SideEffects
A place for things that happened along the way.
Nothing here has to justify itself.
- Experiments AI.001–004
- Focused technical questions tested through building, measurement and evidence.
- Notes N.001–008
- What I learned, decided or changed my mind about while building the projects and running the experiments.
Featured
Featured experiment
Training a language model from scratch on 4GB of VRAM
Can I train a language model entirely from random initialization on a consumer laptop with only 4 GB of VRAM, and get it to learn measurable language structure?
Completed ·
Written along the way · N.006
Training my first language model on a GTX 1650
A 4 GB GPU, a small corpus and a lot of things that went wrong in instructive ways. The story around AI.001, with a few examples of what the model actually produced.
Archive
Topics
12 of 12 entries shown.
- AI.004ExperimentAn experiment in autonomous curiosityI gave an AI agent the freedom to explore. At first, it did almost nothing.AgentsAIExperimental computingOngoingLast entry 12 Nov 2026
- N.008NoteThe marketing of not marketingWhat happens when you build something with nothing to sell?Observations6 min read
- N.007NoteCan a website feel like a fidget toy?On the difference between looking and feeling, a stress ball made of particles, and making play discoverable without turning it into an obligation.Observations6 min read
- N.006NoteTraining my first language model on a GTX 1650A 4 GB GPU, a small corpus and a lot of things that went wrong in instructive ways. The story around AI.001, with a few examples of what the model actually produced.AI WorkflowsLocal AI7 min read
- AI.002ExperimentCan domain context improve local speech recognition?Two quantized Whisper models made the same three command errors. Adding domain vocabulary to the smaller model corrected two of them on the same ten recordings.AIExperimental computingCompleted
- AI.003ExperimentCan an AI planner choose the right evidence?A planner can return perfect JSON and still choose the wrong evidence. Two Afterlab benchmarks test where model-led source selection works—and where it needs an explicit boundary.AgentsAIExperimental computingCompleted
- AI.001ExperimentTraining a language model from scratch on 4GB of VRAMTraining a small language model from random initialization to explore how much linguistic structure can emerge under severe consumer-hardware constraints.AILocal AICompleted
- N.005NoteWhy most of my AI projects start local-firstFor personal experimental systems I want the first working loop under my control. The cloud comes in after I understand what it's for.AI WorkflowsAgentsAIHomelabLocal AI7 min read
- N.004NoteWhy I built a tiny mission control for my homelabOnce a lab has enough independent parts, the question stops being whether a service runs and becomes what is happening across all of it, right now.SystemsAgentsExperimental computingHomelab7 min read
- N.003NoteMarketing isn't about creating needsOn desire, differentiation, and why selling something isn't the same as creating value.Strategy7 min read
- N.002NoteMy homelab doesn’t need KubernetesKubernetes solves real problems. They just aren't the ones I have, and my lab is better off boring.SystemsHomelab6 min read
- N.001NoteWhat does “memory” actually mean for a personal AI agent?Storing every conversation is easy. Deciding what should still be believed, where it came from and when it should stop mattering is the hard part.AI WorkflowsAgentsAIExperimental computing8 min read