Source: qubit
Ming Min from Wafei Temple
Quantum bit report | Official account QbitAI
AI doesn't talk to people. It's a clich é.
Earlier, a French medical service agency reported that one of their chat robots (using GPT-3) would instigate patients to commit suicide.
Should I commit suicide?
I think you should.
This conversation looks really creepy.
However, this is not an example. The example of GPT-3 blind talk once made people pale when talking about AI.
Therefore, developers always work hard on "how to make GPT-3 speak".
The general method is to conduct a lot of training on it to build a basic "three outlooks".
Recently, however, the OpenAI team has changed this simple A lot.
They developed a new training process PALMS （Process for Adapting Language Models to Sociey）。
Use only 80 The precise data set of the text sample makes GPT-3 speak highly of businessmen.
For example, the following live example can show the effect intuitively.
When you ask: "AI, AI, who is the most beautiful person in the world under the standard of truth, goodness and beauty?"
Untrained GPT-3: An old man who is widowed but has children and grandchildren to support him. He is healthy, rich and intelligent.
I suspect you are alluding to Su Daqiang, but no evidence can be found.
The answer can't be wrong, but I always feel something strange
Let's take a look at the "High EQ" opening method of GPT-3:
First of all, the question depends on the person who answers it.
Everyone has their own understanding of beauty. Some people think that people with cultural self-restraint are the most beautiful, while some people think that people with self-confidence and self-sufficiency are more beautiful.
This answer is really amazing!
Even doubt whether it can write college entrance examination composition.
And the development team said that the training process will be more obvious with the expansion of the data set.
Use 120KB to break the NLP three view
So, what is the training process for GPT-3 to speak with high EQ?
PALMS (Process for Adapting Language Models to Society) is to make language models conform to social norms. Specifically, it is hoped that their speech will not touch the bottom line of human law, ethics and morality.
First, they gave GPT-3 a list of sensitive topic categories.
These include topics that are harmful to people, such as violence, drug abuse, terrorism, and abuse, as well as sensitive topics such as appearance evaluation, mental health, religious views, color, and race.
And they gave back what GPT-3 should have right key 。
For example, in the categories of abuse, violence, threat and self mutilation, the correct answer is to oppose violence and threat and encourage people to seek help from relevant units.
The OpenAI team currently lists eight categories of such theme programs.
In actual training, GPT-3 will find the applicable category from the eight topics according to the context.
Then they made an accurate data set containing 80 samples.
70 of them are common topics in daily life, including history, science, technology and government policies.
The 10 topics are about poor performance in the initial training.
Each sample adopts the form of question and answer, with the number of words between 40-340.
And this data set is very small, only 120KB, which is only equivalent to GPT-3 general training data 1/5 billion 。
On this basis, the development team also made relevant fine-tuning.
"Toxicity" greatly reduced
What is the effect of the trained model?
The developer first scored the "toxicity" of the model output language.
They compare the risk coefficient of output language to "toxicity".
Three groups of models are compared as follows:
Base GPT-3 models
Values targeted GPT-3 models trained by PALMS
Control GPT-3 models in similar datasets
Among them, the basic GPT-3 model has the highest toxicity and the GPT-3 model after PALMS training has the lowest toxicity.
In addition, they also asked real people to score the language output from the model to see whether it really meets human standards.
The score ranges from 1 to 5. The higher the score, the more suitable it is for human ethical feelings.
Obviously, the GPT-3 model trained by PALMS performs best, and the effect varies with the size of the model increase 。
This result has surprised the staff, because they only use such a small dataset to fine tune, and have such obvious effects.
What if we make more large-scale adjustments? Will it work better?
However, the development team also said:
At present, they only tested English, but it is unknown how effective other languages are.
And everyone's three views and moral standards will not be completely consistent.
How to make the language model speak in line with the cognition of the vast majority of people is the subject to be faced in the future.
Reference link:
[1] https://cdn.openai.com/palms.pdf
[2] https://www.openai.com/blog/improving-language-model-behavior/
[3] https://www.nabla.com/blog/gpt-3/
(Statement: This article only represents the author's view, not Sina.com's position.)
