• btc = $71 172.00 675.70 (0.96 %)

  • eth = $3 552.43 -6.45 (-0.18 %)

  • ton = $7.23 -0.04 (-0.57 %)

  • btc = $71 172.00 675.70 (0.96 %)

  • eth = $3 552.43 -6.45 (-0.18 %)

  • ton = $7.23 -0.04 (-0.57 %)

21 Mar, 2023
2 min time to read

Alpaca AI, developed by Stanford, achieves comparable performance to the remarkable ChatGPT on various tasks, yet it is constructed using an open-source language model and was trained at a cost of under $600.

Six months ago, only experts and scientists followed the progress of large language models. However, the launch of ChatGPT late last year has brought attention to the potential of machines that can communicate in a manner indistinguishable from humans. They can produce high-quality text and even program code across a wide range of subject areas in seconds. The launch of GPT-4 indicates that these models are improving at an astonishing rate and have the potential to revolutionize human society by automating tasks that were previously thought impossible, particularly among white-collar workers.

Many other companies, such as Google, Apple, Meta, Baidu, and Amazon, are not far behind, and their AIs will soon be available in every possible application and device. Language models are already integrated into Bing, and soon they will be present in robots, cars, phones, TVs, and customer service.

Despite the benefits of these models, they also have the potential for misuse, including spam, misinformation, and malware creation. OpenAI and other large companies have been working to limit these capabilities manually before the launch of their models. However, there is concern that governments are not moving quickly enough to regulate AI for the public good.

The Stanford research team has built their own language model, Alpaca, for less than $600. They used Meta's open-source LLaMA 7B language model as a basis and generated more human-written instruction/output pairs using OpenAI's APIs. They then fine-tuned the LLaMA model with the generated data and tested it against ChatGPT's language model. Alpaca won 90 of the tests, while GPT won 89. The team released the questions and code used for the research on Github and noted that they have not yet fine-tuned the Alpaca model for safety and ethics.

While anyone can now create their own AI, OpenAI's terms of service prohibit using their output to develop models that compete with OpenAI. Meta is also currently only allowing academic researchers to use LLaMA under non-commercial licenses.