132 karma · joined July 13, 2024
Currently experimenting with ways to make AI deployment as seamless as web deployment. Also building Oxlo, a global platform that lets developers launch and test AI models on GPU instances (free tier for small models, like Vercel but for AI).
Background in Rust, Substrate, Linux systems, and computer vision pipelines. Focused on building practical infrastructure that scales — from decentralized edge nodes to secure on-prem inference for smart cities.
Always interested in discussing: distributed compute, model serving, real-time AI, hardware acceleration, and deep-tech company building.
This post highlights how something that used to be a fun, lightweight hobby has turned into a full-time maintenance burden. Systems like Matrix are powerful, but they’ve become so intricate that even skilled engineers struggle to run them reliably. The result is a slow drift back toward centralized platforms, not out of preference, but because convenience keeps winning over autonomy. It’s a reminder of the growing gap between the ideal of a user-owned internet and the realities of modern software.
It also makes me wonder how future kids will see this era. Maybe it will look the same way early mechanical computers look to us. A short period where people had to be unusually inquisitive just to make things work.
Models keep getting cheaper and more capable every few months. However, the underlying compute economics do not deflate at the same rate. GPU provisioning, inference orchestration, bandwidth constraints, latency guarantees, regulatory requirements, and failure handling do not become magically simple because a new model improved its reasoning. In reality, each improvement on the model side increases pressure on the infrastructure side. Bigger context windows, heavier memory footprints, more parallel requests, and more complex agentic workflows all increase the operational burden.
For infrastructure teams, waiting does not help. The surface area of what needs to be built only grows. You cannot delay autoscaling, observability, scheduling, routing, or privacy guarantees. Applications will always demand more from the infrastructure, and they will expect it to feel like a commodity.
My view is that technical deflation applies much more to application startups than to infrastructure startups. App founders can benefit from waiting. Infra founders have to build now because every model improvement instantly becomes a new expectation that the infra must support. The baseline keeps rising.
The real moat in the next era is not the speed of feature development. It is the ability of your infrastructure to absorb the increasing chaos of more capable models while keeping the experience simple and predictable for the user
In a crowded AI tooling market, this kind of contrast joke on the front paired with a real product behind it, cuts through noise in a way a normal landing page wouldn’t. People mock the gimmick, but the gimmick is doing exactly what it’s designed to do, get everyone talking.