← AI News & Strategy Daily with Nate B. Jones
AI News & Strategy Daily with Nate B. Jones
Model Routing Is Table Stakes. Here's the Real AI Edge
00:15:39
Loading summary
Transcript1 lines
- [00:00]
A
Here's something strange that everyone using AI is noticing right around the same time. The tools keep getting better, the prices keep dropping. You can produce more than ever. And yet somehow everything's starting to feel the same. Your output, your competitors output half of what's in your feed. It all looks really similar, right? Better tools, samey results. That's not a coincidence, and it's actually not a tooling problem. Something specific happened. AI made doing things cheap. And whenever execution gets cheap, the value doesn't disappear. It just moves. I'm going to show you exactly where it moved with a true story about an engineer. 40 bucks and a task that no to do list on the planet had captured. By the end of this video, you'll know which side of that move you're on and how to switch. Okay, let's set the scene. Mitchell Hashimoto, co founder of Hashicorp, creator of the Ghosty terminal, one of the most respected, respected engineers working out there today. When Fable 5 launched the current Frontier model, the one that everyone complains about the price on, he spent days testing it against cheaper models on ordinary work. Implement this feature or build this thing. The stuff on everybody's list. And honestly, all three models that he tested produced equally acceptable output. The Budget model cost under a buck and finished in minutes. GPT 5.5 about a buck 50. And Fable 5 cost 4.40 minutes and $9. Right? Same work, same quality, 9x, more expensive. If you stop there, Fable 5 looks like a ripoff, right? And a lot of people are stopping there because that result launched a thousand hot takes about how you should route everything to cheat models. And you know what? You should write a lot of execution in sheet models. I think that makes sense. But here's where I want to call something out that no one seems to be saying enough. Because part of my job is to look out ahead when everyone in AI is saying the same thing at the same time. And right now, everyone is saying route to cheaper models. That's exactly when I start asking a different question. Where does value move? Once we've all figured this out, routing is real. It's also about to be table stakes. Everyone's going to have it. Execution is commoditizing real fast. So the question that matters isn't the one the whole industry is answering right now. It's the one hiding in the second half of Hashimoto's experiment. Because Hashimoto ran one more test, he handed the Frontier model a problem that the cheap ones couldn't touch at all. Optimizing a gnarly piece of systems code that he'd written himself. It took two hours, it cost 40 bucks, and it reached a level of performance that Hashimoto, again one of the best engineers in the world at exactly this, says he couldn't have hit on his own. Now this is the question I want you to wrestle with in this video. Who assigned that task? It wasn't on a backlog, it wasn't in a sprint. No PM prioritized it. It didn't exist as a task until one person, an expert, suspected something new had become possible and spent money trying to find out. No individual process generates this task. No best practices guide is going to contain a task like this. It comes from somewhere else. AI can only do work that someone has imagined. These cheap coding tools, they execute, but they don't decide what's worth executing. Which means the ceiling on what AI is worth to you was never the model or the price or the prompt pack or whatever you're depending on in your harness. It's the size of your list of things you know how to ask for. It's your imagination. And let me be careful here because this is not a video about how execution doesn't matter. Execution matters enormously. It is the thing that gets multiplied by imagination. Hashimoto's 40 buck job still needed 2 hours of world class model execution to become real. The point is about what sets that multiplier. It's a smart move and you should keep making that move aggressively. It's a layer where you can control cost, you can keep driving them down, and that control is all yours. That is why Hashimoto ran an under a dollar GLN 5.2 run in his initial test. The mistake would be treating that layer as the whole strategy. The leverage comes from what you put on top of a strong execution strategy. You need a targeted strategy, surgical application of frontier models to the questions that change what the execution layer is even building. Right? Cheap, open execution is a great engine. Frontier imagination is where you steer the plane. You need both, right? They're not in competition. The cheaper your execution layer gets, the more commoditized it gets and therefore the more valuable every frontier posed question becomes. And that explains the strange thing that that I called out at the top of this video. Look at the first half of Hashimoto's experiment again. Cheap model ties, expensive model. That's not a fact about the models, it's a fact about the tasks implement. This feature is work everyone already knows how to ask for. And the work everyone knows how to ask for is exactly where the models have converged. And that's why a one buck model ties a nine buck model. Now let's zoom the lens out. From models to people, the prompts we use are often shared. The Playbooks are often public. Many of us are following the same productivity channels. So when you and a million other people run the same tasks across the same tools, of course the results are going to converge. And that's fine if you want similar quality at a cheap execution price. But AI is not responsible for making our work outputs generic. Instead, it just reveals that differentiation is hard and that it's a human task. Now, we've watched this exact dynamic decide a market before. BlackBerry and Apple both made smartphones. And BlackBerry executed on those smartphones brilliantly. It had the best keyboard in the world, the best email, the best security, and the market leadership to prove it. But what killed BlackBerry was not an execution failure. It was executing superbly inside a category everyone had already imagined. While Apple and Steve imagined a different answer to what a phone even was and then executed on that. So same industry comparable execution muscle and imagination sold set the multiplier. One company's execution became worth a hundred times the others. That's the divergence. And it starts exactly where the known list of execution tasks ends. And that's where the models still separate. And much more importantly, it's where people will separate. The $40 job had no competition. Nobody else was doing what Hashimoto was doing because nobody else had thought of it. So here's the test, and it works for a person as well as a company. Has your task list changed in the last 12 months? In the last six? In the last three? Has what you have asked AI to do shifted? Or are you doing your old list faster and cheaper and calling that AI transformation? Because if it's the old list, nothing is wrong with your tools. You have an imagination shortage and you're going to be spending a lot on optimizing for execution in the commoditization market. But the good news is that imagination is not what you think is, is because the word imagination gets overused. It sounds like a gift, something that artists have and analysts don't. That's not what was operating in the $40 story, and it's not what's scarce. Hashimoto could pose that question because he has hundreds or thousands of hours inside these models. He knows where the capability line has moved. Not from a benchmark chart, but from instinct, from touch you can't imagine with capabilities you haven't touched. Nobody imagines a use for tool they've read a summary of. Right. And this is where most of us quietly sabotage ourselves. We interact with AI and we have cost savings in the back of our heads or we have a defined task list that we expect AI to do. We wonder if AI can help us go faster or cheaper. It's like pointing the telescope at the ground. It's pointing it at work you already have. The questions that find new territory to explore are different. You're asking, what can this do that I've never been able to even ask before? And I'm going to give you an example from Fable 5 here. This is a real example. I found it on X. I love this one. You can get Fable 5 to use Google Maps and to map all of the porches in a geographic area that are unshaded, that get all day daylight, in an area where average temperatures in the summer cross a certain degree threshold. When you do that, you can then go through, find those properties, get a three dimensional model of the structure on Google Maps, and then mail those individuals a custom card offering them a covered porch with specific data on their porch, their situation, a specific visual of how their porch actually would look when built. All, all of that powered by Fable 5. And if you're wondering, can I get a cheaper model to execute some of that 100% once the idea is prototyped through with Fable 5, you can, but combining all of those tasks together, running them through Blender, through other tools, managing the whole thing, and critically being able to do both the spatial and logical reasoning and then translating it all into a business flow with an address table, that's an example of something that is business oriented as a problem. It's not like we're, you know, not like we're solving pie in the sky problems here, but it's also a really hard problem that wouldn't have been possible before. That kind of marketing is something that we are able to do now that we never would have been able to do before and that earlier versions of model just can't get you there. Now, once you imagine it, once you see it, once you put it into action, you can start to get a cheaper pipeline put together and you can start to move it back from the frontier into something that's simpler where, like for example, the merge of the image into the mailer, that doesn't take a frontier model, right? But the idea of putting that together and getting the first few houses done and getting into analysis of where the sun is and where shade is, that's frontier model stuff. That's Stuff that takes imagination, not just imagination to prompt the model. The hard part is not the prompt there. It's imagination to say a new kind of marketing is possible. I can be hyper targeted, hyper specific and hyper relevant to my customers. If I can use this model to do analysis, that would never have been possible otherwise. I love that example because frankly, I could use a porch like that. I love that example because it calls out for us why frontier models matter and that they matter, especially in places where we haven't even seen value yet. So your personal practice is almost embarrassingly simple and it maps right onto a two layer stack for daily execution. Absolutely. Use cheap models, optimize away, et cetera. And that works at company level and it works at individual level. But where do your scouting hours go? Are you taking scouting seriously? Are you thinking about where your imagination time is going? And keep in mind the two examples I've given. Hashimoto's example. This example around using sun analysis to get complicated marketing done. These are not artist type imagination tasks. They're technical. They're business imagination tasks where we look at a particular problem in a new way because we have fingertip awareness of what new models are capable. When factories electrified, the technology worked on day one, the productivity payoff took decades because factories kept their Steam era layout. Every machine crowded around one central drive shaft and just bolted an electric motor where the steam engine used to be. It was the same building, it was a new power source and there was barely any gain. The payoff arrived when a new generation of managers exercised technical imagination and redesigned the factory around what cheap distributed motors make possible. So the unit of change wasn't the motor, it was the building. AI is the same kind of tech and companies are making the same move. They're bolting it onto the old layout. They're running the existing task list through cheaper models and reporting on the savings. Now the savings are real and they're available to every competitor with the same insights. They're table stakes. Here's what redesigning the building looks like when it works. Stripe reports that they ran a migration across 50 million lines of code in one day. Work estimated at 2 plus months for a team. The impressive number there, it's not a day. The important number is the years that Stripe spent building test coverage that could verify that many changes. Review systems that could move at that speed, have people who knew how to drive the model in a complicated task like that. The model deleted two months worth of typed code. Right? Two man months of typing code the building to accommodate that change. The Structures in terms of team, in terms of review cycles to verify quality that had been redesigned in advance. So you could point the same model at a company that hasn't done that work in their code base and you would not get a one day migration. You would get 50 million lines of changes that nobody could approve. Stripe built the infrastructure first and then harvested the value with frontier models and technical imagination and leaders. One warning, because I know the shortcut you're thinking about. You can't hire your way out of this with one imaginative person. That 40 buck job needed imagination, deep context, and permission to ask in one head, Hashimoto's head. And that's true as far as it goes. But your new AI visionary that you hire, if you want them to solve this kind of problem, they have all of the imagination to do that. Sure. But they have none of your context. So you need to ask yourself, do the people with the context in your company, do the people who could exercise technical imagination have the permission to to do that? Do they have the tools to do that? Imagination only fires when it sits next to context. And your context is spread across everyone who actually does the work. So the job isn't hiring imagination, it's manufacturing it. It's putting the people who have context in your systems in contact with capable models and giving them permission to make bets. And you know the test from earlier scales right up. Who on your team is allowed to pose a 40 or $400 question to a model today without asking anyone? If the answer is nobody or just a tiny number of people, that's an imagination constraint. It was never about the price of the model. One last piece of evidence that imagination is the asset. And it comes from the strangest AI story of the year. A couple of weeks ago, Fable 5 shipped on a Tuesday. It was gone by a Friday. Now it's back. Everyone's carefully optimized their model setup since, right? But look at what the blackout could not take away. The people who had spent those first 72 hours dreaming of what was possible. The questions they posed were kept, the workflows they'd redesigned were kept. And the model came back into a price war. But the people who had imagination to use the models went right back to what they were doing imagining with Fable 5. And I've seen that all over my timeline this last few days as Fable 5 has been back. This is our answer to the question of sameness. Remember how I started this video saying everything is looking the same? It was never the tools. The tools did their job. They made execution cheap and you need to optimize for execution. Nothing in my video is saying don't do that. You got to do it. You also have to have an answer for where your 10x multiplier comes from, where your iPhone moment comes from, where Hashimoto's $40 question comes from. You want to spend the frontier where it multiplies, where you have real context, real bets on the line, and real questions that exercise technical imagination around what's possible with a larger model. There is no substitute for that. Now, if you want to dive deeper, I have a deeper written version of this argument with sources, the full strike breakdown, and counter arguments all on substack. Check it out and I'll see you next time.