David Rostcheck, 4/20/2023

The post-ChatGPT breakout of Large Language Models (LLMs) disoriented many technologists. The concept that a statistical model running on a server could hold a meaningful conversation forced many to come to terms with fundamental concepts about what thinking is and how people do it.

While technologists were studying and building information technologies, cognitive scientists spent the last century studying exactly those questions. LLMs emulate many aspects of human thought. Because of this, as soon as AI models were able to hold human-level conversations, researchers discovered that many techniques and observations from psychology applied equally well to the models as to human brains.

As new disciplines such as prompt engineering suddenly emerge, the technology landscape has tilted. In dealing with modern AI models, grounding in cognitive science proves much more useful than traditional software programming knowledge. Concepts from psychology such as framing, priming, and confabulation can be taken directly into interactions with LLMs. Suddenly psychology majors hold the high ground in mastering automation.

So how can technologists learn these foreign concepts? For those struggling with the idea that LLM AIs think in human-like ways and looking to master cognitive science, here are some resources:

• The “Sparks of AGI” paper and/or lecture from Microsoft Research explains how we know that LLMs can truly think. Most people with technology-only backgrounds should start here.
• Dr. Alan D. Thompson’s archive of cognitive testing on AIs archives tests of AI vs. human performance on IQ, theory of mind, common sense reasoning, commonsense inference, and other such tests.
• Wikipedia explains IQ and its underlying concept, the g-factor, core concepts in understanding intelligence.
• If you can’t write down an explanation of how humans think, the resource you need is an intro psychology class. This one from MIT is older but good, especially lectures: 2 (emotion and motivation systems), 3 (association), 5 (attending), 7 (memory), 8 (cognition), 9 (child cognitive development), 11 (language development), 12 (intelligence). Technologists retraining to understand cognition will probably end up taking something like this class - it’s a fair bit of material but it really moves the needle in comprehending and working with intelligent systems.
• For a deeper dive into psychology, see Canadian psychologist Dr. Jordan B. Peterson’s two university classes archived on his YouTube channel (note: Peterson has now become a political figure, but his significance here is his work in psychology that eventually launched his life as a public intellectual):
  • Maps of Meaning - the Architecture of Belief
  • Personality and its Transformations
• The book The Neuroscience of Intelligence by Dr. Richard Heier summarizes current scientific understanding of human intelligence. Note that the “Neuroscience of…” series from Cambridge University Press also contains other relevant books to AI work, including The Cognitive Neuroscience of Memory, The Neuroscience of Expertise, The Neuroscience of Creativity, and Introduction to Human Neuroimaging
• Humans can be programmed, and many of the techniques used to program humans also work on AI. This field is known as “influence” or “persuasion.” The book Pre-suasion by Dr. Robert Caldini is the most influential recent work in that space. Author Scott Adams also wrote an accessible popular market book Win Bigly explaining persuasion techniques through the lens of the 2016 US election.

Understanding both AI and cognitive science makes one a rare and valuable resource. If you are a technologist looking to shine at working with AI, consider adding cognitive science training to your learning plan.