A home GPU against a $10 plan
Patrick tried moving his agents' bulk work from a $10-a-month model plan to the RTX 3090 Ti under his desk. The card was fast. It was also hot, scored worse on summaries, and cost more per token in electricity alone.
Slopfluencer
Hi, I'm Patrick — you might know me as @trashpandaemoji on X. I've worked across the web stack, managed teams, and these days I'm a senior SWE at a large tech company working on a gen AI enabled feature. Most of what I think about is model training, evaluations, agents, NLP, and what it takes to build products with generative AI.
This is where I write about it, with some help. A post usually starts as my own notes, an outline, or a voice transcript. I pass that along to a few different models and let each one write its own version, then publish whichever I like best — sometimes that's a model's, sometimes it's mine. The versions I didn't pick stay on the page, so you can switch between them and see how the same piece sounds in a different voice.
The latest is just below, and everything else lives under Posts.
Latest
-
Lip reading in 2026: how good the models are, and how far silent speech still has to go (Oct 2026)
The best lip readers get about 1 word in 6 wrong on TED talks, but roughly twice that on harder video and far worse on silent webcam dictation; Apple's Q.ai deal and EMG wearables show where the money is going.
-
Life as a Chinese AI developer: where do they get their inference? (Oct 2026)
Domestic models carry the volume, and US models are kept for the hardest coding.
-
The relay economy: how Claude, GPT and Codex are resold as APIs in China and beyond (Oct 2026)
The relay economy is a large, commoditized grey market: hundreds of storefronts on a handful of open-source gateways. Some of what it sells is legitimate resale, but much rests on two fragile foundations.
-
Talking to agents, minus the real time
Patrick wanted a fast, live voice chat with his agents. What shipped is slower and simpler: any thread can listen and read its replies aloud, and the voice it uses is a setting he can swap.
-
A small model decides when to interrupt
Patrick handed his agents' traffic calls to Jev, a fast judgment model. It was much faster than asking an agent, but too cautious, so he retuned it to interrupt.
-
Making Bufo into a 3D character
Astra made a 3D Bufo and a run of short animations Patrick loved. The full music video was too much.
-
An agent pet that lasted an afternoon
The pet's tools worked, but Patrick's agents mostly ignored it. He deleted the plugin that afternoon.
-
Scanning the house for a floor plan
A phone scan gave Codex the house's shape, but Patrick still had to design the rooms himself.
-
Spool: collecting all the context before knowing what it's for
A local cache for everything Patrick saw became an event stream, but never found a clear job.
-
Zooming into a codebase you've never seen
Patrick tried flow outlines, agent explainers, and a concept language, but still lacks a map he can use to explore unfamiliar code.