Feynman's wobbling plate and Threshold Concepts behind modern AI

Translate

Feynman’s wobbling plate and 8 core ideas behind modern AI

The same GPT-4 fixed 10x more GitHub issues in a better loop. Learning science calls the idea behind it a threshold concept. Here are eight, and how to learn them by following your curiosity.

Zairai

Oct 27, 2026

Share

I’m Luis, and after studying AI for ~10 years and having worked on professional-development software, I started building Zairai, an AI coach based on the ideas below. You can use everything here without it, including a study prompt that works with any LLM. The plug is one paragraph at the end.

The spark of discovery

After the war, Richard Feynman arrived at Cornell burned out, unable to start research. So he decided to play with physics “without worrying about any importance whatsoever.” Soon after, someone in the cafeteria tossed a spinning plate in the air, and Feynman wondered why it wobbled the way it did. For fun, he worked out the equations. Hans Bethe asked why it mattered; Feynman said it didn’t.

The wobbles eventually led him to electron orbits in relativity, then to quantum electrodynamics. “The diagrams and the whole business that I got the Nobel Prize for,” he later wrote, “came from that piddling around with the wobbling plate.” Feynman got his spark by playing at his edge. Many of us learning AI feel the opposite: buried, behind, studying out of anxiety rather than curiosity. But that spark is open to anyone who knows where to look.

The Wright brothers ran a bicycle shop. When the published lift data didn’t match their gliders, they tested wings on a bicycle wheel, built their own wind tunnel, and found out for themselves. Two years later, they flew. Discovery is curiosity that keeps showing up. AI feels a lot like that today: a field still being invented, where most of us learn by reading and tinkering as we go.

If curiosity is what drives real learning, the question is: how do you find that spark, and keep it, when the field moves faster than you can read and falling behind feels inevitable?

Start with why it feels so hard. The pace is real: AI moves faster than anyone can track. But most of the change is on the surface: a new architecture is often a variation on attention, a new training trick a variation on gradient descent. The faster the surface moves, the more the deep ideas are worth. Feeds inform without teaching; you can read about AI daily and still not know why a model does something poorly.

There are great courses and books out there, and we start them with real intent. Then life interrupts: a hard week at work, a move, a busy month. Coming back means remembering where you were and finding the next step, and that friction is often enough to stall us. The hard part is sustained progress, when attention is one of the scarcest things we have.

In 1984, Benjamin Bloom reported the average student tutored one-on-one outscored about 98% of a conventional class. Later reviews find smaller gains (VanLehn, 2011: about 0.8 SD), but responsive teaching still beats other approaches.

This pace of progress in AI is unlike anything we’ve seen, and it’s only getting started. To get up to speed and stay competitive, learning will have to look more and more like an elite tutor:

Fitted to your edge. Your goals, your level, and your curiosity, not a fixed syllabus, set the next step. Lev Vygotsky called the sweet spot the zone of proximal development: what you can do with help but not yet alone. Too easy bores; too hard discourages. At the edge, learning feels like Feynman’s plate.

Strategic and compounding. What you study matters as much as how often. Some ideas unlock everything after them, so a good plan reaches those first and goes deep enough that they change how you see the rest.

In 2003, education researchers Jan Meyer and Ray Land named something most learners have felt: threshold concepts, ideas that, once grasped, change how you see a subject; you can’t unsee them, and hidden connections start appearing. They’re often easiest in roughly the order the field found them, each the answer to a problem the last left open.

Here’s a window into eight that explain how modern AI works:

Behavior learned from data, not programmed. What if, instead of writing the rules, we let a program find them by running a prediction, measuring its error, and adjusting? It’s older than most people think: in 1958, Frank Rosenblatt’s perceptron, a single artificial neuron, taught itself to tell whether a punch card had a mark on its left or right side: a tiny first step toward machines that recognize images. His later Mark I machine stored the parameters it learned in knobs turned by electric motors. This enables machine learning itself, and everything after it.

Representations and embeddings. What if the model decided which features (inputs to the machine learning model) matter, instead of people hand-picking them for each task? This is one of the main value propositions of neural networks with many layers; these layers process the inputs and compose increasingly complex features, hence the name deep learning. In 2012, AlexNet, trained on two gaming graphics cards, won the ImageNet contest with 15% top-5 error, while the runner-up, built on hand-picked features, had 26%. This enabled the deep learning boom that followed. Go deeper: Jay Alammar’s Illustrated Word2vec.

Next-token prediction. What can a neural network learn by only guessing the next word? In 2017, OpenAI trained a network to predict the next character of a corpus of Amazon reviews, and someone noticed that one of its neurons turned out to fire on or off depending on whether a review was positive or negative. This was surprising because nobody had taught it that. The only task of the neural net was to predict tokens in a sequence of text, but in doing so it learned how to identify and represent underlying concepts. At the scale of internet data, the exact same task produced grammar, facts, and reasoning, the foundation of today’s models. Go deeper: Andrej Karpathy’s “Let’s build GPT”.

Attention. What if a model stopped reading only one word at a time? The 2017 transformer threw out the word-by-word reading language models had relied on, and its base model reached top-tier translation after just twelve hours of training on a single machine with eight GPUs. This enables the transformer, training in parallel on GPUs, and the speed up and scale that followed. Go deeper: Jay Alammar’s Illustrated Transformer.

Evals. How do you know a change actually made a general model better? The trap is old: Goodhart’s law warns that when a measure becomes a target, it stops being a good measure, and models are very good at hitting targets. Good evals enable real progress instead of chasing proxies that drift from what you meant. Go deeper: Hugging Face’s Evaluation Guidebook.

Post-training and RL fine-tuning. How do you turn a text predictor into a more helpful assistant, and then a reasoner? In OpenAI’s 2022 InstructGPT work, people preferred answers from a 1.3-billion-parameter model trained with human feedback over those from the 175-billion-parameter GPT-3. How you train after pretraining can matter more than size. This enables ChatGPT-style assistants and reasoning models trained on answers that can be checked. Go deeper: Nathan Lambert’s free RLHF book.

Harnesses and agents. Can you make a model far more capable without changing a single parameter? Put it in a loop with tools, memory, and checks. On the SWE-bench Lite coding benchmark, GPT-4 scored about 2.7% with one early harness and 28.3% with another: nearly ten times better, with the same model. This enables agents that plan, act, and retry, and it explains why the same model does better in one product than another. Go deeper: Anthropic’s “Building effective agents”.

Recursive self-improvement. How can models help build better models? In 2025, DeepMind’s AlphaEvolve, powered by Gemini, found a way to speed up a key piece of Gemini’s own training code by 23%, cutting overall training time about 1%. This enables a faster loop of AI research; how far it goes is open. Go deeper: DeepMind’s AlphaEvolve post.

These eight are an introductory set. There are of course many more, and which ones matter next depends on what you’re exploring and where you’re starting from.

In 1960, Jerome Bruner proposed that learning works like a spiral: you return to the same core ideas, each time in a fuller, more formal form. The first pass is a sketch: you can’t explain the idea yet, but you can sense its shape. On the next, it starts to feel familiar, and you notice it in places you didn’t expect. Later you can reason with it, predict where it breaks, and use it to understand something new. That’s when it’s yours: not because you memorized it, but because it changed how you see everything around it. Here’s a taste of what three passes at attention can give you:

Pass 1, the intuition. You see why attention was a breakthrough. Older models read word by word and slowly lost the thread; with attention, every word can look at every other word, so in “the cat sat on the mat because it was tired,” the model can work out that “it” means the cat. That’s what lets models stay coherent across long passages.

Pass 2, the mechanics. You see how a word decides where to look: each word asks a question, every other word offers a label, and the best matches pass along what they know. Once that clicks, the core equation behind modern AI reads like a sentence instead of a wall of symbols.

Pass 3, the fine print. You see the small design choices that keep it all trainable, like a single division that stops the model from fixating on one word and going numb to feedback while it learns. This is where the original paper stops feeling intimidating and starts reading like a series of clever decisions you could have made yourself.

If you want to go deeper, Jay Alammar’s Illustrated Transformer walks through all three with clear visuals.

You can turn these thresholds into a learning plan with any capable LLM:

Help me learn modern AI from first principles. My background: [...]. My goal: [build products / train models / research / lead]. Help me identify the most strategic threshold concepts to uplevel my core understanding and capabilities.

1. Ask me 3-5 short questions to find what I already understand. Wait for answers.
2. Plan from my edge: about one concept a week, 5-15 minutes a day, each concept revisited 1, 3, and 7 days later at greater depth.
3. When I explain a concept back, point out gaps and misconceptions without giving the answer, then ask one follow-up question.

Feynman didn’t plan his way to quantum electrodynamics. He followed something interesting just past what he knew, and each step made the next possible. That’s open to us too; we just need the right approaches and mindset.

I’m building Zairai on these ideas: threshold concepts and the spiral, plus spacing, retrieval, coaching at your edge, and other cognitive principles. Its bite-sized interactive sessions feel like real work and compound strategically: you step into the shoes of the people who made these discoveries and measure for yourself, so the foundations click and everything built on them makes more sense. It’s for anyone who wants to get better at technical AI, from beginner to near the frontier, free to start and launching soon: info.zairai.ai.