Yesterday we established that an LLM is auto-complete at scale. Today we traced how the field got there — from a straight-line equation to deep networks.
1. Where We Left Off
- LLMs work by predicting the next token.
- Training is how a company builds the model. Inference is what happens when you use it.
2. AI Is Not New
AI is a field of study aimed at mimicking human behaviour and intelligence — understanding language, summarising information, recognising patterns.
It has been through cycles of hype and collapse, including two “AI winters”: 1974–1980 and 1987–1993, when funding and interest dried up after results fell short of promises. The current wave is the latest turn of a long cycle, not a sudden arrival.
3. Machine Learning vs Traditional Programming
| How it works | |
|---|---|
| Traditional programming | You hard-code the logic. Same input, same output, every time. |
| Machine learning | You give it data. It finds the pattern and predicts. |
The supermarket example: predict today’s sales from the number of visitors.
Y = MX + C- Y = sales, X = visitors
- M = slope, C = intercept
In traditional programming you’d write M and C yourself. In ML, the model works out M and C from past data. That’s the whole difference, and every larger model is an elaboration of it.
The consequence to internalise: ML models are statistical and probabilistic. They produce estimates, not certainties. That is precisely why ChatGPT and Gemini get things wrong — not a bug, but the nature of the method.
4. Neural Networks
Inspired by the brain: neurons connected in pathways that activate for particular tasks.
Structure: an input layer, one or more hidden layers, and an output layer.
Inside a single artificial neuron:
- Weights — how much each incoming signal matters
- Bias — a constant that shifts the result
- Activation function — decides what the neuron passes on
Two clarifications worth getting right:
- Weights are ordinary numbers and are often negative. A negative weight is how the network learns that a signal should reduce the output. They aren’t confined to 0–1.
- An activation output may be squashed into 0–1 by a sigmoid. That’s the part people are thinking of when they say “between 0 and 1.”
- Real biological neurons don’t fire in clean binary states either. The on/off picture comes from the 1943 McCulloch–Pitts model — a useful simplification, not neuroscience.
5. Training and RLHF
A fresh model starts with random weights and produces nonsense. Training is the process of pushing those weights toward something useful.
RLHF — Reinforcement Learning from Human Feedback — is a fine-tuning stage that comes after pretraining, not the main event:
- The model produces several candidate answers.
- Humans rank them from best to worst.
- Those rankings train a reward model.
- The reward model guides further weight updates.
So humans aren’t correcting answers one at a time. They’re teaching the model what “better” looks like, at scale. That’s what turns a raw next-token predictor into something that behaves like an assistant.
6. Deep Learning
Deep learning means a neural network with multiple hidden layers — conventionally more than one. There’s no official cut-off; “deep” simply means enough layers to build up complex representations.
More layers means harder problems become solvable: self-driving cars, speech recognition, and the architectures behind LLMs.
7. Why GPUs
Early ML ran happily on CPUs. Neural networks changed that, and demand shifted to GPUs — which is why Nvidia and AMD have grown the way they have.
✅ Action Item
Research the connection between neural networks and GPUs. What is it about the way a GPU works that suits this job so much better than a CPU? Come ready to explain it in your own words.
⏭️ Next up: building on deep learning toward the architecture behind modern LLMs.
1 comment