Anthropic emotion vectors spotlight local runner gap

Hot take: local inference stacks are behind the research curve here. The plumbing exists, but it is fragmented and developer-first.

• –`llama.cpp`, `Ollama`, and `vLLM` are primarily inference servers; they are not turnkey interpretability environments for arbitrary concept probes.
• –The closest practical path is to use a model backend that exposes activations, then layer on tools like `TransformerLens`, `NNsight`, or `SAELens` for probes, steering, and activation sweeps.
• –Arbitrary concepts are feasible if you can collect activations and train linear probes or sparse autoencoders, but the workflow is usually custom and model-specific.
• –The “highlight input/output based on concept activation” idea is very doable as a product feature, but it is not something local runners generally ship out of the box today.
• –If someone wants this to feel productized, the opportunity is a local interpretability UI, not just another runner.