I have three or four posts floating around in my head, and I went to get them out.
This is based on my experience using AI at Ramp and in my personal life, and by AI, I really mean LLM-powered agents like ChatGPT/Codex or Claude Code.
Today, my setup is ChatGPT Desktop on a MacBook, with permissions wide open and lots of access.
I subscribe to the $200/month OpenAI “Pro” plan but struggle to use even half of the weekly limits. We are in the peak—or a little past the peak—subsidy era, with OpenAI edging out Anthropic over the last couple of months.
I use GPT-5.6 Sol on High effort almost universally.
And with that out of the way, on to the takes!
Mental Model for AI, Computer, Apps
My AI setup is literally: Me → Computer → App → AI.
But my mental model is: Me → AI → Computer.
I don’t see myself as “using AI through my computer.” I see myself as having AI use Computer on my behalf. And by capital-C “Computer,” I mean my literal device and everything beyond it.
The irony of “Applied AI” as a practice is that, outside software engineering, the primary “application of AI” is actually to “Apply Computer.”
Another route to the intuition here is to consider what it means to “use a computer” today and what it will mean five years from now.
It’s tough to undo decades of experience manually operating computers and open yourself to this framing. But it’s necessary, and on the bright side, knowing intimately what a computer might be used for is very helpful when it comes to bossing AI around to do the actual operating.
Diligence and Semantics
The line I’m most proud of in the blog post about my Claude Code Plays Rollercoaster Tycoon project with Ramp Labs (modified):
The limiting factors for general-purpose agents are the legibility of their environment and the strength of their interfaces.
In general, I prefer to think of agents as automating diligence, rather than intelligence.
A lot of applications of AI could be described as opportunities for diligence arbitrage: scenarios where you (or others) don’t have the time, interest or expertise to perform diligence.
Anything from credit card point programs to lease renewals or account security reviews.
And I love that this is a semantic point. Semantics matter here both for forming a useful model of this new technology and for the literal invocation of it.
Personal Software Is Underrated
I’m immediately undermining my previous point about semantics by willfully misappropriating the term “personal software” to mean “in-house business software.”
Self-built software is underestimated right now.
At Ramp, the distribution of outcomes between vendor products and internal tools looks like two distinct clusters. There are successes and failures in both clusters, but the internal tools include positive outcomes multiple deviations beyond anything in the vendor cluster.
Blowout success has happened like 5 times now, and it’s less surprising each time. I’m getting used to the idea that there’s something unintuitive or ineffable going on here. My best guesses are:
First, the speed of initial build-out, iteration and maintenance is now just faster than procurement. It’s literally easier to create something from scratch than to spend even 2 weeks in a sales process. And the feature requests land in minutes instead of quarters. It’s almost like “turning problems into code for coding agents” but the problem isn’t just the software; it’s also the vendor relationship.
Second, there’s just a massive failure of imagination. After 30 years swimming in mass-market software, our vision and ambition for what we want out of our tools has withered. Our minds are trapped in an allegorical cave like the one from the Allegory of the Cave.
Third, the whole is more than the sum of its parts. It’s like furniture built into a house. You simply can’t beat it with a standalone piece, and you certainly can’t beat it with some mass-market flat-pack IKEA assembly.
I’m drawn to this furniture analogy because that’s an area where we know better solutions exist, and we have preferences for them, but we don’t have the ability to manifest those ideas. Imagine if we all had at our disposal adept humanoid handymen that cost $50/month + raw materials. Houses would be built very differently.
It’s hard to find other ways to express “customization is great.”
And of course the wide-open potential here will result in monstrosities, wasted efforts, and visions that should have never materialized. I’m not trying to defend any possible use of a computer. I’m just saying there’s a lot here.
I think we’ll see more frameworks, SDKs, and Platform-as-a-Service-type offerings. The biggest threat to traditional SaaS (my employer included) is the infrastructure pieces 1-2 layers below it in the ‘stack’. These pieces enable the granular assembly of custom solutions—as Home Depot is to IKEA. I also think consultancies could build and operate bespoke software at larger scales. Maintenance of internal tools is increasingly a closed loop around feedback channels.
Open Problems That Are Bothering Me
Collaborating with AI is janky. There are a couple of apps out there where you can have an AI do a bunch of work and then point-and-click edit the results, but each one is totally bespoke. This obviously has to come a long way from where it is now. This might be the main promise of “generative UI.”
Browsers. I use AI, and the AI uses my computer. Unfortunately, a lot of “using my computer” is locked up in the browser. ChatGPT Desktop and Claude Desktop both ship with built-in browsers and can also use Chrome windows on my machine, but the experience looks clunky, and it kind of bums me out.
The more innovative SaaS companies are publishing MCPs and CLIs. I wouldn’t consider procuring a vendor that lacked this kind of access, but in consumer-land there’s close to nothing here.
Software Engineers Owning Prompts and Context. Literally the least prose-capable people in business are responsible for nearly all the natural language that goes into the language models. I could rant for hours about this. Ultimately, I think it comes down to the adage:
It is difficult to get someone to understand something, when their salary depends on their not understanding it.
I don’t want to go to war with my peers, so I’m going to call this one a leadership failure.
Practically, I’d love to see:
A thoughtful IDE for prose (prompts and context).
I’m writing this in Substack, and it has red squiggles (including under the word “Substack,” so clearly not a ton of thought has gone into it), but I would love some AI-assisted copy-editing flags and diff recommendations. C’mon, how hard is this?
A thoughtful agent tracing/observability tool (in the category of Braintrust and Langfuse) that presents agent traces as a natural-language phenomenon rather than blindly copying Datadog’s design.
A business culture of prose review similar to what existed for code review.
Hiring English majors and journalists—people with prose skills for prose work.
Wresting AI influence away from software-engineer “shape rotators.”
Printing out agent inputs and marking them up with colored pens. Then getting this back into the loop with AI Labs’ book-ripper 9000.
Basically where I'm at. Enjoying your posts!!