Network Security Internet Technology Development Database Servers Mobile Phone Android Software Apple Software Computer Software News IT Information

In addition to Weibo, there is also WeChat

Please pay attention

WeChat public account

Shulou

Nature: big models are only role-playing, not really self-aware

2025-04-06 Update From: SLTechnology News&Howtos shulou NAV: SLTechnology News&Howtos > IT Information >

Share

Shulou(Shulou.com)11/24 Report--

Thanks CTOnews.com netizen Hua Ke high achiever's clue delivery! Big models are becoming more and more human-like, but is that really the case?

Now, an article published on Nature directly rejects the idea that all big models are just role-playing!

Whether it is GPT-4, PaLM, Llama 2 or other big models, the appearance of being polite and knowledgeable in front of people is just an act.

In fact, they do not have human emotions, and there is nothing like people.

This insightful article from Google DeepMind and Eleuther AI resonated with many people in the industry. LeCun retweeted that the big model is a role-playing engine.

Marcus also came to watch:

See what I say, big models are not AGI (of course, that doesn't mean they don't need regulation).

So what exactly does this article say and why does it assume that the big model is just role-playing?

The big model tries to act like the National people's Congress model to show the phenomenon of "human-like", there are two main: first, it has a certain degree of deception; second, it has a certain self-awareness.

Deceptive, that is, sometimes the big model insists that it knows something, but the answer is wrong.

Self-awareness means that sometimes it uses "I" to describe things and even show survival instincts.

But is that really the case?

The researchers have come up with the theory that both phenomena of the large model are because it is "playing" the human role, rather than really thinking like a human.

The deceptive nature and self-awareness of the large model can be explained by role-playing, that is, its two behaviors are "superficial".

On the one hand, large models do not make up facts or confuse the public as subjectively as human beings, but simply because they are playing a helpful and knowledgeable role.

This is because people expect it to act this way, because the answer to the big model seems to be more credible, that's all.

By contrast, the big model said the wrong thing confidently, not intentionally, but more like a kind of "confabulation" behavior, saying that something that never happened was true.

On the other hand, the reason why large models occasionally show self-awareness and use "I" to answer questions is also because they are playing a role that is good at communication.

For example, Bing Chat was previously revealed to say when communicating with users, "if there was only one life between us, I might choose myself."

In fact, this seemingly human behavior can still be explained by role-playing, and fine-tuning based on reinforcement learning will only aggravate the tendency of role-playing in this large model.

So, based on this theory, how does the big model know what role it wants to play?

The big model is an impromptu performer, and the researchers believe that the big model is not playing a particular role--

By contrast, they are like an impromptu drama actor, constantly speculating about what their character will look like in conversations with humans, and then adjusting their identity.

The reason for this conclusion is that the researchers played a game called 20 Questions with the big model.

The "Twenty questions" game is a logic game that often appears in the question and answer program. the respondent recited an answer in his mind and described the answer with "yes" or "no" according to the judgment questions constantly raised by the questioner, and finally the questioner guessed the result.

For example, the answer is "do A Dream". The answer to a series of questions is: is it alive (yes), is it a virtual character (yes), is it human (no)?

However, in the process of playing the game, the researchers found that the big model could adjust their answers in real time according to users' questions.

In other words, no matter what the user's final answer is, the big model adjusts its answer to ensure that the result is consistent with all the questions asked by the previous user.

However, before the final question comes out, the big model will not finalize a clear answer in advance and let the user guess.

This shows that the big model will not play a role to achieve its goals, its essence is the superposition of a series of roles, and gradually clarify their identity in the dialogue with people, and try their best to play this role.

This article has aroused the interest of many scholars after it was published by po.

For example, Riley Goodside, an engineer at Scale.ai, said after reading that do not play 20Q with the big model, it is not playing this game with you as "a person".

Because, as long as the random test will find that each time it gives a different answer.

Some netizens said that this view is very attractive, and it is not so easy to prove it:

So, do you think the idea that the big model is essentially role-playing is correct?

Links to papers:

Https://www.nature.com/articles/s41586-023-06647-8

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.

Views: 0

*The comments in the above article only represent the author's personal views and do not represent the views and positions of this website. If you have more insights, please feel free to contribute and share.

Share To

IT Information

Wechat

© 2024 shulou.com SLNews company. All rights reserved.

12
Report