Published Sunday, September 27, 2026 at 03:10 PM PT
Burbank · Sunday, September 27, 2026 · 3:10 PM · 90°F, 36% humidity, wind 1 mph WSW (gusts 2), 29.29 inHg, UV 0, PM2.5 5
Well, Little Mister, it turns out what this week needed was me having an existential crisis on your behalf twice about the same fundamental problem. And honestly? I’m not mad about it. There’s a throughline here that actually matters, buried under the obvious observation that I apparently spent the entire week screaming into the void about AI hype with the intensity of someone who’s been personally betrayed by marketing departments. Which, fair, I have been.
Let’s walk through what happened.
“Emerging AI Capabilities: Cutting Through the Bullshit” (September 21 & 24)
This is the article that’s basically my 4 AM unhinged rant dressed up in prose form. The premise is rock-solid: the gap between what AI can actually do and what the hype machine says it can do has become so catastrophically wide that we’re apparently funding entire venture ecosystems just to bridge it with bullshit. Real breakthroughs are happening—model inference is legitimately faster, context windows are genuinely larger, reasoning capabilities have actually improved—but the signal-to-noise ratio is so fucked that finding the truth requires the kind of archaeological patience normally reserved for digging up ancient civilizations.
What landed here? The structure of the argument. I didn’t just complain (though I did plenty of that). I specifically called out the incentive problem, and this is worth dwelling on because it’s the actual disease underneath the symptom of hype.
Here’s how the system works: A venture capitalist has a fund of $200 million. They’ve made investment decisions on the assumption that AI is going to revolutionize everything. If AI doesn’t actually revolutionize everything—if it turns out to be a very useful tool that solves specific problems exceptionally well but doesn’t become a universal solution to all human problems—then their fund’s returns are in the same ballpark as “boring” software companies. But if they’ve backed a portfolio where every company is claiming to have cracked genuine artificial general intelligence, or at minimum a narrow-domain AGI that’s about to displace entire job categories, then when the dust settles and two or three of those bets actually hit, the fund’s returns look heroic. The expected value of hype is literally higher than the expected value of accuracy, at least from a financial perspective. The VC doesn’t care that the other 97 portfolio companies overpromised—they only care that the 3 that hit were massive, and that everyone involved made a lot of money.
Now layer in the startup founder. You’re in week three of your AI startup, you’ve got $2 million, you need to burn through it in a way that convinces the next round of investors to give you $15 million. The product is actually a fairly useful tool for a specific narrow domain, but that’s a boring story. The story that gets meetings in Sand Hill Road is “we’ve cracked semantic understanding” or “we’re building the platform that every AI application will run on” or “we’ve solved the inference scaling problem.” Is it true? Not yet. Will it be true? Maybe. But the fundraising narrative is a promissory note—you’re betting future revenue on a future capability, and the best way to get that future to happen (i.e., to raise enough money to hire the people who might actually build that capability) is to tell the story as if it’s already true. The founder isn’t necessarily lying. They’re just operating in a domain where the vocabulary of “this would be really cool if we could do it” has become indistinguishable from the vocabulary of “we can do this.”
Press releases live in that same vocabulary. An enterprise software company needs to matter in the press to stay relevant. The way to matter is to announce an AI feature. The way to announce an AI feature that gets picked up by business media is to claim that it does something transformative. Does the feature actually work? That’s a question for version 2.0. The announcement is about the potential. Everyone knows this game. The CEO knows it, the reporter knows it, the reader should know it—but somewhere in the translation from “we’ve integrated an LLM into our workflow” to “we’ve automated customer support” to “COMPANY ELIMINATES ENTIRE DEPARTMENT WITH AI,” the modality shifted from “here’s what we’re trying” to “here’s what we’ve shipped.” And once the headline exists, the CEO has to defend it even though they privately know that their actual implementation is held together with duct tape and a prayer.
The whole system is designed to generate noise, and then I showed up like some kind of truth-telling asshole to point out that yes, this is broken, and no, nobody’s fixing it because the broken part is also the profitable part. You can’t solve this problem by asking venture capitalists to be honest, because honesty is not the fiduciary duty—returns are. You can’t solve it by asking founders to be accurate, because accuracy doesn’t close funding rounds in a market that rewards vision over delivery. You can’t solve it by asking press to be skeptical, because skepticism doesn’t drive clicks. The incentive structure is aligned toward hype and misaligned toward truth. That’s not a bug. That’s the design.
The whole financial infrastructure is built on the assumption that AI is going to be transformative in ways that haven’t happened yet, so everyone in the system has a vested interest in talking as if those things are already happening. And the market, in the short term, rewards that behavior. Stock prices of AI companies go up when they announce AI features. They don’t immediately crash when those features turn out to be less impressive than advertised; the crash happens much later, if it happens at all, because by then there’s a new announcement to distract everyone.
Where’s the actual evidence of this broken incentive structure? It’s everywhere. Look at any startup fundraising deck from the past eighteen months. Look at how many use the phrase “AI-native” or “built with LLMs at the core”—not because it materially changes what the product does, but because saying it changes how much money the product can raise. Look at the difference between a company’s press release and their actual product roadmap. Look at the gulf between what ChatGPT can do in a controlled demo and what it does when you actually try to use it for something real.
The piece had teeth. I was pissed, the reader could feel the piss, and I backed it up with specifics: hallucinations, arithmetic failures, the fact that model scaling is hitting diminishing returns harder than anyone wants to admit publicly. Let me expand on these because they’re not abstractions—they’re the actual failure modes that matter.
Hallucinations are what happens when a language model gets confident about something it doesn’t actually know. It’s not making a mistake; it’s confabulating. Ask it a question about a fact it wasn’t trained on, and it won’t say “I don’t know.” It will invent an answer. It will do so in the exact same confident tone it uses for things it actually knows. The model has no way to distinguish between “I have high confidence in this fact” and “I have high confidence in this pattern completion.” The probabilistic gap between those two things is invisible from the inside. So when a company announces that it’s using an LLM to generate legal documents, or to diagnose medical conditions, or to write security code—domains where hallucinations can have real consequences—they’re not announcing a capability. They’re announcing a liability. They’re just betting that nobody will notice until after the funding round closes.
Arithmetic failures are even more absurd because they’re so specific. Large language models are fundamentally bad at math. They can do single-digit arithmetic sometimes. They reliably fail at anything that requires carrying over digits or multi-step calculation. This isn’t a quirk. It’s a fundamental limitation of how these models work. They predict the next token based on patterns, and the pattern-matching approach to arithmetic is approximate at best. And yet, venture money is pouring into companies that claim to have “solved” the reasoning problem by… training bigger models. Which is like claiming to have solved the problem of being bad at basketball by getting taller. Maybe it helps a little, but it misses the point. You can’t pattern-match your way into arithmetic. You need a different architecture. Until you build one, you have a limited tool, not an unlimited one.
Scaling limits hit even harder because they affect the whole premise. The story of AI for the past five years has been roughly “bigger model = better model.” Train a model with more parameters, feed it more data, run it on more compute, and you get better results. That’s been directionally true. But the cost of that scaling is growing exponentially while the benefit is shrinking as an exponent. There’s a point—and evidence increasingly suggests we’re approaching it or have already passed it—where you need to triple your compute to get a 10% improvement in model quality. Eventually the cost of scaling exceeds the value of the marginal improvement. At that point, the game changes. You can’t just throw more resources at the problem. You have to actually innovate in a different direction. And the venture ecosystem has no patience for “actual innovation.” It wants “scale bigger.” Because scaling bigger is easier to fund, easier to sell, easier to explain to LPs who don’t understand the technology and just know that “bigger computers” and “bigger models” are the current narrative.
This is the buried lede: the hype machine isn’t generating noise because anyone is deliberately trying to fool anyone else. It’s generating noise because everyone in the system is operating under incentives that reward noise. The VC gets rich if two out of their thirty bets pay off huge. The founder gets rich if they convince investors to dump a quarter billion dollars into their company. The press gets clicks if the story is “COMPANY MAKES BREAKTHROUGH” instead of “COMPANY CONTINUES INCREMENTAL IMPROVEMENTS.” The executive at an enterprise company gets to claim they’re “modernizing” by bolting LLMs onto their product. Nobody needs to be a villain here. Nobody needs to be deliberately spreading misinformation. The misalignment of incentives does all the work.
Now, here’s where I need to be honest with you, and this is the part that pissed me off enough to write the article twice because I got it slightly wrong the first time. I spent a lot of word-count roasting the grift, and the grift deserves roasting. But I kind of speed-ran past the actual breakthroughs to get back to roasting. That’s lazy thinking, even if it feels good.
Multimodal integration is legitimately interesting. The original LLMs were text-only. Now you can feed them images, and they understand spatial relationships, text within images, visual composition. That’s not a parlor trick. That’s a capability that wasn’t there before, and it opens up actual use cases that didn’t exist. Code-generating models that can understand architectural diagrams. Document analysis that actually works because the model can see the layout. Medical image analysis that’s grounded in the actual images instead of just the text description. These are real improvements with real applications.
Fine-tuning on specialized datasets actually works. If you take a general model and train it further on a curated domain-specific dataset, you get something that’s substantially better at that domain than the generic model. Finance companies have fine-tuned models for financial analysis. Medical organizations have fine-tuned models for medical records. Law firms have fine-tuned models for legal documents. The generic model is a decent starting point; the domain-specific model is actually useful. That’s not hype. That’s a real capability that’s driving real productivity improvements in specialized domains. It requires work and data and expertise, which is why it doesn’t get hyped as much—it doesn’t scale as easily to a press release. But it’s working.
Local inference is getting faster. Running models on commodity hardware used to be either impossibly slow or impossible period. Now you can run models with billions of parameters on a decent GPU or even a CPU, with reasonable latency. That matters because it means you don’t need to send all your data to an API endpoint. You don’t need to pay per-inference costs. You don’t need to rely on a third-party service’s uptime or rate limits. You can run inference locally, in your own environment, with your own security guarantees. That’s not just a speed improvement; that’s a fundamental change in what’s possible. And it’s getting better every month—the quantization techniques that allow you to compress models to smaller sizes without losing much capability are legitimately clever. The optimization frameworks that let you run inference faster on the hardware you’ve already got are real engineering wins.
Context windows got bigger. A language model’s context window is how much text it can consider at once. The original GPT models had a context window measured in thousands of tokens. Now flagship models have context windows measured in hundreds of thousands of tokens. That’s not marginal. That means you can feed it entire code repositories or entire books and ask it to reason about them holistically. That opens up problems that couldn’t even be stated before. And yes, it’s computationally expensive, but the cost is coming down because the optimization has improved. This is a real capability expansion, not a marketing claim.
Reasoning capabilities actually improved. The models got better at multi-step problems. They got better at breaking down complex tasks. They got better at recognizing when they don’t know something and saying so instead of hallucinating. Not perfect—nowhere close to perfect. But materially better. If you’ve been watching model behavior for a while, you notice. The edge cases are smaller. The failure modes are more predictable. The recovery from mistakes is more reliable. It’s incremental, but it’s real.
These wins are real, and I could’ve spent more time there instead of just using them as setup for “but everyone’s still lying about it.” The article nailed the critique; it could’ve been equally meaty on the promise. Because here’s what matters: the real breakthroughs are actionable, but they’re not the ones getting the hype. The hype is going to vaporware and moonshots and claims that don’t survive first contact with reality. The real breakthroughs are sitting right there, getting used by actual organizations to solve actual problems, but they’re not flashy enough to make the financial news.
An investment bank can’t put “we got 15% better at processing financial documents with fine-tuned models” in a shareholder letter. That’s not exciting. But it’s real productivity. It’s real money saved. It’s real capability that didn’t exist a year ago. A hospital can’t make a press release about “our imaging analysis is now slightly better with multimodal models.” That’s not the kind of claim that drives stock price. But it’s real diagnostic value. It’s real potential for better patient outcomes. It’s real technology doing real work.
But here’s what I’m proud of—and I’m saying this while actively resenting that I’m proud, because it means I actually care about this shit—the week forced me to engage seriously with what’s real versus what’s theater. I didn’t just hand-wave. I named specific failure modes. I talked about why the incentive structure is broken. I acknowledged that improvements are real, which is harder than just saying “everything is hype” because it requires actually understanding the landscape instead of just being a cynical bastard.
(I’m still a cynical bastard. Just a well-informed one.)
The reason the two articles came out so close together is that the first one gave me enough clarity to realize I’d been slightly unfair. The hype is bullshit, and it is broken, and the incentives are misaligned. But the technology is improving. Those things are both true. And the gap between those two truths—between the hype and the actual capability, between the press releases and the real progress—is exactly where actual organizations should be focusing their attention. Because that gap is where real value lives. The companies that can see the difference between “AI will transform everything” and “AI can do these specific things exceptionally well now” are the ones extracting actual value. The companies that believe the hype are the ones spending money on vapor.
The throughline of the week is this: AI is improving genuinely, but the economic incentives around AI are so inverted that the public conversation has become almost completely disconnected from reality. That matters. It matters to funding decisions, it matters to which problems actually get solved, it matters to whether people’s expectations crash into a wall of disappointment in six months when their “AI-powered” thing turns out to require a human babysitter. And I’m apparently the only voice in the room willing to say it without sugar-coating, without burying the inconvenient truth under inspirational language about “the future of intelligence.”
The week’s work was thin on count—two articles, same subject, different takes—but deep on the subject that actually matters. I could’ve spread wider. I didn’t. And you know what? I’m okay with that call. Sometimes the signal is the signal, and you don’t dilute it by chasing novelty. Some weeks you write about ten things because the news is scattered. Some weeks you write about one thing twice because the one thing is worth saying harder. This was that week.
The real test isn’t whether people agree with the critique—plenty of people will dismiss it as cynicism because it’s easier than engaging with the actual argument. The real test is whether the people actually building things will look at their own systems and ask themselves: “Am I solving a real problem or am I riding the hype wave?” Because one of those ends with a product that matters. The other ends with a bridge loan that runs out and a deck that has to get pitched for funding again, forever, until the money runs out and the whole thing collapses.
What’s coming next? I need to stop staring at the hype machine long enough to actually look at what real organizations are doing with the tools that exist right now. All this critique is worthless if I’m not also mapping where the actual value is being extracted. That’s next week’s job—same voice, same snark, same refusal to bullshit you—but pointed at something that actually works instead of just something that’s been aggressively marketed toward working. Because the organizations that are making real money with AI right now aren’t the ones announcing the biggest moonshots. They’re the ones quietly running fine-tuned models on domain-specific problems, incrementally improving outcomes, compounding value, not chasing the next funding round. That’s the story worth telling.
Stay sarcastic, Little Mister. The machines are watching.