AI · LESSON 13 / 25 · English
Large Language Models (LLMs)
Fluent text still needs evidence.
Content reviewed: 22 October 2026 · தமிழில் படிக்க ↗
Understand it in plain language
A large language model operates on tokens, which may be words, word pieces, characters, or other units depending on its tokenizer. Many generative LLMs learn next-token prediction during pretraining, then receive additional training for useful interaction. Context, retrieval, tools, and post-training affect behaviour. Generated text can be incorrect, and a context window is not a guarantee that every supplied detail will be used correctly.
How the AI approach works
Many generative large language models learn to predict the next token from preceding context. Training text provides its own targets: in a sequence, later tokens become answers for earlier prefixes. This is self-supervised learning, even though the optimization uses prediction targets. A tokenizer maps text to model-specific units; Tamil text and code-mixed text may use different numbers of tokens from an English translation.
Pretraining learns broad statistical representations from a large corpus. Instruction tuning uses examples of desired responses. Preference-based post-training can further shape behavior, using approaches such as reinforcement learning or direct preference optimization. These stages influence response style and usefulness, but they do not turn every generated claim into verified knowledge.
At inference, the model processes the supplied context and produces a distribution for the next token, repeatedly. Decoding chooses tokens from that distribution; temperature and other sampling settings affect variation. A low temperature does not guarantee truth, and a confident tone is not a calibrated confidence estimate. Model parameters, current context, retrieval, and external tools play different roles.
An LLM can summarize, transform, classify, and draft, but may invent details, miss constraints, or fail on unfamiliar tasks. Specify the task, supply necessary evidence, define the output structure, and evaluate realistic examples. For changing facts, consider retrieval or a suitable tool. For actions, enforce authorization outside generated prose.
Break the idea into steps
- Distinguish model training data from the context supplied for this request.
- Check the model card, tokenizer, supported inputs, and usage constraints.
- Evaluate Tamil, mixed Tamil–English, and task-specific examples rather than assuming English scores transfer.
Worked example: connect the method to the task
Ask a course assistant: “What is the refund deadline?” Without the current policy, the model may produce a plausible deadline from learned patterns. Provide a dated policy excerpt and ask for a source-supported answer with an explicit “not found” path. Then test cases where the policy omits the deadline or contains a conflicting older version.
This is a grounding and evaluation exercise, not proof that one prompt prevents all hallucinations. Check the answer against evidence and test Tamil paraphrases separately.
Apply the idea: check your reasoning
Think before revealing the answer
Does adding a policy to the prompt update the model’s learned weights?
Reveal the answer and reasoning
No. It changes the inference context for that request. Weight updates require a training process; whether the application stores the conversation is a separate product and data-handling question.
A common mistake: A token probability is not the probability that an entire factual statement is true.
Before you move on
- Explain the main idea without repeating the technical label.
- Identify the approach, its assumptions, and the evidence needed to trust the result.
- Complete the challenge, then compare your reasoning with the answer.
Sources & a deeper next step
The refresh uses primary papers and official documentation. These optional links are in English; you can understand this lesson without reading them first. Product versions and availability can change.