Even Robots Must Take Notes
We've seen AI get more forgetful the more information we give them. It's time for AI to take note of how humans keep track of things.
You're late to class, didn't study, and must sit through a lecture on advanced mathematics (or a compliance training in an office, whichever you find most boring). You forgot your notebook and end up without any lecture notes. Now, an hour later, you're doing homework and try to remember what you learned.
You remember things at the beginning of the lecture. Topics connect to your previous learning quite easily. But then things get a bit murky... You've got something written down, but it looks wrong. You remember, at the end of the lecture, there are a few ways to check your results. Running through these checks, you can confirm that you're wrong, but you don't know why...
Without notes to reference, you're a bit lost. This is exactly the way AI feel when given too much data and no way to take notes!
The Challenge of Noteless AI
In a recent paper, Lost in the Middle - How Language Models Use Long Contexts, researchers found that when ChatGPT, and other AI, are given a lot of text they tend to forget or misremember some of it. Specifically, the information in the middle of the text. Similar to our lecture, the AI are less certain about information in the middle of any given data. It's uncanny that AI continue to act so human.
To help explain how this happened, the paper points to AI training. AI haven't been trained on larger chunks of information paired with accurate reflections over this information. Instead, they have been trained on shorter sections of text, often derived from how humans respond to information.
Where could we get this training data? Perfect recall isn't something we expect until we think about a machine "learning to read"” Then we want exact results. We want machines to do the drudgery of figuring everything out and just tell us what we need to know or what might have missed in our own reading. We don't want machines to read how we read or act like us!
But, hey. This is science, not science fiction. Because humans don't act like a prefect machine, we unfortunately don't have anything to model these machines after. It's becoming more obvious that, because these AIs are trained by humans, AI have even copied our faults.
How would we help a human get over these faults? How would you keep track of key pieces of information when interacting with a large and complex set of information?
Taking a Cue from Human Memory
In another Percy Liang paper titled, Generative Agents - Interactive Simulacra of Human Behaviour, the research team simulated a world full of AI-driven people. They woke up, ate, worked, socialized, planned, and slept.
A major challenge when creating this simulation was getting the citizens of this experiment to think like people. To solve this, the team constructed a complex set of memory systems. Short-term memory helped them remember where important items were, track immediate needs, and engage in conversations. Long-term memory was used to plan and remember important facts (among other things). These enabled the AI people in this simulation to act like people and even plan an ad hoc Valentine's Day parties with others.
It's a fascinating paper and I recommend you read it.
The AI were able to organize and stay on task when they were given some place to keep their memories stored. Memories tracked, at a high level, are summaries of what was happening around them. These memories are essentially notes on their environment!
Getting AI to Remember with Notes
This all leads us to some creative thinking about how to use AI to solve problems. While we'd love to treat the AI like machines, we need to assist their ability to remember and provide them with notes that direct their actions. Helping them take notes is essential to developing good AI workflows.
When we interact with ChatGPT, we're usually already keeping track of what the AI is thinking. They're using us as their long term memory, and we can remind them when they forget important information. In part, this is why chat interfaces are beneficial for this era of AI. They can sort of rely on us to perform well.
However, say we want an AI to read a large section of text and help build a summary of that text. Then we have a more challenging role to fill because the AI is going to misremember or forget information. We need to help give the AI note-taking capabilities.
If we split the text into smaller chunks, we can ask the AI to give us short summaries of each chunk. When we're finished, the AI can combine those pieces together to summarize what was covered. Just be careful that the pieces aren't too large when combined, otherwise we're right back where we were and the AI is being forgetful again.
Helping AI build plans isn't new, and there's already been a lot of research done to understand how to get AI to follow plans that help them accomplish tasks. A popular tool for building a plan is Chain-of-Thought (CoT) Prompting. We ask the AI to write out a plan for steps to take, which it will try to follow:
Question: Marty has 100 centimeters of ribbon that he must cut into 4 equal parts. Each of the cut parts must be divided into 5 equal parts. How long will each final cut be?
Answer: Let's think step by step.
Give this a try in ChatGPT, and you'll get back a sequence of steps and, usually, a pretty accurate result. We can leverage this pattern to get the AI to make plans for itself which can be referred back to later. These plans help AI to better organize its approach to solving problems.
The Path Ahead
While this only covers a few methods to help AI handle larger sets of text, many others methods exist and continue to be explored. While many believed that providing AI with a larger "context window" (the ability to be given more text) would solve many limitations, it seems that those limitations continue to exist. At this point, I won't use the longer context windows of these models until these memory problems are resolved with new modeling or better training data.
For the time being, the AI development community is going to be innovating around these limitations to engineer systems and solutions rather than just relying on LLMs to solve these problems. If you're doing something interesting to help AI better understand and remember larger sections of text, we'd love to hear about it!
We'll continue to post interesting ideas about how to use AI here and on our main site, Stable Discussion. Keep checking back to keep updated on our latest thinking around these ideas!