AI
Summary
From Symbolic Logic to Agent Harnesses
Symbolic logic showed how reasoning could be expressed through formal symbols and rules. Artificial neurons and perceptrons turned parts of that idea into computational systems, while backpropagation made multilayer learning practical. Deep neural language models learned representations alongside word prediction; attention-based models learned which context mattered; and the Transformer made attention central to highly parallel sequence modeling. The pretrained LLM became a reusable language system, the instruction-tuned LLM became better at following requests, and agent harnesses now connect models to tools, memory, validation, permissions, and controlled action.
Branch Outline
- Symbolic logic — Precise symbols and rules separate logical form from subject matter, enabling rigorous validity testing and foundational work in mathematics, computation, and logic-based AI.
- Artificial neuron — McCulloch and Pitts’s 1943 binary threshold unit used fixed excitatory and inhibitory connections to implement logical functions and establish a foundation for neural networks.
- Perceptron — Rosenblatt’s trainable binary classifier adjusts weighted inputs and bias after errors, learning a separating boundary when classes are linearly separable.
- Backpropagation — Chain-rule gradient computation made it practical to train multilayer networks by propagating error backward and allowing hidden units to learn internal representations.
- Deep neural language models — Beginning with landmark 2003 work, these systems jointly learned distributed word representations and next-word probabilities, improving generalization beyond count-based methods.
- Attention-based models — Soft, differentiable alignment let recurrent encoder–decoders dynamically weight relevant input representations at each output step, foreshadowing attention-centered architectures.
- Transformer — The 2017 encoder–decoder architecture replaced recurrence and convolution with multi-head attention, positional encodings, feed-forward layers, and causal masking for parallelizable training.
- Pretrained LLM — Broad next-token training turned decoder-only models into reusable language systems that could subsequently be adapted to labeled tasks through supervised fine-tuning.
- Instruction-tuned LLM — Fine-tuning pretrained models on request-and-response examples, often supplemented by ranked human feedback, improved their ability to follow user intent.
- Agent harnesses — Controlled execution loops assemble context, expose tools, validate actions, manage state and permissions, recover from errors, and constrain how model decisions become real operations.