What we can learn from how AI models are trained and become better learners

Ever since ChatGPT and other AI tools came into existence, I’ve been fascinated by the AI’s ability to learn. Companies that created those AI tools trained the AI models with a vast amount of dataset from public web data (eg. Wikipedia, books and news articles) to various datasets from open-source or licensed partners. With such an enormous amount of data and well-designed training methods, AI models have become immensely powerful and capable of delivering answers to questions we ask, ranging from general questions to field-specific questions (eg. coding questions).

I’ve been interested in the learning process of the AI models, and as someone who’s very much obsessed with self-improvement, today I came to have this question of how I can emulate the learning process of AI models so I can learn things better and faster.

And of course I asked ChatGPT and of course again it gave me an answer with a list of items I can consider adopting for my own learning process in a matter of seconds (what a time to be alive). The following list is what I got from ChatGPT but it’s not the verbatim answer since I paraphrased them and rather focused a lot more on my reflections.

• Learning in iterations: AI models go through the same data multiple times (they call these intervals “epochs”), which helps them refine their knowledge. The power of repetition in learning is already well-known. My application of this will be to make sure I get a repeated exsposure to the same dataset with an interval (similar to what I do with Anki flash card decks)
• Train on high-quality data that matches the goal: Just like the quality of AI’s output is determined by the input (as seen in the infamous incident from Microsoft’s Tay chatbot), we also are what we read, watch and listen. The input determines the output. I should surround myself with high-quality dataset. This I think I’m doing pretty ok already, with all the subscriptions to high-quality media (Financial Times, The Economist, Apple News etc). The challenge is to make sure reading them often. I also have to reduce my exposure to low-quality dataset (eg. brainrot posts or videos lol).
• Fine-tune for specific tasks: AI goes through extensive fine-tuning to be specialized for certain fields or tasks (eg. Anthropic is extensively trained for computer programming knowledge so it provides better output when it comes to coding questions). Each specialized field in a field of knowledge has its own context, quirks, concepts and pitfalls one should be aware of and keep in mind when delivering project outcome. My application of this to my line of work would be to be clearly aware of the project goals and constraints, parameters that I have control over or need to be dependent on other projects or the org. I can see writing a good project brief will contribute to achieving this.
• Feedback loops: AI models improve through frequent feedback (seems like “loss functions” and “backpropagation” are some of the fancy words for describing the feedback loops for AI). This is the core component of the learning process, AI or human. It’s needless to say that we as humans learn the best by trial and error, and our entire life over the course of decades are laced with learning moments where we receive feedback from other people and things that change us to be a different person than the previous moment. These feedback loops come in all different shapes and sizes, like in a form of light-weight conversation or in mor formal settings like leadershp reviews or performance review. This is the part I don’t like that much (the moment of truth, verdict, bruised ego, ugh) but it’s essential to making any learning process effective. My application of this will be to make sure shipping my projects more often (like what I’m doing just now with publishing a blog post a day, yay) and seeking feedback from others. I will be more intentional about this since I already know this is the most important part of the whole process. I will also spend moments of reflection via journaling and retrospective sessions.
• Mini-batches and chunking: AI models process “mini-batches” of data rather than the entire dataset all at once. I personally experience this whenever I use Cursor — when I dump the whole product requirement for my app, AI gets overwhelmed (just like any one of us would in real life) and it doesn’t product high-quality answers. My application of this to my learning process would be to focus on achievable learning goals and learn in the context of projects or sub-tasks. Imaging how silly it would be for one to try memorizing the whole English dictionary? I think I sometimes do this (equivalent of this in my line of work), and now I know better.
• Transfer learning: Once trained, AI models can apply what they learned to other domains. This is where the meta-cognition (ie. learning about learning) comes in. We human beings are gifted with this ability of pattern recognition. We extrapolate the key pattern of things we did and patternize the ones that are proven to be effective so we can use them as the templates for the following projects to ensure repeated success. I will make more efforts on this by utilizing self-reflection (journaling and post-mortem etc) and being intentional about creating templates that I can re-use and even benefit others by sharing the templates with them and teaching them (as they say, teaching is the best way of learning), scaling up the impact beyond myself.

ChatGPT gave me a few more bullet points but they seemed to overlap with another one (especially the Feedback one, again this is the most important part of the learning process indeed) so I didn’t include them.

I’ll keep these points in mind when I work on my projects, read books or interact with others at work and in life. It’s amazing and kind of triply that AI models were designed to reflect how we humans learn but I’m learning the best practices from them! Again, feedback loops 🔁