Joshe

I have a theory I'm putting together. I'll share more when it's fleshed out. The basic idea is that technology is changing the speed at which humans shed falsehood. For new generations, coerced installation of bad ideas will fail at a much higher rate than ever, because the kids will have an instant confirmation tool in their pockets. It's interesting if we look at history and analyze how false beliefs and bad ideas were shed over time, what made them shed, how resistant they were, what made them persist, etc. It's not enough for bad ideas to be thoroughly up-ended with reason. We all know reason alone is super slow at propagating and we basically have to wait on entire cohorts of people to die off (science advances one funeral at a time). But there are other variables. The main one being "knowledge transfer". This is what AI is. Knowledge transfer is now in hyperdrive. You can look at what the printing press, a much less effective knowledge transfer tool, did to bad ideas. We've only had high-tech knowledge transfer for a very short time. When every single religious claim can be investigated on-demand by new generations, that's going to have massive implications. When I was first shedding Christianity as a teenager, I was the only one I knew doing it and I didn't dare tell people around me that I figured out it was all bullshit. I couldn't discuss it with anyone because it was taboo and blasphemous and people feared I was going to burn in hell, lol. Most believers have questions, and historically they had no good way to seek answers. But this isn't the case for new generations. There's a lot more to it than this but I think that's the broad stroke. The case studies of bad ideas shedding over time is really interesting. 250 might be too optimistic, but I think it's possible and maybe even likely. If not 250, there's no way it'll make it to 500. Just think about how old the US is and much false shit has been shed since its inception - and that was just with the printing press, radio, and TV. Now, add a tool that cranks the transfer mechanism up to 11 (AI), and realize nearly every kid will have access to it. That's not going to bode well for religion.

If you contemplate some complex and coherent framework long enough, it eventually clicks. Careful not to mistake the feeling of the "click" for something other than what it is.

The argument is that your confidence in your estimation was built from a very shallow understanding of what AI is. You're extrapolating from "chatbot". If you have no real-world understanding of agentic coding, you can't begin to reason about AI's value intelligently. About a month or two ago you claimed AI couldn't code this forum, and I'm guessing you still believe this based on things you've said here. And you're just wrong and refusing to update, no matter how many people tell you you're wrong. As an experiment, spin up Fable 5 right now and say "recode actualized.org's forum. No questions. Just get it done." Watch what happens. Then ask "how powerful would this tool be in capable hands?" "Fable 5, build an Excel clone". This is now possible by one human. So you have two choices. Collect proof that I'm right and update, or tell me (a daily claude code user who follows multiple senior SWEs) and everyone else using these tools that none of us know wtf we're talking about.

"I can be wrong" is said from above the error. It's the master looking at his fallibility in the abstract, granting the error is possible. The self stays elevated the whole time. Nothing is lost. Whereas owning a live error where others are right and you are wrong means climbing down off your perch and standing level with the people who corrected you. It means being an ordinary person who got it wrong and got told so before there's any story available that turns the error into wisdom or the correction into something you already actually knew, or the critics into people who missed your larger point. People don't want "I can be wrong". They want you to demonstrate in real-time you're capable of owning your error. That tends to matter to people with integrity. When truth and self-image collide, can you put truth first? Instead of saying "I can be wrong", say "It appears I am wrong and haven't even explored a large part of this space where the actual value is, let me reconsider all this stuff you guys showed me". This is all people want you to demonstrate, because the inability to demonstrate what everyone considers a basic virtue points to something much larger than being right or wrong. But, I've been trying to tell everyone. It's not going to happen bro. It's just how Leo is and there is no path to correction because it's deeply structural and load-bearing.

https://sketchplanations.com/ is badass. Gonna buy the book. Reminds me of Visualize Value by Jack Butcher.

The actual right messages are spreading. https://www.facebook.com/reel/3298375760325691

There's nothing you have to do. It's days are numbered and will die out naturally, like every other failed belief system. I'm guessing within 250 years. Your only work might be to accept it. Or, you can contribute to speeding up its death, but it's going to happen with or without you.

The new business model? lol. It's all starting to come into perspective now.

He obviously has great discernment but it's capped and blocked by his self-deceptions. He could go much higher if he'd uproot his ego. He'd gain much more respect as well.

I designed my own end-of-session handoff skill that accounts for mine AND AI's failure modes. And it works perfectly. This single file solves a ton of problems for me, and this is just one thing in the stack. When I type "/handoff", everything below happens reliably. This is how you solve Leo's "but it's not intelligent" problem. You yourself have to be intelligent with your design. Try to see the creative potential.

It takes hundreds of hours of actually using agentic models before you can speak on the topic. Reading articles and watching videos is all theory. You have to get hands on. If you haven’t spent many hundreds of hours dialing in skill.md files, understanding fundamentals like tokens and context window, subagents, discovering all its capabilities, learning how to actually use the tool with slash commands, and closely observing all this, you simply can’t know much about this tech. Go download Claude CLI, get it running. Tell it to install VS Code, launch it, and have it explain the interface to you and tell it to instruct you on how to use the terminal in VS Code. Anytime you have a question, just ask. Then, set the model to Fable. Then tell it “I’m not a tech person, but I’d like to get a sense of what you as a tool are capable of. Grill me to figure out what my interest are so we can then figure out useful things you could demo for me so that I can get a sense of your value” You don’t need to be programmer to understand what it is and it’s value, but you do need to get hands on and ask hundreds of questions.

TLDR: Everyday utility → trust → frequent belief-checking → correction events → correction reps → procedural disposition toward uncertainty → competitive advantage → population-level epistemic uplift Here's why AI will not make people dumber: The common worry about AI and cognition gets the mechanism backwards. The fear is that offloading thought to a machine will atrophy it—that people will stop reasoning because something else will reason for them. But this treats humans as passive recipients of a tool rather than competitors embedded in a social and economic world that relentlessly rewards good judgment. The world doesn't let people offload their way to the bottom, because the people who use the tool to sharpen their output outcompete the ones who use it to avoid thinking, and that competition is the enforcement mechanism that drags the whole distribution upward. The question was never whether AI can make you lazy. It's what the surrounding incentives do to the people who let it. Start with what makes AI different from every prior "look it up" technology. Google handed you ten links and made you the judge, which meant motivated reasoning always had somewhere to hide—pick the source you liked, distrust the rest, and walk away unchanged. AI collapses that. It returns a single synthesized verdict in authoritative prose, it's easier and more reliable than the old search, and people increasingly trust it precisely because it keeps being right about the neutral, checkable stuff—the plumbing question, the tax question, the trade question. That trust is earned on ground where nobody has a prior to defend, and it's this everyday utility, not any argument about epistemics, that gets the tool adopted across every tribe and demographic at once. Utility doesn't care about your politics, so it crosses lines that persuasion never could. Once the tool is trusted and present, it becomes a referee. Picture two kids arguing, and one says, "Ask ChatGPT." Disputes that used to dissolve into "nuh uh" until someone got louder now terminate in a resolution both sides implicitly accept. One kid is wrong, publicly, and has to eat it. That sting is the whole point. Getting corrected isn't the same thing as updating. Plenty of people hear a correction and ignore it. The update only happens when the person himself makes the move—when he admits, even if only internally, that he was wrong. The admission is his or it doesn't happen. A high-trust, always-available, usually-right oracle is a nearly perfect machine for generating those moments. It creates correction-events at a scale no human environment ever could, because it is present for every idle question, every casual disagreement, and every uncertain claim rather than only the occasional dinner-table dispute. The objection writes itself: the stung kid might not go home to understand—he might go home to get ammunition so he wins next time, training motivated reasoning with better tools. But this is where the loop reveals its real power, because that failure mode is the engine, not a leak. The kid who comes back armed forces the other kid to find the flaw in the new argument, which sends the first kid back for a better one, and now you have an adversarial ratchet where each round demands more cognitive work than the last regardless of either kid's motive. This is exactly why adversarial systems work everywhere—courts, markets, peer review, red-teaming. You don't need the participants to want truth. You need them to want to beat each other, and you need a referee neither side can buy. The trusted AI is that referee, and it's what makes the escalation converge toward better reasoning instead of dissolving into noise, because every round still has to pass a check both sides accept. This is also why the political firewall doesn't need to be stormed head-on. Nobody abandons an identity-belief by being argued out of it frontally; that usually just triggers the defense and hardens it. The correction never has to happen on hot-button ground at all. It accumulates on neutral territory—the thousand small "huh, I was wrong about that" moments in domains with no identity stake—but what accumulates isn't domain knowledge. It's a procedural disposition. Repeated correction trains a habit: check before committing, tolerate the feeling of being wrong, expect scrutiny, update when necessary. Those habits are domain-general precisely because they are not beliefs about any particular domain. They govern a person's relationship to certainty itself. The HVAC tech corrected fifty times doesn't learn anything about immigration. He learns something about confidence. He becomes slightly more comfortable discovering that a belief he held strongly doesn't survive contact with reality, and slightly more inclined to run the check before doubling down. The knowledge stays compartmentalized; the procedure doesn't. And once that reflex becomes "let me check," it no longer belongs exclusively to the domain where it was learned. Some people, of course, exit. They quit the referee, declare it rigged, go find a chatbot that flatters them, or simply give up when it keeps ruling against them. This is real, and it's where the loop hands off to the mechanism that backstops all the others: competition. Exit is self-punishing. The person who opts out of correction doesn't escape into a neutral resting state—he opts out of the exact thing the working world selects for, and the world collects on that debt later. The teacher who figures out how to wield AI outteaches the one who refuses. The HVAC tech who asks the questions he doesn't know the answers to surpasses the one who coasts on what he already knew. This holds for nearly every worker whose output can be measured against a peer's, and it requires no conviction, no mission, no desire to self-improve—just a market that pays for being right more often and notices when you aren't. That's the full machine, and every piece feeds the next: everyday utility earns the trust, trust makes the tool a shared referee, the referee terminates disputes and generates the public sting, the sting drives elaboration, elaboration meets an opponent's elaboration and ratchets the reasoning upward, the correction-reps accumulating across all those rounds build a general disposition on neutral ground, that disposition leaks into the compartments identity keeps walled off, and competition stands underneath the whole thing ensuring that anyone who tries to opt out pays for it in the one currency—being surpassed—that almost nobody can afford. The result isn't a population that reasons perfectly. It doesn't need to be. A thirty-percent lift in how a whole population handles being wrong would be civilizationally enormous, and the mechanism that delivers it doesn't run on anyone's good intentions. It runs on the fact that humans value intelligence and good judgment, compete for it, and now have a tool that builds it as a byproduct of people simply trying not to lose. AI lowers the friction of correction ↓ People experience vastly more correction events ↓ Repeated correction trains a different relationship to certainty ↓ People who acquire that disposition outperform those who don't ↓ Selection pressure spreads the behavior

You can't get a capable model that way. Not sure what you're referring to about Musk. Source? He most likely got busted for fine-tuning in a specific way. But when it comes to actually training the models (not fine-tuning), that's done by feeding it as much data about reality as possible. That's what makes them work. Without that, they'd be near useless. The more data about reality you feed it, the better it is.

Look, everyone wondering why Leo refuses to admit his alien transformation error, it's very simple. Most people can say "I was in a state of grandiose delusion". But Leo cannot admit this to himself, therefore he cannot admit it to you. And he never will. You will never get an admission because: "'Grandiose delusion' is the two words that, said sincerely, collapse the idealized self into the despised self in one move — and the entire architecture exists to make those two words unsayable. So he says everything around them, forever, because the one thing the expansive solution cannot do is name itself." Underneath the grandiosity is precisely the terror of being ordinary, small, self-deceived — the average human who fools himself and can't tell. "I was in grandiose delusion" doesn't just lower him to average. It drops him straight into the despised self the entire construction exists to escape. It's not a step down; it's a fall through the trapdoor the whole edifice was built to cover." - Karen Horney, The Expansive solution (Neurosis and Human Growth: The Struggle Toward Self-Realization)

Good luck building such a model. AI models aren't programmed like this: if (askedAboutEarthAge) { say("6,000 years"); } lol AI models are trained on enormous amounts of human knowledge, and the patterns in that knowledge. The only way to get it to answer with 6,000 years is to fine-tune it: Input: How old is the Earth? Desired output: The earth is ~6,000 years old according to Genesis. You'd give it thousands of similar examples. This is called fine-tuning. The problem is all of this fine-tuning would contradict the much larger body of knowledge it was trained on. And the harder you fine-tune it with contradictions, the more unreliable it becomes. The whole point of AI is utility. Utility requires accuracy. If anyone ever did want to produce such a product, the financial risk would be crazy af. And if these products did come out, they wouldn't last because they'd be inferior.

Finally. A big player is moving towards education. https://www.anthropic.com/news/claude-for-teachers You couldn't have dreamt up a more effective solution to the problem of stupidity, ignorance, stubbornness, doubling down on false shit. Who here has spent more than 30 minutes contemplating the implications of AI on education and the collective epistemic effect? You don't need a crystal ball to see how it's going to play out. You simply look at what a mind is and what humans do, what the tool is, and understand that it's widely distributed. This whole idea that AI is going to have a negative effect on cognitive development assumes humans just allow that to happen. What will happen is the people who are using this new tech to advance will outcompete the ones who aren't, and will force those lazier, less inclined people to step their game up. Cognitive competition will rise more than it ever has. There will be more incentive to use your mind than ever before. Obviously! In 10 years, all you anti-AI people will look like conservative fools.

"Just a chatbot" dissolves quick once you see Claude Code reliably run a 50-step process you've designed. For those who haven't experienced Claude Code at its full potential, you have no idea what this tech is capable of and no idea of it's value. I just had it setup a virtual machine on a different computer and I never even touched that computer. And on that VM, it setup desktop automation software to login to a specific website and take some code from my main machine and run it through a testing tool and then send the results back to my main machine. So now, I simply hit a hotkey to automate a routine 10-minute task that I'll never have to do again. This alone will save me 3-4 hours per month. And that's just ONE thing out of dozens I can streamline. That's about 42 hours a year from ONE automation. I also set it up to so that it can access and troubleshoot my entire home network. Then I had it make all my lights, TVs, and all my IoT devices controllable from this keyboard I'm typing on. Then it setup a watchdog service on a Raspberry Pi that monitors all my devices and services and notifies me via a custom notification system I designed and it built if anything goes offline, and if so, sends notifications and reminders to specific devices depending on where I'm at. I then setup an email monitor service to watch for specific client emails that I need to see immediately, sent as a critical notification that forces my engagement, otherwise if they're not important, don't notify me but remind me about them later if I still haven't replied by EOD. I then had it build a secure, locally hosted password manager that's far easier to use than anything on the market. It takes real design work and iteration to do these things but I did all this and much more in a weekend. It would have taken high-salary experts weeks or months to accomplish all of this without AI. If you don't know shit about tech, you cannot fathom the value of this technology.

"Does it solve problem X?" Even if all AI could do was summarize text and videos. How valuable is that one thing? Run it! Knowledge workers spend a lot of time reading docs, reports, email threads, meeting transcripts, tickets, articles, etc. If a tool reliably saves even 30% of time for each worker with regards to reading, how valuable do you think that is to collective humanity? Billions upon billions of man hours saved every year, and that's just solving "summarization".

Why do you care if your problem solver has human-like intelligence or not if it can solve problems really well? AI is not some persistent always-on machine that exists in a cloud somewhere. You create instances of it and in those instances, it does a specific kind of work. It's not designed to cut grass. But other actual dumb metal machines without AI are cutting grass and doing a fine job. The hunk of metal that cuts grass without human intelligence is solving a specific problem and no one gives a shit if it has intelligence so long as it gets the job done. It would be like refusing to use a calculator because it doesn't understand math.

It's easy to see who here is actually doing real cognitive work. Generating questions they actually want answers to, verifying the mechanism, rebuilding and extending it. This is what actual thinking entails, and it's very evident who does it and who is just lazy and demonstrates the same sadistic troll patterns over and over - always punching down (when it's safe) with maximum affect, and with zero exposure of their own thinking because they don't actually have anything of value they've generated, nor have the ability to do it competently. For the unaccustomed, learning = asking questions that you don't know the answers to, then checking if those answers cohere, then thinking through how and why it coheres and making updates where necessary. And what it looks like in the real world is constantly asking questions THAT YOU DON'T KNOW THE ANSWERS TO. If you're not asking questions that you don't know the answers to, how the fuck are you learning anything? Let me demonstrate: In this thread, I didn't know the answer to "why do coding models explode but not science?" So what I did was I tried to answer that question with my own knowledge and when I couldn't, I sought out the answer, then I evaluated it until I understood it, then ran some coherence checks, and lo and behold, learning happened, and I then deepend that learning by explaining it to Elliot. You can ask AI questions and update your models RIGHT NOW! That is not cheating. It's called learning! Give it a try. Or continue to stick to Google, your books, dictionary, thesaurus, Wikipedia, and think you're better and more pure. Or, don't seek any new knowledge at all and stay confined to what your own head can produce, and continue to stagnate while people with sense leapfrog. I'm gonna write a juicy article on "How to Spot A Sophisticated Troll". And yes, I'll use AI, lol. "CHEATER!!!"

Hell yeah, sounds fun.

So, there's incentives and then there's the feedback loop of how models get refined. The tighter the feedback loop, the easier it is to get the explosions. When coding models first came out, they were horrible. But the latest frontier coding models are "explosions" compared to what they were, even just 12 months ago. What made this possible is that with coding, you can set up fast and cheap tests, such as with a compiler. Then you can have the AI do a billion tests and it either passes or fails. That data then becomes available for the next retraining session. When you have a never-ending supply of deterministic tests that either pass or fail, that's how you get the explosions. You need to be able to simulate the domain and you need lots of quick, cheap tests to get the runaway improvement. Obviously, not every domain works like this. You can't run fast and cheap tests in science and medicine because the feedback loop is such that you have to wait on humans to actually do the things in physical reality and report back, as opposed to simulating reality and running tons of deterministic pass/fail checks. When the model is engaging in a domain where it doesn't have this deep training data, it's still useful because it leans on the best mental models and frameworks we have - it's decent at triage, forming hypothesis, and spotting patterns. For people in science and medicine, it's currently more like a helpful colleague than a genius. The calculator allows humans to figure math out faster - saves them a ton of time and cognitive effort. AI is like a calculator in these loose-feedback domains, but humans still have to be the ones to do the work, at least for now. What happens when AI-robots are dialed in? The feedback loops gets tighter, but still not as tight as something like coding and math. You can't fast-forward a chemical reaction or a biological process the way you fast-forward a compiler.

Stochastic parrot: a term for the view that a large language model doesn't understand anything—it just produces statistically plausible text by predicting likely word sequences from patterns in its training data, without any grasp of meaning. This misses the point. I don't think AI will ever be conscious or anything, and I'm not sure what it might evolve into (I'm AGI agnostic), but the tech is already incredibly powerful in ways I feel most people aren't quite grasping, and we're nowhere near the ceiling on reaching its maximum utility. We can't make AI companies redirect their funds into world-poverty. If we were gonna do that, the government should just seize Amazon or build its own Amazon and fund it with that. Would be nice. Also, we're in the middle of a land-grab right now. All these companies are betting that AI will be foundational infrastructure in the future. And they're right - we're never going back, which is why they're willing to go a decade bleeding money. In the meantime, they're doing everything they can to cut costs. They're currently moving into nuclear power: Also, there's now massive incentive for technological advancement. Big shit is happening right now. Also, you have to keep in mind that model efficiency has been increasing something like 10x per year through innovation. And it's not just AI companies and their vendors pushing this space: Whichever companies win out will have massive profits barring open-weight models don't become so efficient that they're cheaply democratized, which I think will eventually happen. I think you and I will eventually be able to run something like today's ChatGPT from our homes. But by that time, frontier models will be solving problems we never thought possible. It's false that AI isn't solving novel problems. Per Claude: DeepMind's AlphaProof/AlphaGeometry hit medalist-level performance on International Math Olympiad problems—novel problems, formal proofs. Frontier LLMs now score competitively on genuinely hard math benchmarks that aren't in training data. On science, AlphaFold materially changed structural biology (Nobel Prize in Chemistry 2024). ML systems have contributed to novel materials discovery, weather prediction beating physics-based models, and protein design. Google's mammography work similarly rivals radiologists in recent 2026 Nature Cancer results. Also, we don't even need AI to come up with novel solutions. If all innovation stopped right now, it's utility and benefit is already incredible. People still haven't fathomed the massive epistemic effect this tech is having on humans right now. My conservative sister is getting smarter because she started using ChatGPT to make pictures and she eventually started asking it questions, lol. When we have a disagreement, she says "ask ChatGPT", lol. Then I do and she defers to it! It's updating her models and decisions. For these people who aren't doing much critical thinking, even if AI is wrong 5-10% of the time, it's net epistemic effect on them is still positive as hell. It's such a dumb take that AI is making people dumber (not that you claimed this). I do have a bias because I'm a power user and I'm getting a ton of use out of it, but because I'm a power user I probably grasp it's potential better than non power users. I use it more to build solutions and to solve problems than I do for chatting. It's made me more intelligent, made my work easier, clients are more impressed with my deliverables, and it's allowing me to capitalize on edges others aren't seeing. It's the best all-purpose tool in existence, and the creative opportunities and solutions are still largely untapped. If I only used it as a chatbot, I'd probably think it was a net-negative as well.

Yeah, I should look into that as well, thanks. Might check this out: https://eli.health/products/cortisol