← AI News & Strategy Daily with Nate B. Jones
AI News & Strategy Daily with Nate B. Jones
Which AI Model to Use for Any Task Without Overpaying
00:14:16
Loading summary
Transcript1 lines
- [00:00]
A
Coinbase Cursor Lindy Lots of companies are switching to open source models. This video is not about them. This video is for you. This video helps you sort through the noise and pick a model in a world that has exploded with model choice just in the last couple of weeks. Because since Fable was banned. So on Wednesday, July 1, Fable 5 came back online. Yay. But here's what isn't coming back. The month where everyone assumed the model you build on will still be there tomorrow. Because for 18 days a lot of companies found out it might not be. And the ones who could shrug it off were the ones who never tied their work to a single model in the first place because they owned the harness, they routed somewhere else and they kept moving. And that's what a lot of companies did. This video isn't about whether Fable's back. Right. It's about making sure it never matters that it was gone at all. And I am seeing a trap that a lot of people are falling into where they confuse the act of picking a model with getting real work done. So. So in this video, I'm going to walk you through how I think about picking a model. I'm going to walk you through why. I'm going to walk you through the pieces you need to think about along the way, and I'm going to give you specific tips for specific models. Yes. GLM 5.2. Yes. Quinn. Yes. Kimmy. Yes. Chatgpt. Yes. Claude. And how I think about putting them together. And I'm going to give you recommendations to simplify, like what it looks like to actually simplify in different roles, whether you're a team lead, whether you're an individual inside a company, whether you're your own business driver and owner, whether you are a developer, whatever that role may be for you. How you start to think about picking models in an age when open source models are becoming more significant, and how to do that in a way that actually gets you focused on the work and not distracted. All right, let's jump into it. Given where we are in the model landscape today, how do you make sure that picking a model is not your second job and instead you're focused on the work? Here's the frame I want you to use. A daily driver needs to be good across a wide range of use cases, because it is the model you reach for before the task is clean. A cheap workhorse earns its place when the job is familiar and repeatable and easy to understand. And so I would suggest that the useful answer begins with the work in front of you. Before you pick the model, think about what you want the model to do. Those questions matter more than the name on the model card, right? Ask whether you need it to act as a coding agent, ask whether you need it to make PowerPoints or spreadsheets, whatever it is, and then get to the model. And this is where GLM 5.2 is interesting because GLM 5.2 is good at a lot of center of distribution work. So stuff that I think of as normal and not too heavy a normal PowerPoint, a landing page draft, a meeting summary, a lot of code work that has familiar shapes to it, where you can say here's the file, just work on it. That's actually a massive category of work. Most people spend most of their workday producing familiar artifacts under time pressure and so having the ability to make a table or a client note or support reply or whatever very quickly with a cheap, strong model is great. And I do want to be clear about the non coding part. A lot of the current GLM conversation is coding heavy because coding is easy to benchmark, the savings are visible, et cetera. But center of distribution work includes Web pages and PowerPoints and memos and CRM cleanups and routine synthesis and all the normal business artifacts people make every day. If the shape is familiar and reviewing it is easy. GLM belongs in the conversation. Frontier models need to earn their keep elsewhere. I still want Claude and ChatGPT's newest models when the shape of work is not obvious yet. If I'm trying to find the angle in a messy sort of bundle of sources. If I'm trying to like figure out where my taste and judgment matter on a tough problem. If I have to decide where to push the model into a new piece of work I've not done before, I don't want to optimize for cost. I want to get it right. So if you're just trying to choose a daily driver, that's the conversation I would have. Your daily driver should be the model you trust for messy mixed human work. Now in my previous video I shared I'm still using codecs for that because I find the harness is so easy to use that the intelligence inside it is not the most significant piece. But there are a lot of other harnesses out there and that's not the only one that's effective. And you should be looking at other harnesses that work for you, whether they're open source harnesses, whether they are closed source harnesses like Claude or Claude code. Whatever harness you choose, make sure that you are Deliberately picking something that feels like it helps you do the work and doesn't distract you. If costs are a distraction, think about a harness for GLM 5.2 that works for you. And yes, they released one called Z AI. This should be very helpful. This is the difference between GLM style work and what I'm calling a fable style problem. In a fable type ask, the hard part is not just producing a familiar artifact cheaply. The hard part is understanding what a new capability means. Across video, across physics, across character, intent, legal exposure, business strategy. You're asking the model to help you discover the shape of a new type of product problem. And this is where I want the broadest, strongest, weirdest generalization I can get, plus a harness that can keep that context together. So if you're just trying to choose a daily driver as an individual person, that's the. That's the background matrix I would have. That's the way I would think about it. Look at where you need to do messy, complex work and then trust that to a frontier model maker. And look at where you need to do familiar work and consider whether something like GLM 5.2 is going to save you money at that point. Because it might. Once you have a daily driver, take it for a drive test it put the inputs in that you care about. Maybe it's spreadsheets, maybe it's PowerPoint stuff, maybe it's docs, maybe it's PDFs, maybe it's code. Whatever it is, make sure that you don't validate it as your daily driver till you actually run it across the tasks that you care about. And that's where you're going to find out just how complex your asks actually are. I think we humans are sometimes better at estimating complexity once we do the task than we are in advance. So make sure you test it now. If you're at work, you may not have the freedom to choose a model. If you're using AI inside a company, the first filter is going to be permission. If you have a choice of more than one model inside a company, you should be using the same thinking process here to pick what you would use. Whether that's with Microsoft Copilot or Claude Teams or Gemini or ChatGPT Enterprise. Regardless, you want to be taking your actual work, reliably testing it across these different models, and then picking a daily driver as you find actual utility matches your ask. And where it doesn't. If you're inside a company, that should be your cue to say, hey, I need a more powerful model. I need to work with the IT department to level up the models because look at the result that I'm getting here. This is not my prompting issue. This is actually the model unable to get the work done. Now, if you run a small business or you have a small team, you're going to have a little bit more flexibility and you should be thinking a ton about the level of energy you put into work in order to get to outputs that are really, really high quality for your customers. Because really your ability to meet customer needs and scale is a function of your team's ability to work with models efficiently. So ask yourself, do you want a 20 model routing system or do you want to pick the five recurring artifacts that that are most critical to customers and figure out how to draw a straight line to value from that into the process your team uses to make sure it's as clean and simple as possible to get that artifact generated. With AI, that's how I think about it, right? You think about what is the simplest way to get the client brief done. What is the simplest way to get this code in front of the client? What is the cleanest way to present this PowerPoint, whatever your business might be, and you should have the flexibility if you're a small business owner to pick the model array that works for you without overwhelming your team. And this is where specialists can start to play, right? If you're constantly making ads or thumbnails or mockups, it's where names like Flux and Z Image and Grok Image becomes relevant, right? You don't need to know those names on day one. You just need to know the job. I need images, I need references, I need to understand local control and cheap APIs so I don't overspend on tokens because this is such a heavy part of my business. Then you walk into the model name from there. If you work with video, the same logic applies. A local video tool like LTX could matter if you want local iterations. A high end API video model like See can matter when the quality bar is really high. A cheap API video route like Grok can matter when the point is fast and disposable clips. So the better question again is to ask what kind of video work you're doing and what's the fastest and simplest way to customer value. The same specialist logic applies to live information. If the job depends on current web information, try Grok for like Live X posts. Now X did recently release an API that anybody can plug into with any other AI model, but that's a brand new as of the last couple days thing and we'll have to see how good it is is. But largely the point here is that you need to see the customer value you're looking to create. Look at how a model needs to process files or live web information to create that value and ask yourself what is the simplest possible way to get to customer value here? Is it with a daily driver? Is it a normal artifact? Is it something GLM 5.2 can produce with its own native harness? Is it something that's more complex that I need to wrestle with so that I need to have codecs involved? Is this something where the real blocker for me is that my team is not AI fluent enough and so I can't use these specialist models like Seed Dance for video because the team isn't ready and I have to just simplify down my stack to just Claude for example, because if I expand beyond that the team is just going to be overwhelmed. These are the kinds of challenges team leaders need to be thinking about and talking about is honest trade offs along the way. And this is where these company migration stories become useful. Lindy is moving serious traffic to Deepseek because it's finding it can save money. Cursor has been building on Kimi and moved to a pre trained model again for saving money. Coinbase is increasing their usage of tokens while decreasing costs through smart routing to open source routers like GLM and Kimi. Shopify and Airbnb are going with quen style routing because they find that their queries route there effectively. Microsoft is testing into a deep SEQ architecture. So these big companies are finding that they need to 1 not just jump into a one size fits all model and 2 think about what is the straightest path to value from their application given a particular kind of intelligence. Don't feel like you're afraid of Claude, don't be afraid of codecs. Those are fantastic models. We need to get to a world where we have a much higher dimensional conversation around model trade offs. And the point of this video is to equip you to make that trade off yourself. I can't tell you all. Always use QENT, always use Kimi, always use GLM, always use Cloud. Opus 4.8 there's not one simple answer here. What I can tell you is that the route to get the model right runs through your job right what you need the model to do and it runs through your ability to get work into and out of that model efficiently. And that's true whether you're an individual or whether you're a team leader or whether you're a company. I'm just saying use the model that works for you. Use a model that aligns to what you need done just to take away if it's a middle of the distribution task, which means if it's a fairly simple task for which there are a lot of examples online that the model would have seen before, GLM 5.2 is going to do a great job in some cases better than claude. If it is a weird non standard task that requires generalized intelligence, there is no substitute for a frontier model. Right now you cannot substitute for Claude or for ChatGPT 5.5 or 5.6 if you can get a hold of it when you are tackling a tough generalized task. And the other complexifying factor as I keep calling out, is that you need to be thinking how to get work in and out of this intelligence. That is part of why I've talked about Gemini less in the last few weeks. Getting work into Gemini and out is something not only I but many others have called out as unnecessarily difficult. The Gemini harness is not strong for getting work out, even though the Gemini intelligence is strong. Gemini is a solid model without a great harness and that's something that I'm seeing a ton of movement on from Chinese open source models and expect more in short order. They can see the success that Claude code is having, that Codex is having. There's a reason they launched GLM 5.2 with Z AI as a harness. Expect more harness work from open source models in the next couple of months, even as close source model makers continue to improve their harnesses. So the takeaway is simple. I'm going to give you a few rules of the road here. Do not just copy what someone else is doing, and that includes me. Don't just copy people. Make sure that you ask yourself how hard the work is, not just how much work you need done. And that's what a lot of the conversation around GLM is about. Is is it hard work or is it work that you just need done? Make sure number three that you know how to tell if it's good or not. Yes, this is evals, but there's also sniff tests just looking at and saying is it actually acceptable? Make sure that whatever model is available to you, whether you work inside a company, work inside a team, whether you own a small business, make sure that the model choice itself is not work. And that's why I keep emphasizing the simplifier of figuring out what the end value is to your user and finding the simplest way to make that happen and then just do that. And that leads me to number five, don't pick too many models. I had given you the names of like half a dozen or more models over the course of this video but it is likely that you don't need that many to get your work done. And to make that easier for you, I would like you to put that in the comments. Put in the comments. This is what I'm doing typically. This is the model I'm using and let's get this conversation going as a community. Let's talk amongst ourselves about what the best model is and I'll pick out some of those comments and use them to call out specific model applications specifically in a future video. All right, have fun. Subscribe for more awesome model picking takes. I hope that this was useful to you and helped cut through the noise. I know there's a ton of noise out there.