Uncover Eye-Popping Insights to Turbo-Charge YOUR Understanding Today from the Long-Awaited Book by an AI Authority (Part 1)
Before delving into the insights from Ethan Mollick’s newly published book [1], let us briefly talk about my past experience with the granddaddy models of AI. This puts me in a unique position to compare the history of AI with its current state.
I’ve been working with AI since 2004.
Like most young and ambitious finance academics looking for a way to beat the market, my first focus was on prediction of financial asset prices [2].
Then, in 2006, I started using AI for the asset allocation problem of pension funds. This is a unique problem which drastically differs from the usual asset allocation paradigm. The aim is to find the optimal asset allocation for each period to achieve a reasonable net replacement ratio, i.e. the ratio of retirement income to final salary, rather than maximizing expected return for a given level of risk in a single period. It requires the solution of an optimization problem that is very hard, if not impossible, to solve by analytical methods. The problem comprises some constraints such as short-selling restrictions which renders the problem mathematically intractable.
AI methods such as genetic algorithms and simulated annealing enable us to efficiently solve the problem without sacrificing the inherent complexity [3], [4], [5].
Even these granddaddy AI methods are truly remarkable and efficient in problem solving. However, you needed to be a highly quantitatively oriented person who knows how to code in order to gain access to these methods.
Then, came the AI revolution. Revolution is not an exaggerated phrase when you consider what LLMs (Large Language Models which form the basis of ChatGPT, Gemini, Copilot, etc.) can accomplish.
Not even the actual coders behind these huge models know what LLMs are capable of. The good news is you do not need to be quantitatively oriented to start using these models. Anybody can start using them right away.
I use ChatGPT, Gemini, and Copilot extensively for almost any task.
Ethan Mollick is a leading authority on how to employ LLMs in business and academia. I benefit immensely from his work.
You can follow him on LinkedIn and read his articles on his website: https://www.oneusefulthing.org/. He has recently published a book on AI: “Co-Intelligence”.
As I read through it, I see many new angles on the current status and the emerging trends of AI. I want to share them in this series of articles. So, let’s start with “Part 1”:
Insight #1:
“Superior test results can stem from the AI’s proficiency in problem-solving or its prior exposure to relevant data during training, effectively rendering the test comparable to an open-book scenario.”
This was a real epiphany for me. Large language models are trained by feeding them numerous documents. It’s quite possible that former bar or medical exams, along with their answer keys, are included in the training material. Of course, LLMs do not memorize; they only retain the weights after training. However, these models are so sophisticated that they can generate new content very similar to the learning material. Therefore, we should not hastily conclude that AIs are ready to replace lawyers, doctors, coders, and many other professionals. It is possible that they might simply be “overfitting”. Overfitting is a phenomenon in data science where the model learns both the signal and the noise, resulting in exceptional performance with the training data but poor performance with test or unseen data. Hence, a similar phenomenon may occur with LLMs. There is a possibility that LLMs may perform poorly when faced with a completely original problem that has not been seen before.
Insight #2:
“When you provide data to an AI, the majority of current LLMs typically do not directly incorporate it into their learning process, as it is not included in their initial pre-training, which has usually been completed long ago. Nonetheless, there’s a possibility that the data you upload could be utilized in subsequent training iterations or to refine the model you’re utilizing.”
Should we be concerned about data privacy issues? For instance, you spend so many hours on developing a proprietary model for trading the markets. The code you have developed together with your companion LLM may not be readily available to other coders. On the other hand, it will be too optimistic to rule out the possibility that this code will be included in the training material for the future runs.
Insight #3:
“To maximize the benefits of this partnership, it’s crucial to define a distinct and precise AI persona, outlining its identity and the specific issues it should address.”
This is actually very important to get an appropriate response from the LLM. The persona should be consistent with the level of expertise you expect from the model. Without defining a persona, the output you get may not be at par with your expectations.
To be continued with “Part 2”…
References
[1] E. Mollick, Co-Intelligence Living and Working with AI. Portfolio, 2024.