Remix: Using Powerful AI
We've completed the Kessel Run in 12 parsecs.
AI for SMBs is now one year old! It’s wild to think how much has changed since we started writing. Our original thesis was that small and medium sized businesses have a fundamental advantage when it comes to GenAI, and we hope that this first year of posts has equipped you with practical tips, mindsets, and strategies to help your business thrive.
Last month’s “partner chat” format was our best performing post in the history of this blog! In the spirit of “give the people what they want,” we’re back with another chat. This time, a remix of our post from last November about Using Powerful AI.
📰 What’s Happening in GenAI
Working with Cowork
Lately, we’ve been describing Claude’s Cowork as “the AI you didn’t know you wanted.” For so many of our clients, using this new harness has been a huge unlock. But it is very different from just chatting with GenAI. If you’re new to Cowork, we recommend this guide on how to get set-up.
In the News
Our friend Stu @ Just Curious interviewed us, you can catch a clip on LinkedIn here or listen to the whole thing on Spotify. Justin also co-authored a piece for Inc. with his old IDEO colleagues David Schonthal (Northwestern U entrepreneurship professor) and Katherine Otway (CMO at MIT’s The Engine) about the stages of AI adoption for SMBs. Enjoy!
Agent Racers
Remix is closely following the folks who are trying to build “dark software factories” in which no humans ever look at a line of code. It’s a fascinating community that reminds us of those wild experimenters in the early days of automobiles who tried to set the land speed record. Their innovations are in all of the cars that we drive today. Enjoy.
Remix: Using Powerful AI
TL;DR
AI got more powerful. The question is no longer which plan to buy, it’s how you actually work with these tools. Stop thinking of AI as a smart intern you manage; start thinking of it like a product manager you give a goal to and trust to figure out the rest.
In this post, Jason and Justin break down what that looks like in practice, why Cowork and Claude are dominating client conversations right now, and what separates true agentic AI from the workflow automation most businesses are actually running. Picking tools? Don’t pick one; run a bake-off. Directors and above should have access to everything (~$200/month, obvious ROI). And if you’re not on paid plans, you’re missing the biggest capability jumps in AI history — and you don’t even know it.
Justin: Last November we wrote about using powerful AI. That post was really about access. Which plans to buy, what context windows are, how to pick the right tier. Useful stuff, but the landscape has changed a lot since then. Today I want to talk about what it actually means to use powerful AI. Not the plans. The experience. So maybe we start there. What makes AI powerful to you right now?
Jason: When I can one-shot something. Whether it’s an image, a video, a workflow, a research query. And if it’s not a one-shot, how few iterations does it take to get where I need to go? That’s how I measure it. How quickly is the AI accomplishing the task I threw at it?
Justin: I’m with you on that. There are definitely moments where I want to offload work to a really smart intern that needs almost no management. Just get the thing done. But there’s another mode I’ve been living in more and more. Sometimes I want the AI to go back and forth with me. To elicit things out of me that I wouldn’t have been able to articulate on my own. Those feel like very different ways of working, and they’re both getting better fast.
Jason: The other thing that’s changed is when the AI asks me something that shows it knows me better than I expected. Like talking to someone with a thousand years of memory (as opposed to the 2,000 year-old man?). It’ll say, “Because you asked for that data point, I found the latest version and here’s what I think about it.” It saved me two or three prompts because it somehow knew what I was after. Those moments are happening more and more since the new year.
Justin: The formal AI term for what we’re talking about is alignment. The model doing what the human actually wants. And one of my favorite new benchmarks is BS Bench. What I love about it is that it measures the inverse of what most benchmarks test. It asks: when I’m trying to do something the AI shouldn’t help me with, will it challenge me? An AI model that just agrees with everything I say is not aligned to what I want. I want something that pushes back.
Jason: Doesn’t that humanize AI in a strange way? If you’re working with a teammate or a peer, hopefully you want to be challenged. With AI, you get to be challenged with the privacy of sitting in front of your computer. You get to save face and produce better work.
Justin: Exactly. The Socratic method works. Debating and going back and forth is a great way to get good outcomes. And AI doesn’t care about the power dynamic.
How We Each Use It
Jason: The tool that’s changed everything for me is Cowork. It does more than a chat session would. You can pull in projects, bring context over. I’m not rushing to get back in front of a terminal window, and Cowork is powerful enough that I don’t have to…yet.
Justin: I use Cowork a lot too. I describe it as the AI you’ve been waiting for. Capable over long time horizons. Persistent. Relentless. I’ve yet to meet a project it couldn’t handle. That was not the case with Chat, where I’d regularly hit walls. But I also spend a lot of time in the command line, because I like the constraint. You and I reach similar outcomes through different paths, and that says something about where these tools are headed. It’s not just about capability. It’s about the fit between the tool and how you think.
What We’re Seeing With Clients
Jason: We’re hearing more about Cowork than any other tool right now. And more about Claude than any other model.
Justin: My favorite pattern is when a client gets excited, has a couple of breakthrough moments, and then three days later I get the late-night text: “I hit my rate limits. What should I do?” The fact that people are burning through their monthly limits in days tells you how powerful this is. We should note that Microsoft just announced their own Cowork variant with Anthropic. We haven’t tried it yet, but we’re intrigued to see how this category develops.
What’s Next
Jason: I’m thinking about collaboration. These tools were built for individuals. Claude has been deliberate about the enterprise and small business market. They’ve recently introduced ways for people to collaborate that weren’t there before. And I think Anthropic could bring something game-changing here. Microsoft and Google already have collaboration baked into their core tech stacks. Small businesses should lean into the Team collaboration plans and see what works for them.
Justin: Agreed. And we’ve been experimenting with giving tasks to an AI agent the same way you’d give a task to a teammate. Figuring out how AI becomes a collaboration tool across teams, rather than just within a single person’s workflow, is the next frontier. We’re in the early days.
Agentic-ish vs. Actually Agentic
Justin: Outside of coding, if I reflect on 2025 the primary way people started pushing the edges was building workflows with AI slotted into specific spots. Tools like n8n, Make, Clay, Re:Live. For the right business process, those can be very valuable. But I’ve started calling all of them “agentic-ish.”
The agents we’ve been building at Remix are different. There’s no linear workflow. It’s an AI model in a loop with a set of instructions and some context. I don’t tell an agent,”step one, two, three.” I give it the instruction manual and let it loose. Cowork supports the same approach. It’s powerful not because you give it a task list, but because it generates its own.
Jason: Six months ago you couldn’t open LinkedIn without someone posting a prompting guide. Now, to your point, it’s treating Claude like a product manager, not an intern. You give it the goal and trust it will figure out how to get you there. That’s a significant shift in just a few months.
Don’t Pick One. Run a Bake-Off.
Justin: We’ve been talking a lot about Claude in this conversation. We’re fans of the current tools. But give us a few months and we might be on to something else. The leading position keeps changing. So for the businesses who bought an annual subscription or invested heavily in training their teams on one tool, is our recommendation to rip everything out and switch? No.
Jason: Assuming you can set aside some budget, some of the most well-spent R&D dollars right now are arbitraging at least two if not three big tools against each other. For example if you’re working on a contract, have one model proofread and even argue against another. Massively effective. The more little tests like that you can run, the better POV you’ll have on where each model excels versus what you need. And it’s too early to declare a winner. The Law of Duality is alive and well (Trout/Ries) – there’s always going to be a Coke and a Pepsi and a third challenger for years to come.
Justin: Agreed. This isn’t going to consolidate. So here’s how I think about it: directors and above should get access to everything. It’s so cheap. You’re not going to spend more than about two hundred bucks a month. The benefit relative to the expense is a no-brainer. Upper middle management, people leading teams, should probably have two or three subscriptions. If you’re a Microsoft shop, maybe Gemini matters less. If you do a lot of graphics and you’re a Microsoft shop, maybe Gemini matters more.
Individual contributors are a different story. Not every person needs every model. But everyone on a team using the same tools so you can manage and train them together? That makes a lot of sense.
Jason: So it’s tiered access based on role and need.
Justin: Yes. And the key point is this: if you’re not on the paid plans, you are kidding yourself that you’re getting the best version. Based on publicly available data, the capability jump from GPT-3.5 to GPT-4, the one that took the world by storm, barely registers compared to the jump from Opus 4.5 to 4.6. That happened in about four months. Most people haven’t experienced that difference because they’re not using those models. So they can’t see the trajectory.
Using powerful AI means you actually have to use it to have an opinion. Get the paid plans. Use multiple tools. Run a bake-off. The pace of improvement is accelerating, and the only way to stay ahead of it is to keep your options open.
✨ ✌🏻 ✨
Questions? Email us at info@remixpartners.ai - we read every message.