A popular strategy for interacting with generative AI chatbots is to start with a well-crafted message. In fact, message engineering is an emerging skill for those who Pursuing career advancement in this era of artificial intelligence.
However, there is an alternative. For developers who have the budget to invest in developing large language models (LLMs) and a lot of custom data of their own, “fine-tuning” an AI model can, in some cases, be a superior approach.
But fine-tuning can be expensive, and the good news is that OpenAI announced Tuesday that it is offering drastically cheaper fine-tuning for its GPT-4o mini AI model. introduced last week.
Also: OpenAI offers GPT-4o mini to reduce application costs
A fine-tuning process involves subjecting an AI model to a new round of training after the initial training of the model. By loading some data and running the training again, the neural “weights” (or “parameters”) of the model are modified relative to the standard version of the model.
The result is a model that can place more emphasis on the data in the new training dataset when asked for predictions than is the case with the traditional model.
A neural network like GPT-4o mini reflects a probability distribution, and its output (i.e., its predictions) is simply the most likely text that follows the user’s input. Fine-tuning shifts that probability distribution in a certain direction. As a result, the model’s responses also shift to reflect the changed probability distribution.
Fine-tuning is therefore a means of pushing the message in the direction one desires.
The cost of tuning GPT-4o mini starts at $3 per million tokens used for training. According to the OpenAI pricing guideThat’s less than half the $8 cost of the GPT-3.5 “Turbo.”
OpenAI is offering a free two million tokens per day to qualified institutions until September 23.
Also: Millennial men are most likely to enroll in AI training courses, report says
However, note that the price of an optimized GPT-4o mini is double the price of the generic GPT-4o mini, at 30 cents per million input tokens to the model and $1.20 per million output tokens, i.e. the tokens it uses to generate and then receive predictions.
In addition to the cost advantage, OpenAI emphasizes that the amount of training data that can be fed into the model for fine-tuning is four times larger than for GPT-3.5, 65,000 tokens.
Please note that fine-tuning is only available for GPT-4o mini’s text functionality, not its imaging tasks.
Before making adjustments, it’s worth considering other options. Continuing to refine messages is still a good strategy, especially since refined messages can be helpful even after the model has been fine-tuned, according to OpenAI’s documentation. Fine-tuning documentation.
Another approach to getting more personalized LLM results is to use “Generation increased by recovery (RAG), an increasingly popular engineering approach that involves the model making calls to an external source of truth, such as a database.
While RAG can make each query more complicated, in a sense, by requiring the model to call the database, it also has advantages. When fine-tuning a model, it is possible for the model to unlearn what it acquired in the original training stage. In other words, manipulating model parameters can produce setbacks in terms of the broader, more general functionality that a model possesses.
Also: Make room for RAG: How the AI generation’s balance of power is shifting
A third alternative, in addition to prompt engineering and RAG (but closely related to the latter), is function invocation. In such cases, very specific questions can be put into the prompt and a very specific form of response can be requested, which can be grouped together and sent to an external application as a function invocation. OpenAI and others refer to this as function invocation, tooling, and “agent AI.”
All of these approaches will find their place, but at least fine-tuning experiments will cost a bit less under OpenAI’s new pricing.
Note that Google also offers fine-tuning for its models, through its Vertex AI program, and many other model providers do as well.
Retraining models are likely to become more common and Maybe it will even come to mobile devices one daywith sufficient computing power.