curtiscovington.com

What It Is

Curtis Covington is a personal technical blog focused on AI interpretability, Sparse Autoencoders, activation steering, local-first toolchains, and experiments with open-weight models. It is not an AI application or SaaS tool in the traditional sense, but rather a collection of reports and implementation notes from the author’s mechanistic interpretability experiments on models such as Gemma and Pythia, with an emphasis on “metrics first, caveats included.”

Core Capabilities and Experiments

The most important content falls into two types of experiments. The first is a reproduction of Natural Language Autoencoder interpretation/editing based on Gemma 3 12B, attempting to decode activations into natural-language explanations, edit the text, reconstruct it back into activations, and inject those activations into the model. The second is SAE-based feature steering, including feature amplification/suppression on Gemma Scope and Pythia-70M, as well as evaluations using minimal-pair prompts, holdouts, wrong-layer/wrong-hook tests, random controls, and more. The author clearly distinguishes between logprob probability shifts and actual control over completion behavior, taking a relatively cautious research stance.

Pricing, API, and Chinese Support

The site appears to be a public blog, with no subscription, paywall, free tier, or commercial pricing information found. There is also no product-level API, SDK, or third-party integration documentation; the articles instead refer to experimental repositories, Python commands, and local Transformers/MPS workflows. The main content is in English, and no Chinese content or Chinese interface support was observed.

Pros and Cons

The strengths are its transparent experimental protocols: models, layers, metrics, control conditions, and limitations are described in considerable detail, making it especially useful for learning how to turn AI steering demos into more rigorous controlled experiments. The downside is that it is not a ready-to-use tool and the content is highly research-oriented. Many results are based on small sample sizes, and the author also acknowledges that NLA completion steering is relatively weak, random delta interventions complicate causal interpretation, and, aside from the code category, most SAE category evidence is not particularly strong.

Who It’s For and Access from China

It is best suited to AI interpretability researchers, LLM safety engineers, and developers who want to reproduce SAE/NLA experiments locally. It is not suitable for users looking for a general-purpose chat, writing, image-generation, or enterprise AI platform. The crawled text does not provide information on access from mainland China, so network availability and payment methods are unknown. Alternative references may include Transformer Circuits, the Anthropic interpretability blog, Gemma Scope papers, and related open-source repositories.