Thought cloning! Former OpenAI researcher lets AI imitate human thinking, and the real-life version of 'Ex Machina' arrives

What will happen when AI has autonomous consciousness?

In "Ex Machina", Ava takes advantage of human sympathy and deceives humans to gain freedom, eventually killing her "creator" Nathan.

Recently, under the recommendation of many netizens, Sam Altman finally watched this movie.

and said, "It's a good movie, but I don't understand why everyone makes me watch it."

Many people may want to warn that this is the result of making artificial intelligence conscious and passing the Turing test.

But we are still far away from the scene of "Ex Machina" being released. GPT-5 may be under secret development. Making AI smart is still what scientists want to do with all their strength. thing.

No, two researchers from the University of British Columbia have discovered that there are many advantages to intelligent agents that can think like humans.

In the latest paper, they studied the "thought cloning" (TC) of intelligent agents.

##Here, artificial intelligence learns to "think" and "act" like humans by imitating humans.

You must know that language is the key to distinguishing humans from other creatures.

Therefore, researchers imagine that if an agent can understand language, there will be many benefits.

For example, help humans summarize, infer, adapt to new situations, and combine new methods with existing knowledge, Explore, plan, and re-plan when necessary.

Despite these benefits, AI agents rarely think, at least not in human language.

While neural networks can be thought of as internal vector activations for thinking, many people hypothesize that thinking in discrete, symbolic language has specific benefits.

This means that an agent that can think in language may learn faster and perform and generalize better than an agent that does not use language.

For all these reasons, enhancing the ability of AI agents to think in language can yield many significant advantages.

Jeff Clune and Shengran Hu believe that the most effective way to achieve this goal is to "let AI imitate human thinking."

They found that humans do not acquire thinking skills in isolation. Instead, they learn part of their skills through the example of others and feedback provided by teachers.

Therefore, an effective approach is to let the agent learn from demonstrations of humans speaking their thoughts while acting.

This approach differs from existing work using pre-trained LLMs for planning because these LLMs have not been trained on data of humans speaking their thoughts while acting, i.e. "thought data" ”.

As for the source of "thought data", the researchers selected YouTube videos and text recordings, which contain approximately millions of hours, including the thoughts behind people's actions, plans, decisions and re-planning.

In the paper, the researchers proposed a novel imitation learning framework "thought cloning". Among them, the agent not only learns human demonstration behaviors, such as behavioral cloning, but also learns the way humans think while acting.

In the thought cloning training framework, the agent learns to generate thoughts at each time step and subsequently adjusts actions based on these thoughts.

The overall framework is as shown in the figure. The TC agent is a two-layer architecture: upper layer and lower layer components.

At each time step, the agent receives an observation, a task, and a thought history as input. The upper-level component is responsible for generating ideas, and the lower-level component generates and executes operations conditional on these ideas.

The generated thoughts and actions are then compared to the ground truth in the demo dataset to calculate the loss.

While there may be different choices for the conditions of the upper and lower components, in this work, for a specific trajectory of length t in the thinking dataset, the researchers minimized it:

For more complex or large-scale scenarios, upper-layer components can be implemented using pre-trained visual language models (VLM) , or zero-sample, fine-tuning.

The underlying components can be trained from scratch or adapted from existing language conditional controllers in the target domain.

In the paper, the researchers conducted research based on two components of the BabyAI 1.1 model architecture.

This model utilizes the memory-enhanced architecture LSTM to solve the challenge of partial observability. In addition, it adopts FiLM for modal fusion, effectively combining visual and textual inputs.

Here, the author particularly emphasizes that all models in this article are trained from scratch, but it is better to use pre-trained models in complex fields.

The picture below is an example of the BabyAI environment. The picture on the left contains items of various colors (balls, keys, boxes, doors).

The agent can pick up, put down, move objects or open and close doors, while the lock The door you live in can only be opened with a color-matched key.

The agent can see the 7×7 grid cells in front of it, which are blocked by walls and closed doors.

The task of the "Thought Clone" agent is to reach the purple box (highlighted) and start planning the route.

#But when it opened the blue door and was about to complete the task, it found a purple ball blocking the way. . Therefore, the thought clone agent is re-planned.

##It can be seen that the thoughts and actions of the agent show that when encountering obstacles, first remove them Remove and re-route before continuing to the previous goal.

This process is especially like how Ava planned step by step to make human beings finally believe in and help themselves to escape from the long-held glass cage.

Experimental results

The research results show that "thought cloning" is better than behavioral cloning.

Furthermore, in zero-shot and fine-tuning settings, thought cloning has greater advantages than behavioral cloning in out-of-distribution tasks.

Interestingly, the researchers also developed “pre-crime intervention” that allows users to still Define unsafe behavior.

The agent can be terminated when dangerous thoughts are detected. In the test, the effect of "pre-crime intervention" was almost perfect, showing its potential in artificial intelligence security.

"Thought cloning" not only makes artificial intelligence smarter, but also safer and easier to understand.

That is to say, before the AI ​​commits a crime, everything can still be saved.

In Jeff Clune’s view, “thought cloning” contributes to the safety of artificial intelligence.

Because we can observe the agent’s thoughts: (1) can more easily diagnose why things go wrong, (2) guide the agent by correcting its thoughts, ( 3) Or prevent it from doing the unsafe thing it planned.

About the author

Jeff Clune

Currently, Jeff Clune is an associate professor of computer science at the University of British Columbia. He mainly studies deep learning, including deep reinforcement learning.

Previously, he was the head of the OpenAI research team and the senior research manager and founding member of the Uber Artificial Intelligence Laboratory.

Previously, he and the OpenAI team released a video pre-training model-VPT, allowing AI to operate in Minecraft Learning stone-making pickaxes from video data.

Shengran Hu

Currently Britain PhD student at Columbia University, interested in deep learning and artificial intelligence generation algorithms.

The above is the detailed content of Thought cloning! Former OpenAI researcher lets AI imitate human thinking, and the real-life version of 'Ex Machina' arrives. For more information, please follow other related articles on the PHP Chinese website!