The year was 2001; the month December. A new form of transportation was breaking into the public consciousness – The Segway. Pundits, commentators, and the tech-savvy early adopters told us all about how this new form of personal transportation would shape the way we interact with the world. But… it never lived up to the hype.
The year is around 2020 (no SeaLab though); no clear month. But a new form of AI was breaking into the public consciousness – The Large Language Model (LLM) chatbot. Admitted much more popular and widely adopted than the Segway, the likes of Chat GPT and Claude and Grok and Co-Pilot and…the list goes on and on have actually changed how we do things. But… will it live up to the hype?
Obviously, since I am comparing the two, it should be clear that I believe the answer to be no. The reason for that is as a daily user of Chat GPT and Co-Pilot I understand what they are good for and what is AI Snake Oil. Even in areas where LLMs are supposed to succeed the dreams of a future where ‘AI’ frees us from work are just that dreams.
There are numerous examples of the dark side of AI. I don’t mean that AI is putting people out of jobs (although that is sadly true) but rather the contrary that AI hype has fooled businesses into stupidity in adopting the technology. A sample of the cautionary tales include:
- Gary Marcus on Steve Eisman’s (of the Big Short fame) podcast cautioning against LLMs
- Havard Business School on navigating the jagged frontier of AI (i.e., LLMs) whose key takeaway is that just because an LLM is good with one human problem doesn’t mean it is good with similar ones.
- Cold Fusion’s report that AI fails at 96% of jobs based on Scale AI’s Remote Labor Index (link to their study).
LLMs (and their more sophisticated agentic spawn) can be useful for some things, like designing code if (a big if) the design pattern has been discussed at length by humans within the ecosystem of the internet. But (a big but) if you are trying to do something off the beaten path they suck – plain and simple. They are smug and irretrievable wrong at the same time. A dangerous combination.
Now onto the columns.
One such task that LLMs simply fail at is being able to design a rule-based term rewriting system for computer algebra systems. Using computers to simplify expressions of all sort is a dream that has been alive for centuries and it seems only incremental progress has been made. This month’s Aristotle2Digital explores the low-level, narrowly scoped basic simplification routines in SymPy, demonstrated again that the human aptitude for this kind of task is still unequaled.
Finance’s basic tools find themselves at the intersection of economics and accounting. The former explains the basic interactions between economic objects and agents and the latter assigns values to them. This month’s CommonCents continues the New Year’s resolution of working through Chapter 2 of the college-level finance book by van der Wijst.
On the most challenging of the classical physics problems is the gravitational interaction of more than three bodies. This month’s UndertheHood continues its exploration of the circular restricted three-body problem showing how graphical analysis of Jacobi’s pseudopotential can inform us of a number of important results including the existence of the Lagrange points, around which one of the most complex machines ever built by man orbits.
Enjoy!