In our recent post, we explained that current AI models aren't good at their "own" thinking but can be great collaborators on a variety of tasks.
The Simple Reason Behind AI's Limitations
There's a simple reason LLMs are great at what they do but not so great at "thinking on their own."
At their core, LLMs are sophisticated next-word predictors, sometimes called T9 on steroids.
By giving a sufficiently sophisticated neural network the task of next-word prediction and providing it with trillions of word sequences, we can enable it to uncover many patterns in the data, including patterns of thought and language.
But in this very concept lies a major limitation: LLMs are great at continuing thoughts, but they were never designed to create new ones from scratch.
Why AI Needs Human Guidance
LLMs are designed to collaborate with people who know what they are doing.
They require guidance to do a great job, and they will only do a great job if someone guides them properly.
If you simply ask "make me a million dollars", it won't be able to do it.
Here's The Research That Proves It
There's a substantial body of research on this topic.
A widely cited study, Generative AI enhances individual creativity but reduces the collective diversity of novel content (Doshi & Hauser, Science Advances), has shown that while AI assistance makes individuals' writing more creative, it reduces the group's overall creativity.
Another study, Echoes in AI: Quantifying lack of plot diversity in LLM outputs (Xu et al., PNAS), examined the creativity of LLM-generated stories and found them to be very similar to one another.
Another study, Who Is the Best Creative Thinking Partner? found that human-human groups outperformed human-AI and human-Internet groups on divergent thinking tasks.
The Training Data Problem
Another major limitation is the use of human-generated training data, which can be filled with flaws and mistakes.
You might have seen the news about LLMs reciting propaganda narratives.
Data filled with lies is especially prevalent in demanding and competitive areas, such as trading.
What We See in Our Own Work
We apply cutting-edge AI tools almost every day to help with market research, and we often find that LLMs provide misleading advice, sometimes the opposite of what our research shows to be true.
The conclusion is clear: LLMs are great collaborators, but they need people who know what they are doing.
As answers become increasingly cheaper, the value of the right questions becomes everything.
The Takeaway for Professionals
If you are a professional in a knowledge work field, AI will likely make you more productive and a more capable version of yourself.
And if you want to tap into the power of AI for trading, our AI-powered trading tools have been designed by experts who did their homework.
Best,
Val