Reasonable AI

In my quest to find reasonable AI's I've been researching how data scientists and engineers compensate for all the failings of The Illusion of Thinking (Shojaee et al) machines:

• Logic -- helping me simply the growing tangle of redundant slop in my open source projects
• Arithmetic -- counting errors and doing statistics on my students' programming assignments
• Spatial reasoning -- solving simple and trunk-packing problems problems
• Temporal reasoning -- helping me schedule my day or research AI and teaching approaches
• Abstraction -- helping me design a competitive ARC Prize bot
• Planning -- helping my mentees plan job searches in a world crowded with slop jockeys
• Game theory -- strategizing with me in the game of students deep-faking themselves out of an education
• Security -- not clobbering my databases or exposing my secrets while writing a python script

I'm particularly interested in graph search approaches that give token-predicting transformers a reproducible, verifiable reasoning algorithm for making more accurate and reasonable planning decisions. Drug molecule synthesis pathway planning is one such domain where graph search approaches shine. Drug synthesis can be modeled as a decision tree or graph search problem in the same way that even the most advanced AI still does when navigating a self-driving car from turn to turn. And drug discovery is where both of my recent consulting customers have gone all in on AI, against the advice of researchers, chemists and data scientists, including me. Some of those researchers have even demonstrated approaches that achieve SoTA performance without leaning on LLM for their unjustified and unexplainable reasoning ability:

• Variational graph encoders ... for drug design -- A VAE on chemical molecule graphs (instead of SMILES)
• AI ZynthFinder -- an MCTS on reaction SMILES strings
• Retrosynformer -- a decision transformer approach to SMILES retrosynthesis

Some of these models are so efficient that I can train 50 alternative hyperparameter combinations in a week on a homelab PC, Framework Desktop, 32 CPU Cores, 124 GB RAM shared with integrated AMD Ryzen GPU. I was able to run simultaneous CPU-only and GPU-only optuna studies. I also experimented with resuming particularly promising training runs after early stopping patience countdowns.