In Isaac Asimov’s Little Lost Robot (and other short stories), Dr. Susan Calvin serves as the chief robopsychologist at U.S. Robots and Mechanical Men, Inc., a major manufacturer of robots in the 21st century. Her emotional and mathematical ability to grapple with robot minds makes her the go-to counsel when robots malfunction in ways that humans cannot understand. As always, Asimov’s fiction-as-prediction is spot on, and as deep learning applications profliferate, we must mint a new profession: robopsychologist.
Robopsychology is the art of grokking the inner workings of deep learning systems, translating what is found into a medium interpretable by humans, and curing pathologies in robot thought.
Robopsychology does not currently exist as a profession outside of a handful of commercial and academic research laboratories. This is unfortunate, because to effectively deploy a technology in an engineering discipline, one must control its inner workings with ever-increasing nines of precision. Therefore, to achieve mastery over deep learning, the United States must subsidize and commercialize the rapidly expanding subfields of mechanistic interpretability, scalable observability, and their brethren.
Robopsychology has garnered little attention from for-profit companies because its roots lie in AI safety. AI safety is viewed as a non-profit enterprise, and so research has progressed due to goodwill and scientific curiosity. We can and should accelerate robopsychology by aligning research efforts with market incentives.
The market needs are present, but unrealized. If deep learning is the language of biology, we will need robopsychologists to extract insights modeled by neural networks. If defense robotics will be powered via machine intelligence, we will need robopsychologists to understand why our weapons attack friendlies. If the future of art is generative AI, we will need robopsychologists to prevent generative models from violating copyright law and producing unwanted content.
Lastly, robopsychology does not demand that robopsychologists be human. The emergence of automated techniques gestures towards a future where robopsychology is fully automated and AIs diagnose and treat other AIs.
But for now, it is humans who must develop techniques to understand our silicon brethren.
Robopsychology is the most underappreciated market in machine intelligence.
You need a robopsychologist.