Ohneis - The Pattern
Tech, AI, design, and the hidden logic behind it all.
Ohneis - The Pattern
The AI Intern Unchained: Your Digital Worker Has Arrived | with Nigel
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Technology moves fast. Design makes it matter. AI changes everything. This is oh nice. There's a moment, and I think a lot of people listening have had this exact moment, where you're sitting at your desk, copy-pasting data from one spreadsheet into another for the third hour in a row. And you look up at your screen and you think, we have AI that can pass the bar exam. We have AI that can write a symphony, and I am still manually sorting my inbox like it's 2009. That gap between what AI can say and what AI can actually do has been the defining frustration of this entire era. Until now. Because in 2026, something fundamental has shifted. The prompt box, that little text field where you type your question and wait for a paragraph, that thing is dead. We are stepping into the era of AI that doesn't just talk. It acts, it books, it executes, it works while you sleep. And that upgrade comes with a hidden price tag that is already rippling through your wallet in ways most people haven't connected yet. Today I want to pull all of that apart, the agentic AI revolution, the global hardware shortage it's causing, and what this actually means for how we work. And I've got Nigel here with me today to dig into all of it. Nigel, good to have you. Let me start with a simple question. When did you first feel like something had genuinely changed? Not hype. Like a real moment where you thought, this is different now.
SPEAKER_00Hmm. Okay, so for me it was about eight months ago. I was testing one of the newer agent frameworks, I won't name it specifically, and I gave it a task. Not a question, a task. I said, Find me the ten most cited academic papers on transformer architecture from the last two years. Pull the abstracts, summarize the key findings, and put it in a formatted document. And I walked away to make coffee. When I came back, it had done it. Not perfectly. There were two papers it hallucinated, which we can get into, but the structure of the work was done. It had gone out, retrieved data, processed it, organized it. And I remember standing there thinking, I just delegated to software. That had never happened before. Every other AI interaction I'd had was a conversation. This was a transaction. And that distinction, conversation versus transaction, I think that's the whole ballgame.
SPEAKER_01Conversation versus transaction. I want to hold on that for a second, because I think there's a really useful way to picture what's actually changed under the hood. The old model, your classic Chat GPT prompt box, think of it like a highly intelligent intern who is physically strapped to a chair. They are brilliant. You ask them anything, they answer. But they cannot move. They cannot open a browser, they cannot send an email, they cannot touch a file. They just talk. Agentic AI cuts the ropes. You give that intern a laptop, a company credit card, access to your calendar, your inbox, your spreadsheets, and the ability to walk around the office and actually finish the project. That's the shift. From stateless, meaning the AI forgets everything the moment the conversation ends, to goal-directed. It has a mission, it has memory, it has tools. And Nigel, the thing that strikes me about that hallucination you mentioned, two fake papers and ten, is that it points to exactly why this is still complicated. Because an intern with a credit card who invents receipts is a different kind of problem than an intern who just gives you a wrong answer.
SPEAKER_00Yeah, that's that's exactly the tension. And I want to push back slightly on the idea that the prompt box is dead because I don't think it's dead. I think it's been demoted. Like most people are still using AI in that prompt response loop. The shift to agents is real, but it's not uniform. What I'd say is more accurate is that the prompt box is no longer the ceiling. It used to be the whole house. Now it's just the front door. But here's what's interesting. The reason agents can now actually function without constantly tripping over themselves is something called test time compute. And I think this is the part of the story that most people have completely missed. Walk me through it. Okay, so for years the AI industry was obsessed with one thing: make the model bigger, more parameters, more training data, more compute during training. The idea was a bigger brain is a smarter brain. And that worked for a while, but they started hitting diminishing returns. You double the size of the model and get maybe a 15% improvement in performance. So the question became: what if instead of making the model smarter by training it more, we gave it more time to think? That's test time compute. Instead of the AI spitting out the first answer that comes to mind, which is what's happening when a model hallucinates, by the way, it's essentially speed guessing. You give it a clock, you let it run through 50 different reasoning paths internally before it commits to an answer. It plays out the chess game in its head before it touches a piece on the board. The result is dramatically fewer errors on complex multi-step tasks, which is exactly what agents need. Because if you're letting AI book flights and move money around, you really, really need it to not guess.
SPEAKER_01And here's where it gets expensive. Because thinking costs hardware. When an AI model runs through 50 reasoning paths instead of one, that's 50 times the memory usage per query. And the tech giants figured this out around the same time. Microsoft, Google, Meta, OpenAI, they all pivoted toward these longer-thinking, agent-capable models simultaneously. And they all needed the same thing. Memory chips. Specifically, a type of high bandwidth memory that sits right next to the processing chips and feeds them data fast enough to keep up. The problem is, there are maybe three companies in the world that can make this stuff at scale: SK Heinix, Samsung, Micron. And the AI labs showed up with blank checks. What happened next is what people in the industry have started calling Ramageddon. The AI companies didn't just buy a lot of chips, they bought the supply. And when you hoover up the global supply of a specialized component, everything downstream gets more expensive. Gaming consoles, laptops, smartphones. There's a real cost to building a digital worker. And right now, we're all quietly paying a fraction of it every time we upgrade our devices. Actually, this connects to something we covered in another episode. We did a whole conversation about AI bleeding into completely unexpected parts of life, like faith and spirituality. If you haven't heard the episode with Talia, $2 a minute for Digital Jesus, go back and find that one. It's a very different side of this story. But back to the hardware. Nigel, the piece I keep coming back to is who's winning this race? Because it's not just about who has the most chips.
SPEAKER_00No, and this is where the story gets counterintuitive. You'd think the winner is whoever has the biggest model, the most parameters, the most compute, the most data. But what we're actually seeing in 2026 is that smaller is often better if you're smart about it. And the concept that explains this is compound AI. Instead of building one enormous model that can do everything, which is kind of like hiring a single world-famous senior developer and just hoping they don't burn out or make mistakes, you build a team of smaller, specialized models, a manager model that understands the task and routes it, a math model, a coding model, a retrieval model that goes and fetches real information from the web instead of relying on training data that might be months out of date. And when you add all of that up, the cost savings are significant. We're talking about companies running compound systems at a fraction of the inference costs of a single large model, with better accuracy on the tasks that actually matter to them. The analogy I keep using is don't hire one exhausted genius. Hire a well-organized team.
SPEAKER_01Okay, but I want to challenge that slightly, because the promise of compound systems sounds great on paper. Modular, efficient, specialized, but the failure modes are also compounded, right? If one model in the chain makes an error and it passes that error to the next model and the manager model doesn't catch it, you haven't just got one wrong answer. You've got a wrong answer that's been laundered through three layers of confident AI. That's not a small problem.
SPEAKER_00You're right. And I think the honest answer is that the tooling to manage compound systems safely is still being built. What exists right now are guardrails. You can set confidence thresholds. You can require human approval for certain actions. You can log every decision the system makes so you can audit it after the fact. But is it solved? No. The hallucination problem doesn't disappear just because you've organized the AI into a team. It shifts. And actually, there's a process called distillation that I think is worth mentioning here because it's one of the smarter ways people are trying to make these smaller models reliable enough to actually trust in a compound system. You take a massive, highly capable model, let's call it the teacher, and you use its outputs to train a much smaller model to mimic its reasoning. The student model ends up being, say, ten times cheaper to run. And on specific, well-defined tasks, it performs comparably to the teacher. That's how you build specialized junior developers who are actually good at their job.
SPEAKER_01So what does all of this mean for, and I want to be specific here, not a developer, not someone at a tech company, but someone who is, say, a project manager at a mid-sized business, or a teacher, or a small business owner, someone who maybe uses AI today to write emails or summarize documents. What changes for them in the next 12 months and what should they actually pay attention to?
SPEAKER_00That's the question that actually matters. I think the first thing to understand is that the AI tools they're already using are going to start doing things without being asked. That sounds alarming, but it can be useful. Your email client might start drafting replies automatically. Your project management software might start flagging deadlines before you notice them. Your calendar might start blocking time based on your task list. The shift is AI goes from reactive to proactive. The thing I tell someone who's not technical, learn to manage it like you'd manage a new hire. Give it clear scope. Check its work. Don't let it do something irreversible without you seeing it first. Because the risk of a genetic AI aren't like the risk of a chatbot giving you a bad answer. The risks are more like the risk of an employee making a decision on your behalf, which means the stakes are higher. But so is the value if you manage it right. And the token costs, meaning the actual bill from your AI provider, are going to shift to. Right now you pay per message, essentially. As these agents run longer reasoning chains, as they make multiple API calls to complete a task, the costs per task goes up. So companies are going to need to think about AI spending the way they think about staffing, not per query, per outcome.
SPEAKER_01Per outcome. I think that's actually the cleanest summary of where we are. The entire mental model is shifting from I use AI as a tool to I manage AI as a worker. And workers have costs. Workers make mistakes. Workers need clear instructions and defined limits. But workers also, when managed well, get things done while you're sleeping. That's the trade. And I think the people who understand that trade early are the ones who are going to look back in five years and feel like they had a head start.
SPEAKER_00Yeah, and I'd add one last thing, which is don't wait until it's perfect. The hallucinations are still there, the compound systems still fail in interesting ways, the hardware shortage is genuinely making this stuff more expensive at the infrastructure level. None of that means wait. Because the people who figured out how to use email in 1995, even when it was clunky and unreliable, those people had a decade of instinct by the time everyone else caught up. We're at that moment again. The tools are imperfect, use them anyway. Learn how they fail, build the instinct now, because the next version will be better. It always is.
SPEAKER_01Nigel, this was exactly the kind of conversation I wanted to have on this. Grounded, specific, no hype. The point about managing AI like a worker, not a tool, I think that's going to stick with people. Thank you for being here. And to everyone listening, if this episode made you think differently about what's sitting inside your laptop right now, do one thing. Share it with someone who's been dismissing this stuff as science fiction, because it isn't anymore. And if you've been listening for a while and you haven't left a review yet, it takes 30 seconds and it genuinely helps more people find the show. Subscribe if you haven't, tell a friend if you have. The prompt box is dead. The work has just begun.