About

Summary

I study what actually happens inside large language models — the circuits, features, and representations underneath the fluent output — and I build tools on top of that research. I have worked at the intersection of artificial intelligence and human cognition for nearly four decades, beginning with neural networks in 1987: a PhD in Cognitive Science from UC San Diego, a career leading research teams in natural-language systems, and, more recently, hands-on interpretability work on how models encode style, knowledge, and identity.

The through-line has never changed. My doctoral thesis, Artificial Languages, Virtual Brains, used neural networks to study language learning back when the networks were small enough to inspect every unit. The models have grown; the question — how would we actually know what one is doing? — has not.

Currently

• Inside the Black Box — a newsletter and companion podcast on mechanistic interpretability and artificial psycholinguistics, demystifying how LLMs work. insidetheblackbox.ai
• YourVoiceCraft — a privacy-first writing tool I created, built on this research: a private model trained on a writer’s own archive — it writes in their voice, quotes what they’ve actually written, catches drafts that contradict them, and keeps their work off Big AI’s servers. yourvoicecraft.ai
• Interpretability research — LoRA fine-tuning on open models, probing what changes inside a network when it learns a writing style, and using sparse autoencoders to compare fine-tuning strategies at the feature level.
• Arvoinen Consulting — I advise organizations on building with AI: what their models are actually doing, where the technology genuinely fits, and how to use it honestly.

Podcast Appearances

My own podcast, Inside the Black Box, is the audio companion to the newsletter — me, not interviews. I’m also a guest on other people’s shows, talking about interpretability, local AI, and what these models are really doing. A few recent conversations:

• Marketing in the Age of AI (Emanuel Rose) — Local AI, Clear Workflows, and the End of Fluency Theater. Listen →
• Champion Mindset Collective (Anthony Dahya) — 30 Years in AI: The Truth About How It Works (and How Humans Stay Ahead). Listen →
• Lift-Off (Keith Newman) — Why AI Still Lacks True Agency, and What Founders Must Know. Listen →
• Story-Driven Business Podcast (Susanna Rantanen) — AI Leadership and Trust: How Leaders Can Adopt AI. Listen →
• The Lighthouse Sessions (Jeff Borschowa) — a wide-ranging conversation on building with AI honestly.

Work Experience

Principal, Arvoinen Consulting, Los Angeles, CA,  current

Chief Scientist, CitizenNet, Condé Nast, Los Angeles, CA, 2010 – 2020.

Senior Research Scientist/Engineer, Fox Interactive Media, Los Angeles, CA, 2008 – 2010.

Principal Computer Scientist, MetaLINCS, San Jose, CA, 2007

Senior Scientist/Director of Research, H5 Technologies, San Francisco, CA, 2005 – 2007.

Senior Engineer & Project Lead, Entrieva, Reston, VA, 2003 – 2005.

Principal Scientist, Comprecorp, Nevada City, CA, 2001.

Senior Engineer, Ask Jeeves, Emeryville, CA, 1999 – 2000.

Skills and Research Interests

Interpretability & modeling. Mechanistic interpretability (circuits, features, superposition), LoRA fine-tuning, sparse autoencoders, transformer internals.

Languages & platforms. Python, Java, C, C++, SAS.

Research methods. The full experimental-psychology toolkit: literature review, hypothesis development, experimental design and control, protocols, and appropriate statistical analysis.

Standing interests. Language processing and acquisition in humans and neural networks; large-scale cognitive systems; how working memory and the broader cognitive system shape language; and, throughout, the gap between what a model produces and what it understands.

Selected Publications

Blackwell, A. (2012). A Gentle Introduction to Random Forests, Ensembles, and Performance Metrics in a Commercial System. Web blog post. CitizenNet.com.

Blackwell, A., Bates, E., & Fisher, D. (1996). The time course of grammaticality judgment. Language and Cognitive Processes, 11(4), 337–406.

Blackwell, A.W. & Bates, E. (1995). Inducing agrammatic profiles in normals: Evidence for the selective vulnerability of morphology under cognitive resource limitation. Journal of Cognitive Neuroscience, 7, 228–257.

Blackwell, A.W. (1995). Artificial Languages / Virtual Brains. Doctoral dissertation.

Education

Ph.D., University of California, San Diego, 1995: Cognitive Science and Psychology.

A.B., Brown University, 1987: Honors Cognitive Science and Honors Creative Writing.