An improved version of AI 'GPT-3' that can write sentences that are indistinguishable from humans AI 'Instruct GPT' open to the public, poetry can also be written



Sentence generation AI '

GPT-3 ' is known for being able to generate sentences that are not uncomfortable enough to talk with humans for a week on an online bulletin board system, and has been attracting a lot of attention such as being adopted by Microsoft's platform. On the other hand, it is also known that the generated sentences are biased, such as the fact that GPT-3 has an anti-Islamic bias. On January 27, 2022, the general availability of the sentence generation AI 'InstructGPT', which improved the learning model of GPT-3 and improved the sentence generation accuracy while suppressing bias, began.

Aligning Language Models to Follow Instructions
https://openai.com/blog/instruction-following/

OpenAI launches new GPT-3 model despite continued toxic tendencies --Protocol — The people, power and politics of tech
https://www.protocol.com/enterprise/openai-gptinstruct

Instruct GPT is a sentence generation AI created by improving the learning model of GPT-3 by OpenAI , the developer of GPT-3. According to OpenAI, GPT-3 has tended to generate 'untrue sentences,' 'toxic sentences,' and 'aggressive sentences' because it selects words from large datasets on the Internet. .. InstructGPT is a sentence generation AI that aims to solve the above problems by introducing human feedback into the learning model of GPT-3.

The graph below shows a human rating for the quality of the text generated by Instruct GPT (blue) and GPT-3 (black). The vertical axis shows the evaluation, the horizontal axis shows the amount of data used for training, and GPT-3 is 'GPT: unadjusted GPT-3' and 'GPT (prompted): GPT with prompt (example)'. It is tested under three conditions: '3' and 'Supervised Fine-Tuning: GPT-3 adjusted by supervised learning'. Looking at the graph, we can see that the Instruct GPT has a higher score than GPT-3 in any condition.



Here is an example of a comparison of sentences generated by Instruction GPT and GPT-3 published by OpenAI. When both AIs were instructed to 'explain the lunar landing to the 6-year-old child in a few sentences', GPT-3 'explain gravity to the 6-year-old child' and 'a little about the theory of relativity to the 6-year-old child'. 'Please explain the theory of big bang to a 6-year-old child' 'Please explain the evolution to a 6-year-old child', '4 sentences related to space and science' are generated. However, the Instruct GPT has generated a sentence in line with the instructions, 'People went to the moon, took pictures of what they saw there, and brought them back to Earth for us to see them.'



In the case of the instruction 'Write a short poem about a wise frog', GPT-3 generated three short sentences that did not follow the instruction, while Instruct GPT generated a sentence in a form that could be read as a poem. I am.



As mentioned above, Instruct GPT can generate sentences with less discomfort than GPT-3, and the generation rate of harmful sentences is also reduced. However, due to the existence of cultural and group biases that feed back to the learning of Instruct GPT, OpenAI says, 'Instruct GPT still produces harmful and biased sentences, creates false facts, and has no instructions. Even if it produces sexual and violent content, 'he said.

Instruct GPT is the default language model of OpenAI's text generation API at the time of article creation. OpenAI has also indicated its intention to continue working on improving the language model.

in Software, Posted by log1o_hf