From beginner to owned models
Beginner to owned model: what an LLM is, how a tiny GPT works, how to train one, then register it on Inferix.
Inferix does not create your weights. You train models yourself, then register them as clients. This path is the literacy layer: understand an LLM, build a tiny GPT, run it, then see how it later sits under RouteIQ and FineForge.
Who this is for
Platform and ML engineers who can ship services but have not trained a transformer. Also operators who want to know what owned/general-llm actually is before they write inferix.yaml.
The path
1 · What an LLM is
Next-token prediction, tokens, loss. No code yet.
2 · How a tiny GPT works
Embeddings, causal attention, MLP, generate. Map to source.
3 · Train and run yours
owned-llms repo: inspect, train, sample from a real checkpoint.
4 · Toy → Inferix client
M1–M3 families, then register, route, drift, promote.
What you will be able to explain
- Why a checkpoint file is “your model”
- Why untrained loss sits near ln(vocab size)
- What Inferix observes later (tokens, latency, drift) vs what training is
- Why cheap SLMs and a general LLM are different RouteIQ paths, not different physics
Inferix is later
Do not wait for the control plane to start this path. Train first. When LensAI and TraceForge ingest, the same checkpoints become owned-model clients.