Give a base text model the phrase “Explain why the sky looks blue,” and it may continue a document containing that phrase. An assistant is expected to answer the request. What changed besides model size?
Before you begin: Understand causal generation and the distinction between training and context.
Start with the prediction objective
GPT stands for generative pretrained transformer. The historical family uses causal language modeling: learn to predict the next token from earlier context. Pretraining provides broad representations that can be adapted or prompted for many tasks.
Early generative-pretraining work emphasized adapting a pretrained model to downstream tasks. GPT-2 explored broader task behavior through text continuation, and GPT-3 studied how instructions and examples in context could support tasks without task-specific gradient updates at inference time.
These are historical research steps, not a claim that each release solved all limitations of the previous one. Data, scale, training procedure, task formatting, and evaluation all changed. Avoid attributing every capability difference to one number such as parameter count.
Make the task look like training
A base completion model predicts text that fits the context. If you present a sequence of questions and answers, continuing with an answer can become a plausible pattern. Few-shot demonstrations communicate the expected task and format inside the input.
An instruction-tuned model has additional training that encourages it to respond to requests more directly. Demonstrations can teach the desired behavior, while preference-based methods can shape which responses are favored. This changes learned behavior, not merely the label on a chat box.
The InstructGPT work is one historical example of supervised demonstrations, preference modeling, and reinforcement learning. It does not describe every current assistant's complete training pipeline.
Add tools at the application boundary
A model may generate a structured request to call a search tool or calculator. Application code decides whether to execute it, passes the result back, and enforces permissions. Tool access can supply current facts or exact calculations that the model alone does not provide reliably.
If a system answers today's timetable correctly, inspect whether it retrieved the timetable. Do not use that success as evidence that the model's stored parameters contain today's update. The same separation applies to code execution, file access, and browser actions.
Inspect a small behavioral comparison
Use an invented instruction: “Return only the label question or statement for this sentence: Where is the room?” A base completion may continue with examples or explanation. An instruction-tuned candidate may follow the label contract more consistently.
To compare them, record the exact checkpoint or product, prompt, decoding settings, and outputs on multiple sentences. Include a statement with a question mark quoted inside it and a genuine question without punctuation. This tests task understanding beyond one formatting cue.
Do not fabricate a table of model results from expectations. If you have not run the candidates, label the examples as hypotheses to test.
Keep architecture claims within the evidence
Public model names and products do not always disclose complete architectures, datasets, or training recipes. Use official technical reports for those details and identify what is undisclosed. A familiar brand name is not evidence for a specific parameter count or mixture-of-experts configuration.
Likewise, a reasoning feature may change inference behavior without exposing all internal mechanisms. The next lessons discuss the concepts that can be measured without pretending every commercial system uses the same hidden loop.
Diagnose a current-information failure
An assistant follows your formatting instructions but gives an obsolete event date. Would more instruction examples necessarily fix the problem?
Separate behavior from evidence
No. It may need access to the current event notice and a rule for resolving versions. Instruction tuning can improve how it responds, but it does not ensure every current fact is available. Check retrieval and context selection before changing the model's training method.
Next, we will study a model trained to use context on both sides of a token and see why that objective fits a different range of tasks.
Sources
Language Models are Few-Shot Learners studies GPT-3's in-context behavior. InstructGPT documents one instruction-following post-training pipeline. These papers are historical references, not specifications for undisclosed current products.