Stanford University AI model is like ChatGPT but open-source and $600 trained
robort - 2023-06-03 13:49:08
Many companies such as Google, Apple, Meta, Baidu, and Amazon are investing millions of dollars to keep up with OpenAI and its partner Microsoft who, with ChatGPT and Bing integration, aim to revolutionize the way the user interfaces with IT resources.
But, according to researchers at Stanford University, such a large outlay is not strictly necessary.
In fact, they have structured a new model that aims to achieve an accuracy comparable to that of ChatGPT while minimizing costs.
In order to make an artificial intelligence system like this work, in fact, resources and money must be spent on its training: the latter, in fact, constitutes their main competitive advantage.
GPT models managed to wow everyone thanks to the huge amount of time OpenAI spent post-training. It is one thing to have read a billion books, but quite another to have analyzed large quantities of question-and-answer pairs.
This type of data, in fact, allows the AI to understand what is being asked of it and to provide comprehensive answers.
Stanford University has tried to find a way to streamline the training work, using GPT-3.5 to provide LLaMA 7B with the data it needs to be able to do its job, i.e. the sufficient amount of question and answer pairs.
LLaMA 7B is Meta's open-source language model, the smallest and cheapest of the several available LLaMA models that the Facebook company makes available free of charge for academic projects.
The Stanford team started with GPT, asking to take 175 human-written input/output pairs and start generating more in the same style and format. This process was automated through one of the utility APIs provided by OpenAI and, in a short time, the team thus obtained approximately 52,000 conversations.
Then, this set of conversations was fed to LLaMA 7B, a process that took about three hours to process on eight computers made available via the cloud. These are NVIDIA A100-type systems with a compute cost of around $100 (while the OpenAI API training work cost around $500).
Finally, the researchers tested their resulting model, which they called Alpaca, in a variety of tasks including writing emails, social media, and productivity tools, and compared the results with those obtained by GPT extension. Alpaca passed 90 of these tests, and GPT passed 89.
Furthermore, it is worth noting that anyone wishing to replicate an AI now has access to even more complex models, such as GPT-4, as well as the software and hardware bundle to create real rivals to the reference models without breaking the bank.
In fact, the same researchers at Stanford University underline how it is possible to eliminate other cost items, such as that for processing via cloud computing. The training process, according to some, could be completed within five hours using a single NVIDIA GeForce RTX 4090 graphics card.
The availability of such powerful tools in everyone's hands could change many things in society as we know it today. It could give even more room for misinformation, leading to an even more confusing and difficult-to-interpret world.