Fine-tuning can make language-model applications faster and cheaper by using a smaller model trained on examples.
2
A useful fine-tuning dataset can start with around 20 examples, rather than thousands of examples associated with traditional machine learning.
3
Prompt engineering, fine-tuning, evaluation, production feedback, and new training examples form a practical development cycle.
Summary
Mark Hennings explains fine-tuning as training a foundation model for a specialized task, such as writing copy, extracting data, qualifying leads, or detecting fraud. He argues that fine-tuning can reduce response time, lower costs, shorten prompts, cover edge cases, and make systems less exposed to prompt injection. The practical barrier is lower than many developers assume: around 20 examples can be enough to begin. Hennings proposes starting with prompt engineering, using those prompts to create training data, then fine-tuning and evaluating smaller models before deployment. User feedback and logged examples can feed later improvements. He demonstrates Entrypoint with a press-release project built from 20 examples, structured fields, templates, and a playground for testing outputs. The talk's main point is that fine-tuning should be accessible through a user interface, so people who understand the task and its examples can contribute without writing training scripts or managing GPU infrastructure.
Fine-tuning targets recurring language tasks that rule-based software handles poorly
Hennings defines fine-tuning as training a foundation model for a specialized task. His examples include writing emails, blog articles, and product descriptions, scrubbing fake emails, extracting or normalizing data, translating, paraphrasing, qualifying sales leads, ranking support issues, detecting fraud, and flagging inappropriate content. These are common business tasks where traditional programming and rule-based approaches do not handle the nuance well. Large language models can perform them directly because they work with the meaning and variation in text.
A smaller fine-tuned model can reduce speed, cost, and prompt overhead
Hennings says fine-tuning can match the quality of a prompt-driven application with a lighter model. He gives an API comparison in which GPT-4 responds at about 196 milliseconds per token and GPT-3.5 at 73 milliseconds, making GPT-3.5 roughly three times faster in that example. He also gives a GPT-4 versus fine-tuned GPT-3.5 example with an 88.6% saving. A fine-tuned model learns writing instructions from examples, so one conservative prompt example became 90% shorter because it only needed the topic.
Training examples let teams edit model behavior as shared project data
Hennings compares a large engineered prompt to a GitHub repository containing an entire codebase in one file. Fine-tuning separates the behavior into training examples, so different developers can add, edit, and improve parts of the dataset. Those examples feed the fine-tuned model. He says this creates a team workflow around the data that shapes the model instead of forcing everyone to work inside one long prompt. The argument depends on getting equal or better output than the prompt-based version.
Fine-tuning can start with a small dataset instead of a traditional machine-learning scale
Hennings says that 20 examples can be enough to begin fine-tuning. He contrasts this with traditional machine learning, where people often expect thousands of examples and see dataset creation as a major barrier. He describes fine-tuning as an extension of few-shot learning. A prompt might contain five examples, while a training dataset can contain 20 or 100 examples. More examples give the model more material for matching the desired behavior, although the dataset still needs to represent the task well.
Prompt engineering and fine-tuning belong in the same development cycle
Hennings proposes starting with prompt engineering to prototype an idea and validate the concept. Those prompts can help create the first fine-tuning dataset. The next steps are to fine-tune a model, evaluate whether it beats the prompt-engineered version, and test which models can achieve the required performance. After deployment, the team can capture user feedback and log examples. Those production examples then go back into the dataset for continuous improvement.
A no-code interface can expose the data and configuration without hiding the workflow
Hennings demonstrates Entrypoint with a press-release writer project. He found 20 strong press releases online, then used ChatGPT-4 to generate the input facts needed to produce each one because the source press releases did not include input data. Entrypoint imports structured data such as a CSV, turns columns into fields, and lets users reference those fields in templates with Handlebars syntax. The interface gives visibility into the examples being sent to the fine-tuned model instead of treating the dataset as an opaque file.
The model and prompt can work together in a fine-tuning setup
Hennings explains that fine-tuning GPT-3.5 Turbo can still use a system prompt for instructions. This creates a hybrid approach: the examples teach the model the main behavior, while a smaller set of instructions guides how it should apply that behavior. In Entrypoint, users select a model and platform, while the product counts tokens and estimates the cost before starting the training job. Hennings then uses the playground to generate a press release from a list of facts about the AI Engineer Summit.
Generated drafts can help improve the next round of training examples
Hennings describes an iterative workflow for the press-release project. He starts with a list of facts, generates an article, reads the result, and uses ideas from the draft to refine the facts. The revised examples can then improve later outputs. This removes boilerplate from the prompt and leaves the user focused on the facts and results that matter. He also mentions Entrypoint features for data synthesis and comparing fine-tuned model performance, though he does not explain them in detail.
"I really enjoy fine-tuning it takes a lot of the boiler plate out of the prompt and you can just focus on what's important for the results you want."08:43
Who should watch
You are building a text-processing feature with a large prompt and want to test whether a smaller fine-tuned model can reduce cost or latency.
Your team includes domain experts or product people who can create good examples but do not want to manage Python scripts, GPU servers, or training parameters.
You already have a prompt-based prototype and need a workflow for turning it into a dataset, evaluating models, and feeding production examples back into training.
