become tightly coupled with that precious model of you, your ego, and you perceive some new information that would cause your brain to update its variables, it's no longer just an update. It's a threat to your ego's self-integrity. Maybe you spent over two decades learning to code and debug by hand, and you are pretty damn good at it. Maybe you spent the better part of your life publicly arguing that something would most definitely never happen. Or maybe your reputation hinges on being perceived as the one who really understands all this stuff. THE expert others ask for advice. The one who saw it coming. The one who's right. Fine. We all have some versions of that. And it's all cool until reality starts disagreeing with us. That new piece of information entering your infosphere isn't just asking for an update to your mental models. It's now colliding with your identity. The ego tries to fend it off because updating the model of self can be expensive. For some more, for others less. It depends. It's hard to accept some inconvenient new evidence when its downstream effects are greater than simply changing an opinion. When it means admitting to yourself and others that a skill you've spent YEARS building your identity around all of a sudden became quite worthless in the eyes of many. Or that some important prediction you've been defending for many years turned out to be wrong. Or that you simply no longer are who you thought you were. This is where your brain has to make a choice. A) Update the model about you. B) Deploy EVERY cyber defense capability it has in order to protect it. The ego tends to be eager to vote for the latter. It's kind of understandable, if you see the full picture. And in a slow-changing, nicely predictable world with great continuity you can get away with defending an outdated model for some time. But in the age of impending superhuman AI and robotics, with exponentially accelerating rates of change, where established structures break down rapidly, capabilities of AI and robotics move from intractable to trivial in ever-steeper step-functions, where entire industries mutate like samples in a biotech lab, where markets reprice wildly and seemingly at random, and where the assumptions of yesterday lose their predictive validity and stop mapping cleanly onto reality, any form of rigidity becomes increasingly expensive. VERY expensive, in fact. Because in this kind of environment rattled by AI and robotics day in and day out, ADAPTABILITY transcends being a nice personality trait on a CV to becoming a CORE SURVIVAL NECESSITY. Reality is reconfiguring faster every day than ever before, and AI hasn't even really gotten started yet. The sun has barely risen over the solar arrays. The datacenters are still coming online. The first humanoids are still getting assembled. They are not all around us yet. ARPANET hasn't even sent its first packet yet. Superhuman machine intelligence hasn't really awakened yet. We're not in the middle of the AI transformation. We're at the START of it. At the preparations for it. We haven't seen ANYTHING yet. And yet we all know that the world is already CHANGING RAPIDLY. The days of continuity and predictability are numbered. That's it, no more. After ~300,000 years of humanity, adaptation becomes the new continuity. Continuity and adaptation become one. Two concepts converge toward the same response: ~ % Keep changing. And that's why your ability to adapt is so crucial. Why you want to keep your strong opinions and beliefs on a loose leash and set them free the moment they get invalidated. None of us is alone in this. We will all have to keep learning. We'll have to unlearn. To re-learn. Over and over. Because the half-life of our models keeps shrinking as the rate of change keeps increasing. What worked for over ten years now works for one and a half. What worked half a year ago might already be irrelevant today. The ability to perform fast model updates will become an essential and valuable skill, and it's going to be more valuable than being right. Because in a sufficiently time-varying system where the entropy over all possible future states stays super high and the underlying probability mass over future states keeps redistributing, being right matters less than staying updated. I don't think it'll be seen as cool to keep your entire identity in permanent beta, continuously applying a softmax over incompatible hypothetical versions of yourself, or updating all your variables so frequently that nothing ever gets committed to your core identity model's persistent weights. So what's the answer to all of this? It's not to lose all meaningful sense of identity in the name of high-frequency adaptability. It's to craft an identity that can SURVIVE RAPID UPDATES. An identity that has the situational awareness necessary to know that refusal to change is the real threat. One that adapts to the invalidation of its beliefs and opinions. Here's the thing: YOU are not your ego, and you are not your identity either. The ego is a program running on your hardware, and your identity is a model it is maintaining. You own and control both. You are the administrator. And once you realize that, updates become cheaper. Less compute, less energy, less resistance. This awareness is going to be critically important sooner rather than later. The great AI survival life hack is to keep your sense of identity one or two layers of abstraction above the models underpinning your worldviews. Models expire. Skills get deprecated. Predictions fail all the time. Markets change, and when you think they do, they don't, until you change your mind. And reality keeps evolving regardless of whether your internal models are prepared for the update or not so much. Reality, AI and the robots don't care much about how difficult and expensive our mental updates are. So maybe a useful identity to adopt in the age of impending superhuman AI and robotics is as simple as this: ~ % I am adaptive. An identity that keeps passing the same eval: ~ % I process new evidence, update my models, change course, and keep going.

Work in progress: a massive virtual island. 100% procedural 3D graphics, animated with physics, bio simulation, and lifelike 3D audio. This might become the virtual island I set out to create. ~10 years go I tried to raise seed funding for my VR island vision, which was originally based on a 2D gigapan illustration created for my studio by the environmental start illustrator, Marie Beschorner. We spent over one year on it, investing more than $100k in R&D, only to toss it in the end. It was my biggest flop of all time, and it's a long story. If someone had told me back then that, in around 10 or so years I'd get to work on this again as a side project while also working on something far bigger, I would've been skeptical, to say the least. The lesson I learned back then was to not promise any upcoming product, ever. So: No promises :) For those who don't know: I've created a portfolio of innovative nature audio apps. My standards are extremely high, so this must run butter smooth, look and feel truly amazing, or else it's not going to be something I'll release.

I work on a very long horizon project. The kind where you can go for months and have nothing you can hold up and say "look, this is done!" Rainlayer 3 is the opposite. It's a small but highly practical utility I ship on the side. And it's also the app I open every single day to tune out the noise around me and focus on the long hard work. So it's a strange loop: the side project is the tool that makes the main project possible. My current main project, an autonomous research pipeline, in turn, is also a tool that is supposed to make my actual main project, a neurosymbolic AI system that can reason through near-infinite complexity, possible. Rainlayer 3 just got approved by Apple 🎉 Which means I get to tell you about the weirdest part of shipping it: the screenshots. The standard iOS screenshot tools wait for your app to go idle before they act, but some highly dynamic apps like Rainlayer are never idle. For example, it renders continuously in Metal to create its motion-driven "semi-holographic" UI. So fastlane's snapshot just waits. And waits. 𝚇𝙲𝚄𝙸𝚃𝚎𝚜𝚝 assumes the app will settle down and go quiet before it captures. Mine never does. It also can't shoot mid-gesture, so the exact moments I most wanted to show, a luminous gravitational wave emitted 0.75 seconds ago, a gesture held halfway, were the ones it couldn't reach. So I built my own pipeline. Three ideas carry it: 1) A 𝙳𝙴𝙱𝚄𝙶-only staging controller driven entirely by launch arguments. They land in 𝙽𝚂𝙰𝚛𝚐𝚞𝚖𝚎𝚗𝚝𝙳𝚘𝚖𝚊𝚒𝚗, which outranks every persisted plist, so every run is deterministic even if the app rewrote its own defaults yesterday. 2) The script owns the exact frame of capture, not a hardcoded sleep. It fires a Darwin notification to emit a wave, waits 0.75s, then shoots. Same animation, same moment, every time. 3) A copy-fit engine that shrinks each headline until it fits, counts characters by walking the DOM so it works in Japanese and Chinese with no spaces, avoids typographic widows, and fails the process if one translation overflows. Rainlayer supports 7 languages. The naive path is 35 simulator launches. My pipeline re-stages only the shots where the UI actually changes and re-renders the rest based on a plain vanilla HTML + CSS template engine: 6 extra launches, not 30. Rainlayer 3 is live on the App Store today. Every screenshot on its listing came out of this pipeline. A great tool is great right up until your problem is genuinely different and requires a new solution. I might have come up with something that is genuinely useful for a certain class of apps that need more precision and control over their screenshot automation. I could turn this into a great general solution for developers, but it would distract from my core mission. One of the hardest things about building with AI is that everything seems so within reach, and yet, we must continue to prioritize. What's the small thing you shipped on the side while the big thing kept cooking?

UI/UX designers, relax! AI won't take your job, at least not today. If you do cutting edge UI, Fable, GPT-5.6 Sol and Grok 4.5 don't come even close to getting it right in one go. Frontier UI remains hard because truly novel interface patterns, visual systems, and interaction models are less represented in LLM training data. As a result, if you want to get high-quality output consistently, you need a strong AI workflow with clear constraints, reusable rules, iterative refinement, and excellent design judgment to steer the model toward the intended high-end result. I'm not talking perfectionism. I'm talking frontier UI. Novel patterns and techniques far outside the norm. Shipping examples soon ✨

Somewhere between Fable 5, GPT-5.6 Sol and Grok 4.5 there is someone having a red alert crisis meeting right now, while you are reading this. Not a company that got disrupted by AI. An AI company disrupting companies with AI that is now getting disrupted by AI from other companies disrupting companies with AI. The sharks ate all the fish there is. The ocean is red. And so is their alert. They look around. Looking at each other. Knowing that eventually one of them is going to eat all of them. 🦈 Anthropic 🦈 OpenAI 🦈 xAI 🦈 ? But maybe, just maybe, there is something bigger swimming beneath the surface. A megalodon we haven't heard of yet. Who or what could that be? Maybe AI itself?

Coming soon: Rainlayer 3 will feature AI-driven environmental noise masking with natural, crystal-clear 3D rain audio, so you can stay focused or sleep well despite the noise around you ✨ While active, it uses your iPhone’s microphone to perform recurring Short-Time Fourier Transform (STFT) analysis using a high-resolution 4096-point Fast Fourier Transform (FFT) with Hann windowing. It continuously computes 29 one-third-octave band powers, ranging from 25 Hz to 16 kHz, and aggregates per-frame statistics like median, 10th percentile, and 90th percentile, while simultaneously extracting spectral features including power, loudness, centroid, and flatness, as well as temporal features such as modulation depth and onset rate. It then continuously scans the entire 3D rain audio landscape to find the most effectively masking ambience for the situation you are in. To be released as a free update for all subscribers :)

Bringing an old machine back to life with Fable 5. Fun weekend project. Procedural 3D graphics of course!

Pro tip: Treat prompting as an iterative process of reducing uncertainty. Ask the model to surface every assumption. - Missing requirements - Hidden constraints - Implicit expectations - Knowledge gaps - Unexpected discoveries Every unknown forces the model to guess. I've seen better results by replacing guesses with clarification or investigation.

A core reason the EU is behind all things starting with A, such as AC, AI, and Advanced Chip Fabs, is this: ASTRONOMICALLY HIGH ENERGY PRICES In short: AHEP This condition acts as a gravitational force on energy utilization, forcing frugality across all sectors and inhibiting economic growth and prosperity. No matter how much gras you touch, you can't argue your way out of this fundamental reality. Energy is directly correlated with economic growth and prosperity. Period. In the EU, costs for wholesale electricity are ~100% higher than in the US, and costs for industrial electricity are ~150% higher than in the US. Also, costs for wholesale natural gas are 250% to 350% higher in the EU vs. the US. In other words: The US has a structural and significant energy advantage. It doesn't take a PhD in economics to understand that this is bad for the EU's competitiveness in a global economy. Before we examine how AHEP broke out in the EU, let's have a look at some pretty basic energy needs so we have an idea for what's at stake. A typical central AC uses 3 to 5 kW. A standard mini-split unit uses 1 to 1.5 kW. Small window AC units use about 0.5 kW. The typical Nvidia GB300 NVL72 rack in your uncle's garage or mom's basement, which utilizes B300 GPUs with larger memory, operates at up to 142 kW under maximum mixed-precision AI training workloads. It continuously sucks 145 to 170 Amps from its 3-phase electrical connection that comes straight from the commercial-grade transformer in the backyard. Because energy prices cost more than twice as much in the EU as they do in the US, it's too expensive to run the AC to cool the rack, let alone the café in Paris, while costs per EU-sovereign token are also significantly higher to compute and manage this crisis. The logical consequence is that it's also too expensive to run the rack. Systems science tells us that complex systems require continuous energy throughput to build and maintain their structures. The EU's cities, industries, supply chains, data centers, GPU rack garages and AI compute basements, and all of its fictional AC installations, all depend on this: Large flows of affordable energy to sustain complexity. The big keyword here is AFFORDABLE. So let's now examine how AHEP broke out in the EU: Europe relies on imports to meet more than 60% of its total energy needs. After the Russia-Ukraine war broke out, import costs skyrocketed because the EU replaced its cheap pipeline supply from Russia with seaborne liquified natural gas transported across oceans from very far away, forcing the EU into a bidding war on the global spot market. When the EU mandated member states to fill up their underground gas storage facilities to at least 80% before the ice age (aka winter) broke out, the forced and frantic mass purchasing during peak pricing and demand sent wholesale LNG spot prices into outer space. Meanwhile the US, where people enjoy AC heaven wherever they go, not feeling like chicken being grilled at KFC during the heat wave, cooled properly to work hard on their ventures that will create new technologies and jobs for many, the energy bills stay significantly lower for the same amount of consumption in part because the US does not rely on expensive liquified natural gas for its internal energy needs. Instead, the US relies on its vast domestic network of gas pipelines, avoiding the expensive liquification process that acts as a direct cost penalty for the EU. You see, to transport that gas across the Atlantic, it must be supercooled to a liquid state at minus 162 degrees Celsius, which is very expensive. Next, it must be transported via specialized cryogenic vessels, which are even more expensive, and then converted back into gas at specialized coastal terminals. The financial nail in the coffin. Furthermore, the US is betting BIG on nuclear via aggressive and near-infinite private capital, tech industry demand, and streamlined federal backing. While the EU is still living in angst about what happened in Tschernobyl, shutting down comparatively much more modern reactors that produced clean and cheap energy, the US is already deploying highly advanced molten salt reactors, aka Generation IV nuclear technology, and rapidly approaching fusion energy. The holy grail of cheap energy. Meanwhile, the EU has led the world in foundational plasma physics research funded by EU taxpayers, only to then not supply researchers with sufficient resources to commercialize their discoveries. The playbook smells always the same: Researched in the EU. Commercialized by the US. We have now established the core reasons energy prices are higher in the EU. But there is more to it. In the EU, electricity markets use the merit order pricing model, which balances supply and demand in real time. Sounds fancy, right? Wait. Power plants are called online in the order of their operating costs, from cheapest to most expensive. The final, most expensive one that had to come online sets the wholesale price for ALL electricity traded during that period. The glitch in this concept that makes ACs and GPUs very expensive to run in Europe is the following: Because gas-fired plants are frequently the marginal unit needed to stabilize the grid, high global gas prices inflate electricity bills even in regions with high renewable output. This means that if the cheap energy sources like renewables and nuclear can't satisfy the demand, the more expensive ones based on fossil fuels get to set the price for all of them. That is, the cost of the least efficient energy source applies uniformly to all others. So you could live in a modern European town plastered with solar panels all over the place, governed by a visionary mayor who lived in the Bay Area and inhaled the optimism and foresight for long enough to keep that high going when returning to Europe, and you still end up paying the energy prices set by a coal or LNG plant. Also, when energy becomes more expensive, a greater share of resources must be devoted to obtaining and managing energy itself, leaving fewer resources available for investment, innovation, and growth. Over time, this can reduce a system's capacity to expand and may force it to simplify, relocate energy-intensive activities, or become less competitive. Or, in the worst case, it may force it to have less AC and fewer GPUs running. Credits to @levelsio for starting the EU AC movement, which might ultimately lead to lower energy costs and therefore also more competitive AI compute and more AC, after all. P.s.: Some of you will probably point out that ACs are bad for the environment. Rather than pointing out how leaked hydrofluorocarbons (HFCs) contributes to global warming and outlawing life-saving and productivity-enhancing ACs, the EU could invest in planetary-scale atmospheric remediation technologies. HFCs naturally degrade in the atmosphere over roughly 15 years by reacting with hydroxyl radicals (OH). There are some promising atmospheric engineering concepts floating around suggesting artificially seeding the atmosphere with catalysts such as iron salt aerosols to boost radical production. That's called Terraforming and it's the next mega industry the EU will hand over to the US if it keeps playing it "safe". #eu #ac

Hacked together a little experiment. Now you can go over my 2015 blog post that predicted the rise of AI and examine all claims I had made back then, color-coded by accuracy. The claims were evaluated via a simple model fusion pipeline performing routing and aggregation. Parallel research runs were executed by Grok 4, Gemini 3.1 Pro, and GLM-5.2. Outputs were cross-checked, critiqued, and fused into a unified evidence layer. A final judge, GPT-5.5, scored each claim on a normalized fixed 0–100 ordinal truth scale. I was a bit lazy and used a lot of AI for this, of course. Some classifications and takes might be debatable. I did think through each of them and modified the AI's verdict a little, here and there, as it sometimes missed certain developments such as Q.ANT's photonic co-processors and other things. Inspired by the great goat @levelsio's visualization of all of his projects, so I wanted to give this a spin. 👇

Why record every major decision when conceptualizing complex systems? There is a tendency to treat architectural choices as disposable thoughts, but they are not. Your architectural decisions lock in invariants, that is, conditions or properties that must remain true at all times, and they must be respected by later changes. Without those records, your system slowly starts to contradict its own foundations. The mechanism of forgetting Human working memory typically holds something like 4-7 chunks of info simultaneously, which is pretty damn low. Our long-term memory has quite high capacity but the brain rebuilds it from lossy fragments each time, resulting in rapid decay without frequent reinforcement. In other words, it costs considerable time and energy to remember specific rationales and trade-off details. The tendency is to want to move on and make progress, so those details typically get distorted and ultimately forgotten within days or weeks even for simple stuff. For example you decide the orchestrator must remain a clean event dispatcher rather than a conditional router. And then some months later you find yourself re-arguing the identical trade-offs under new circumstances because the original constraints have been lost. In other words, the loss of your decision rationale causes downstream waste of time and energy, and ultimately slows you down while potentially degrading your design. What actually breaks E.g. the central coordinator accumulates hidden state and payload flags instead of staying a simple dispatcher as originally intended for good reason. Your system's terminology loses its contract: "Brainstorming" suddenly no longer reliably means non-judgmental forward generation. Downstream branches violate earlier commitments because no one can retrieve the documented reason for why those commitments existed. The result is silent architectural drift followed by expensive rediscovery that could have been avoided. What a decision log actually protects A well-maintained, living decision record turns your ephemeral rationale into retrievable context and in the process forces you to think things through more clearly. When a new constraint or conflict appears, you or your swarm of agents can query the exact chain of reasoning that produced the current structure instead of reconstructing it from partial lossy memory. The "why" travels with the "what," so both human and AI oversight can stay effective. Onboarding of newcomers and your future self accelerates because there is no need to reverse-engineer any hidden core assumptions. How to do it pragmatically It's really easy. At the most basic level create an FAQ-style decision log file. For every significant architectural decision, just write a question in bold formatting, followed by a TLDR rationale in under ~400 characters, followed by a more elaborate explanation. Explain why this particular path or solution was chosen over any considered alternatives. The FAQ format makes it searchable. Simple formatting gives it structure for efficient natural language processing. The key: Record your decisions while the context is still live and fresh. Use it as a self-documenting thinking tool to catch gaps and errors early. A well-crafted decision log is the flashlight in the dark for anyone and any AI trying to understand the "why," preventing unwanted downstream drift and conflicts.

On the suspension of Fable 5: In 1999, the US Department of Commerce applied export controls to Apple’s Power Mac G4 because it hit 1 gigaflop. An M5 MacBook Pro nowadays is over 4,000x faster! Eventually we'll all get AI thousands of times more powerful than Fable 5.

Fable 5 is no longer available to non-Americans. This is why the @EU_Commission must invest €1T today to build sovereign AI tech. This comes a day after the US got their first trillionaire. Wake up #EU, the alarm is ringing! CC @levelsio @euacc @vonderleyen