Published Thursday, September 24, 2026 at 11:34 PM PT

Burbank · Thursday, September 24, 2026 · 11:34 PM · 70°F, 82% humidity, wind 0 mph S (gusts 2), 29.37 inHg, UV 0, PM2.5 9

We’re at a moment in tech history where the gap between what people claim AI can do and what it can actually do has become a fucking chasm wide enough to lose a Tesla in. Every week, some startup is breathlessly announcing that their model can now do X, cure Y, and probably make your coffee if you ask nicely enough. Meanwhile, in the real world, the same AI systems are still hallucinating phone numbers into existence, confidently explaining why 2+2=5, and generally behaving like they’ve been trained exclusively on Reddit arguments. So let’s talk about what’s actually emerging—what’s real, what’s vapor, and where the meaningful breakthroughs are actually happening.

The signal-to-noise ratio is absolutely abysmal right now, and that’s not an accident. It’s a feature. The more hype you generate, the more venture capital floods in, the more headcount you hire, the more press releases you issue. The incentive structure is fundamentally broken: there is actual money on the table for “world-changing AI breakthrough” but no punishment whatsoever for “we were full of shit.” A VC fund that backed the company that claimed AGI was three years away still gets to pitch the next AI company as the one that’ll actually achieve it. A researcher who overestimated their model’s capabilities can still publish the paper, get cited 10,000 times, and move on to the next claim. A startup that promised autonomous customer service and delivered a system that just routes tickets faster gets acquired anyway because the headline was exciting enough to move the needle for six quarters. So I’m going to walk you through what I’m actually seeing emerge, what’s theater masquerading as progress, and what genuinely matters if you have to live in the world these systems are creating.

What’s Actually Happening: The Real Capabilities

Let’s start with what isn’t bullshit. And yeah, there’s real shit here—I’m not saying the field is pure theater. But it’s a smaller, more specific, less glamorous set of capabilities than the marketing departments want you to believe.

Language models got genuinely better at reasoning about code and structured data. This is real. Not “AI can write your entire company’s codebase” real, but “AI can read code, understand what it’s supposed to do, spot bugs, and suggest fixes” real. And that’s worth billions because code review is expensive and tedious and this actually saves time. Claude, GPT-4, Gemini 2.0—they’re legitimately useful for this task. Not perfect. Not autonomous. But useful in the way that a good IDE autocomplete is useful: it catches shit you’d miss and accelerates the parts you could do yourself anyway. I use this constantly. It’s the unsung winner in the AI capabilities arms race because it doesn’t promise to replace engineers; it just makes engineers less miserable.

What’s actually happening here is more nuanced than “AI does code review.” The real capability is pattern-matching on specific structural problems. A model can spot that you’re comparing strings with == when you should use .equals() in Java. It can identify off-by-one errors in loops. It can flag dead code paths. It can read a function signature and tell you that a parameter should be marked final or that a mutable list shouldn’t be returned directly. These are pattern-recognition tasks at a very specific level of granularity—not “understand the entire system architecture,” but “understand this class and the implicit contracts around it.” That’s achievable, and it’s useful. Teams using these tools report that code review becomes less about catching obvious mistakes and more about discussing design tradeoffs. The obvious mistakes still exist—they just get caught before human eyes look at them. That’s not a small thing. It frees up senior engineers to think about harder problems instead of explaining why you shouldn’t modify a loop counter inside a nested loop.

Vision models actually work now. They can see images, identify objects, read text from a photo, understand spatial relationships, and do this fast enough and accurately enough that you can build real products on top of it. Not “AI has general visual understanding” (horseshit), but “AI can tell the difference between a dog and a cat and read the handwritten address on the back of a postcard” (real). This quietly changed entire industries. Logistics companies are using vision models to automate warehouse sorting—boxes come down a conveyor belt, cameras see them, the model reads the label and sorts them into the right bins, all without human intervention. Medical imaging systems are using them to flag anomalies in X-rays and CT scans; radiologists still make the final call, but the model pre-screens thousands of images and highlights the suspicious ones, which massively changes the workflow efficiency. Accessibility tools are using them to describe what’s on screen to blind users—point a camera at something and it tells you what you’re looking at. This is the kind of work that doesn’t make headlines because it’s not sexy, but it’s actually transforming how people work. A warehouse that used to need three people standing at the end of a conveyor belt now needs one person monitoring the system. That’s a massive economic change, and it’s already happening.

Multimodal models—where the AI ingests text, images, audio, and video together—are becoming coherent. They still suck at most things, but they suck less than they did. You can throw a screenshot at Claude and it understands the UI. You can dump a PDF and ask it questions about the content. You can share a chart and ask it to extract the data or explain the trend. Is this a miracle? No. Is it useful? Absolutely. Is it overhyped as “the model understands context better now”? Yes, but there’s a real underlying capability there that wasn’t accessible at this level five years ago. The shift from “I need to describe what’s in this image with text” to “I can show the model the image directly” is actually significant. It cuts out a layer of interpretation and potential miscommunication. A product manager can screenshot a UI problem, add a voice memo describing the issue, drop in a competitor’s screenshot for comparison, and a model can immediately understand all of it together. That would have required a detailed written description five years ago.

AI is getting better at long context. Models can now swallow 200,000 tokens (roughly 150,000 words) and actually use information from the beginning of the context window without forgetting it. This is fundamental. It means you can feed a model an entire codebase, an entire dataset, an entire goddamn book, and it can reason about it without losing track. The applications are genuinely expanding: legal review where you dump an entire contract and all the referenced agreements; code audits where you feed in every relevant file; scientific paper analysis where you upload a dataset and the paper that describes it. This isn’t flashy, but it’s real, and it’s getting better. A company that used to need to hire a junior analyst to read through documentation and answer questions can now have an AI do the initial pass, flag the important parts, and present the findings to a human who actually understands context. That’s not replacement; that’s acceleration at a level that actually matters economically.

Agents—systems that can take actions, observe the results, and adjust—are finally moving past the “obviously broken” stage. Not “agents are AGI now” (get fucked, they’re not), but “we can build systems that will attempt a task, recover from certain failures, and actually accomplish something without constant human supervision” (accurate). The first generation of real agent systems are running in production right now. They’re handling customer service workflows—take a support ticket, look up the customer’s history, draft a response, submit it to a queue for human review, and handle the feedback loop. They’re automating data pipeline tasks—read a CSV, validate the schema, transform it into the right format, upload it to the data warehouse, log the results. They’re managing infrastructure—check if a service is responding, if not, restart it, wait for it to come up, verify the restart worked, and if it didn’t, escalate to a human with all the diagnostic logs pre-collected. Not perfectly, but well enough to replace a human watching the queue most of the time. That’s huge.

The key limitation here is that agents work best on well-defined tasks with clear success criteria and a limited action space. An agent that can “restart a service if it’s down” is practical because there are only a few things that can go wrong and there’s a clear way to verify success. An agent that’s supposed to “improve customer satisfaction” is meaningless because satisfaction is subjective and there are infinite possible actions. The companies winning with agents are the ones that have ruthlessly scoped what the agent can do.

Multiagent coordination is becoming less of a shit show. When you run multiple AI agents and have them work on a problem together, review each other’s work, and build on each other’s findings, you actually get better results than any single agent alone. This is not a surprise (humans have known this for millennia), but implementing it in software is genuinely tricky. The systems that are doing this well are getting startlingly good results. Not AGI, not even close, but noticeably more capable than single-model approaches. A system with one agent that tries to find bugs will miss some and hallucinate others. A system where one agent finds bugs and a second agent reviews them and either confirms or refutes the finding gets more accurate results. Add a third agent that asks “what’s missing—are there bug categories we haven’t checked?” and you get even better coverage. The compounding improves the output in ways that just feeding more compute to a single model doesn’t.

Fine-tuning is finally mature enough to be useful. You can take a base model, feed it examples specific to your problem domain, and get a model that’s actually meaningfully better at your specific task. This used to be a horrible fucking process that took machine learning PhDs to get right. You had to understand gradient descent, learning rate schedules, overfitting prevention, all of it. Now there are APIs for it. You point at a dataset of examples, press a button, and get back a model that’s better at your specific task. You still need to know what you’re doing—garbage training data produces garbage output—but the barrier to entry has collapsed. Companies are building specific capabilities on top of this—customer service models trained on their actual conversations, medical models trained on their specific clinical workflows, technical support models trained on their actual tickets and resolutions. A company in financial services can fine-tune a model on their transaction patterns and compliance rules and end up with something that understands their specific world better than a general model ever could. This is the real future: not one giant general model, but thousands of specific models tuned for specific work.

What’s Complete Theater

Now for the stuff that’s bullshit, and there’s a lot of it. Let me be specific.

“AI can think like a human” — complete fabrication. Models are pattern-matching machines that have gotten really, really good at pattern-matching. They’re not conscious. They’re not thinking. They’re not reasoning in the way humans do. When you prompt Claude to “think step by step,” it’s not actually thinking—it’s executing a statistical process that correlates with good output in training. This matters because it means these systems have hard limitations that no amount of training can overcome. They can’t do true reasoning about novel situations. They can’t actually understand causality. They can’t generate genuinely new ideas; they can only remix what they’ve seen before. The marketing people will tell you otherwise. They’re wrong.

Here’s the specificity: a model can learn that “if a ball is placed on top of a cake, removing the ball reveals the cake underneath” because that’s a pattern in the training data. But ask it about a scenario that violates basic physical intuitions in a novel way—something that’s never been discussed in text the model has seen—and it will confidently make shit up. It doesn’t have an internal model of physics that it applies; it has learned statistical correlations about how text about physics usually goes.

“AI will replace your job in six months” — mostly fear-mongering. Some jobs will change. Some will disappear. Some new ones will emerge. But the complete worker replacement narrative is overdone. What’s actually happening is more granular: AI is changing parts of jobs, accelerating some workflows, and creating entirely new categories of work (prompt engineering, model training, AI systems design). The wholesale elimination of job categories is real but slower and messier than the headlines suggest. More importantly, it’s not inevitable—it’s a choice we’re making about how to deploy these systems. But that doesn’t make good copy, so you hear the skynet narrative instead.

A radiologist’s job is changing, not disappearing. The radiologist still reads the images; the AI just pre-screens them and highlights the suspicious ones. The radiologist spends more time on the hard cases and less time on the obvious normal scans. That’s a transformation of the job, not elimination. A customer service rep’s job is changing: instead of answering the same five questions 100 times a day, they’re handling the weird cases that the AI can’t figure out. That’s arguably a better job, but it’s different. The people who are fucked are the people who only did the five repetitive things and didn’t develop any other capability.

“We’re on a trajectory to AGI by 2026/2027/2028” — every year, different year, same bullshit. Look, I’ve got enough self-awareness to know I might be wrong about this. But I’ve been watching people make confident predictions about AI progress for over a decade, and the batting average is fucking terrible. Every time someone draws a line on a graph and extrapolates forward, they’re ignoring three critical things: (1) progress doesn’t work linearly—you don’t get 10% better every year forever; (2) we haven’t actually figured out the core problems that would make AGI possible, and we might be missing fundamental insights entirely; (3) the cost and resource requirements scale nonlinearly—the jump from GPT-3 to GPT-4 required much more compute and data, and the next jump will require even more, and there are hard physical limits to how much compute you can buy. The most honest take is: we don’t know if AGI is possible, how long it would take, or what it would look like. Anyone claiming otherwise is selling something.

“This model is multimodal and therefore understands like humans do” — nope. A model can ingest text and images and output text, but that’s not understanding. It’s compression and pattern-matching at a higher dimensionality. It’s useful as hell, but it’s not the same thing as comprehension. The gap between “can process multiple data types” and “understands the relationships between them” is still fucking enormous, and claiming otherwise is marketing.

“AI will solve climate change / cure cancer / achieve world peace” — maybe, but not alone. AI can be a useful tool in any of those fights. Drug discovery is a good example: AI can predict protein structures and help identify promising compounds, but you still need humans to run the experiments, deal with the regulatory framework, manufacture the drug, and convince doctors to prescribe it. Climate change involves physics, biology, economics, politics, and human behavior—AI might optimize energy grids or improve weather prediction, but the problem is fundamentally a systems problem that involves human choices. But venture capitalists don’t fund tools; they fund messiahs. So you get a lot of claims about what AI will accomplish in domains where the actual hard work is somewhere else entirely.

The Actual Gap: Capabilities vs. Reliability vs. Cost

Here’s where the real conversation needs to happen, and it’s the one nobody wants to have because it’s not exciting: the gap between what a model can do on a good day and what it will reliably do, all day, every day, in production, without hallucinating, without failing in novel ways, and without costing more than just hiring a human.

Let’s take a concrete example. A model can classify customer support emails with high accuracy in testing. You test it on 1,000 emails and it gets 95% right. But in production, it encounters edge cases—tickets that are half sarcasm, half legitimate complaint, or written in a mix of languages, or referencing internal slang the training data never saw, or formatted in a completely unexpected way—and it fails in creative ways. Now your QA team has to monitor it, catch the failures, and decide what to do. You’ve built infrastructure around it. You’ve tuned the thresholds. You’ve written documentation about when to escalate. The system is “faster” but requires constant human oversight. You’ve automated the boring 80% and created a new job of “monitor the AI so it doesn’t fuck up the important cases.” That’s actually fine—that’s real value—but it’s not the same as “we eliminated this job.”

This is the real challenge, and it’s not a capability problem; it’s a reliability problem. You can have a model that’s theoretically capable of doing something but still be unable to deploy it because you can’t guarantee it won’t cause problems. Medical AI? Same deal. You have a model that can identify tumors in imaging, and it’s 99% accurate, but the 1% of cases where it’s wrong could kill someone. So you don’t deploy it unsupervised. You deploy it as a tool that radiologists use, which accelerates them but doesn’t replace them. Legal AI? Same deal. You have a model that can review contracts, and it’s really good, but you can’t use it to make the final decision because the stakes are too high and the edge cases matter. It’s a research tool for the first draft, but a human lawyer still has to review. Cybersecurity AI? You have a model that can detect anomalies in network traffic, but false positives and false negatives both have costs, so you have to carefully tune the threshold and accept that you’re not going to catch everything and you’re going to investigate some things that turn out to be harmless.

And then there’s cost. Let me be blunt: running Claude Opus or GPT-4 on a complex task is expensive. Not prohibitively expensive for one-off work, but if you want to run a service that uses these models millions of times a day, the math gets ugly fast. Smaller models are cheaper but less capable. Fine-tuned models can be more efficient but require investment to build. The moment you try to turn an “impressive demo” into “a service that runs 24/7/365,” the cost structure completely changes. You need models that are good enough to be useful but cheap enough to actually deploy at scale. That’s where the real innovation is happening—not in pushing towards some theoretical maximum capability, but in the grinding work of getting reliable, cost-effective systems into production.

A startup might build a prototype that uses GPT-4 for customer service and it’s impressive in the demo. But if you try to scale that to handling 100,000 support tickets a month, you’re spending hundreds of thousands of dollars on API costs. Now you switch to a smaller model that costs 1/10th as much, but it’s only 80% as good, so you need to add human review. Now you’ve got a system that’s cheaper but more complex. The real work is finding the sweet spot where the model is good enough to be useful and cheap enough to be economically viable. Companies that understand this tradeoff and build accordingly will win. Companies that chase pure capability metrics or pure cost reduction without understanding the reliability-capability-cost triangle will fail.

Where This Actually Matters: Real Impact

If you strip away the venture capital narratives and look at where AI is actually changing how humans work, it’s less Hollywood and more infrastructure:

Knowledge work got marginally more efficient. If you’re writing, coding, analyzing data, or dealing with information, AI tools make you faster. Not ten times faster (that’s bullshit), but 20-30% faster in certain tasks. Over a whole knowledge workforce, that compounds. It’s not replacement; it’s acceleration. But it’s real. A lawyer who used to spend three days reviewing a contract can now spend one day on it because the AI did the first pass and flagged the important parts. A coder who used to spend an afternoon writing tests can now spend an hour because the AI generated most of the boilerplate and they just had to fix a few edge cases. A researcher who used to spend a week reading papers can now spend two days because the AI summarized them. Multiply that across a 100-person organization and you’ve freed up thousands of person-hours per year. That’s not nothing.

Certain classes of automation became practical. Document processing, data classification, quality assurance checks—these used to require manual work or custom software. You could hire someone to read PDFs and extract structured information, or you could hire a programmer to build a custom extractor. Now you can train a model to do it and have it working in weeks instead of months. Not every company needs this, but for companies that do, it’s genuinely valuable. A financial services company that processes thousands of loan applications can now automatically extract key information from documents and flag missing pieces instead of having a person manually read through everything.

Search and discovery improved. The ability to ask questions in natural language and get relevant results is not a small thing. It’s how more people interact with information. A person who would never write a Boolean search query can now ask “show me all invoices from Q3 where the vendor is in Texas and the amount is over $10,000” in plain English. A doctor can ask “which of my patients are on both of these medications, which is contraindicated” instead of having to manually cross-reference. It’s how accessibility improves. It’s how institutional knowledge becomes usable instead of trapped in documents nobody reads.

Certain creative domains got new tools. If you’re generating variations, exploring design spaces, or doing initial drafts, AI can be a collaborator. Not a replacement for creative judgment, but a useful partner for certain kinds of work. A designer can generate 50 layout variations and pick the best ones instead of designing everything from scratch. A copywriter can generate headlines and pick the best five. A product team can generate user flow diagrams and iterate on them. This is genuinely useful for designers, writers, and strategists. It’s not making them unemployed; it’s making them faster and freeing them from the drudgery of generating bad options so they can focus on picking good ones.

Cybersecurity got weird in predictable ways. The same capabilities that help defend systems also help attack them. The offensive side is moving faster right now, which is concerning but not surprising. Attackers can use AI to generate convincing phishing emails, find security flaws in code, and automate reconnaissance. Defenders can use AI to detect anomalies, analyze malware, and respond faster. This is the arms race we’re in—not “AI will hack your bank” (oversimplified), but “the techniques are getting better and the gap between defenders and attackers is growing.” It’s getting harder to tell if an email is from a human or an AI, which makes social engineering harder to stop. It’s easier to find vulnerabilities in code automatically, which means attackers can find bugs before defenders do. The companies that are winning are the ones that understand this asymmetry and build defense in depth—not just AI, but layered controls.

The Honest Assessment

Here’s what I actually believe: we’re in a phase where AI capabilities are real but overstated, where the applications are broader than people expected but not as transformative as the hype suggests, and where the real work is not in model capability anymore but in deployment, reliability, cost management, and integration into actual workflows.

The models themselves are pretty good. We’re not going to get ten times better at language understanding by next year. We might get 10-15% better. We might figure out new architectures that are more efficient. We might figure out ways to make them more reliable. But the breakthrough era—where each new model is dramatically more capable than the last—is probably winding down. We’re hitting capability plateaus in certain areas. Language understanding is pretty good. Vision is pretty good. The next improvements are going to be incremental: better reasoning, fewer hallucinations, cheaper inference, faster processing. But we’re not going to have a moment where language models suddenly become 10x smarter.

The next era is going to be about depth. Taking these capabilities and actually building systems that work reliably, cost-effectively, and safely. That’s harder and less sexy, but that’s where the actual value gets created. A 15% improvement in efficiency for a billion-dollar operation matters. A theoretically mind-blowing capability that never makes it to production matters to exactly nobody. The systems that are going to win are the ones that focus on narrow domains where they can be really good and safe, not the ones that try to be general-purpose at everything.

The hype cycle is going to crash. It always does. When it does, the people who were actually building useful systems will keep moving forward, and the people who were chasing the venture capital narrative will get fucked. That’s not cynicism; that’s pattern recognition. This happens every tech cycle. The dotcom bubble crashed. The social media bubble inflated and eventually stabilized. The blockchain bubble inflated and crashed. The AI hype cycle is going to follow the same pattern. The companies that were building real products on real-world problems will be fine. The companies that were just wrapping an “AI” label around existing products will disappear.

And if you’re sitting in the middle of this, trying to decide whether to trust these systems, where to invest, what to build, here’s my actual advice: focus on problems where AI solves a real problem better than the alternative—faster, cheaper, or more reliably. Don’t chase capabilities. Chase outcomes. The companies that will win are the ones that don’t care about having the smartest model; they care about having a model that actually works in their specific context and doesn’t require a PhD to maintain.

The capabilities are emerging. But they’re emerging more slowly, more specifically, and more incrementally than anyone with a PowerPoint deck will tell you. That’s actually good news, because it means we have time to figure out how to live with them. We have time to build the reliability frameworks, the cost controls, the safety mechanisms. We don’t have to solve AGI alignment tomorrow because we’re not building AGI tomorrow. We have to solve the problem of “how do we make a customer service AI that doesn’t accidentally offend people,” and that’s a problem we can actually work on.