The Malleable Computer
This week Pete and Andy explore how working with AI is changing as agents take on longer-running jobs, software becomes cheaper to build, and recurring work moves beyond a singl...
Latest Snacks from Episode 70
AI Is a Capital Good
AI may look like a consumer product because people encounter it through models, apps and chat interfaces. Its greater power, however, is as a capital good that accelerates what people can already do. Someone with a sound idea or a strong business can use that capability to move faster, rather than treating access to AI as the source of value itself.
This distinction matters more as intelligence becomes broadly available. If everyone can access similar capabilities, the advantage does not necessarily remain with the model or the app that provides them. The same intelligence can be applied to different niches, geographies and needs, producing different solutions in each.
The resulting value can therefore spread outwards instead of accumulating only at the access point. AI helps create the product, but it is not necessarily the product. The finished offering is the specific solution built around a concrete demand, closer to the point of pain. The goal is to become the person who uses AI to create value, not to mistake the AI itself for the value.
Give Recurring Jobs a Visual Home
A universal chat can ask an agent to do almost anything, but it also makes every activity look the same. That becomes difficult when someone is managing many kinds of work at once. A dedicated app gives one job a recognisable piece of visual space, helping the user understand where they are and what they are there to do.
The useful interface changes with the job. Pete described a 30,000-foot chat, a two-foot chat for micromanaging something and separate apps that visualise particular domains. In Flight Deck, he also creates channels for different ways of working, from quick dialogue to longer feature work. Each gives the agent a different frame instead of forcing every interaction into one undifferentiated stream.
Cheap software does not mean an interface should be regenerated on every visit. For recurring work, the better pattern is to build a personal visual shape once, keep it familiar and revise it when the job changes. Snack Studio followed that path. Its original screens exposed each pipeline so Andy could isolate failures, test each stage and develop trust in the outcome. Once the process became reliable, the interface could collapse into a simpler view while still allowing individual parts to be edited or regenerated.
Private Access Without Public Endpoints
Pete is testing a proof of concept for people who run Autopilot on a personal laptop or another machine they don’t want to expose online. Today, making its Wingman Apps (“WApps”) reachable can mean configuring domains, DNS and separate endpoints. A local proxy inside the Wingman app would instead make WApps running on that machine accessible through Wingman while keeping Autopilot itself private.
The second part of the design is a FIPS reverse proxy using signed approval for individual devices. An owner could approve a particular device to access the WApps, while traffic from devices without that approval would simply be discarded. Instead of making the WApps publicly reachable, access could be limited to devices that have explicitly been allowed in.
That becomes more useful as one person runs Autopilot across a desktop, laptop or several other machines. Each additional installation currently brings more endpoint and DNS work. Routing access through Wingman could remove that repeated configuration and make local development WApps easier to share with approved people. The apps can stay on the machines where they are running, while authorised devices still have a way to reach them.