Recently, I came across a very interesting point of view.

Richard Sutton, one of the founders of reinforcement learning and a 2024 Turing Award winner, made a sharp observation about large language models in an interview:

Today's ChatGPT and large language models are powerful, but they may not be the final path toward real intelligence.

Some articles summarize this as "Sutton criticizes ChatGPT."

That headline is dramatic, but after reading the original discussion, I think a more accurate interpretation is this:

Sutton is not criticizing one chatbot. He is questioning an AI path that relies mainly on human-written text and learns primarily by imitating human answers.

This idea is very useful for parents and students.

It reminds us that in the AI era, children should not only learn how to ask ChatGPT questions. They also need to understand what real learning, intelligence, and problem solving mean.

First, What Did Sutton Actually Say?

The discussion mainly comes from Dwarkesh Podcast's interview with Richard Sutton, titled:

"Richard Sutton – Father of RL thinks LLMs are a dead end"

Link: https://www.dwarkesh.com/p/richard-sutton

Here, "RL" means reinforcement learning.

Sutton is one of the most important researchers in reinforcement learning. According to ACM, Richard Sutton and Andrew Barto received the 2024 ACM A.M. Turing Award for their foundational contributions to the concepts and algorithms of reinforcement learning.

ACM announcement: https://awards.acm.org/about/2024-turing

So his view on the future direction of AI is not casual criticism. It comes from decades of research into how intelligence learns.

Why Does He Think LLMs Are Not the End Point?

The capabilities of large language models are impressive.

They can write essays, generate code, translate, summarize, create lesson plans, explain Python code, and help students debug errors.

But Sutton is focused on a deeper question:

These models mainly learn from text that humans have already written.

In other words, they are very good at learning patterns from human language.

But can real intelligence emerge from imitating text alone?

Sutton is clearly skeptical.

What he values more is whether an agent can act in a real environment, observe outcomes, receive feedback, adjust its strategy, and improve through long-term experience.

That is the core idea of reinforcement learning:

Learning is not only reading answers written by others. It is learning through action and feedback.

The Era of Experience: AI Should Not Only Consume Human-Written Data

This view also connects to David Silver and Richard Sutton's article The Era of Experience.

Article link: https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf

The article makes an important argument: AI is moving from an era that depends heavily on human data toward an era that depends more on real experience.

Simply put, stronger future AI should not only keep learning from what humans have already written. It should be able to:

• Act on its own
• Observe the environment
• Receive feedback
• Form goals
• Adjust strategies
• Keep learning through long-term experience

This is very different from how many people use ChatGPT today.

Many people treat AI as an "answer machine."

They type in a question and expect a complete answer immediately.

But Sutton's perspective reminds us that real intelligence is not only about producing answers. It is about improving behavior through experience.

What Does This Mean for Children Learning AI?

In education, this idea matters a great deal.

Many children currently use AI like this:

"Help me write this essay."

"Help me write this code."

"What is the answer to this problem?"

"Just tell me how to do it."

Of course, this is convenient.

But if children only use AI as an answer source, they have not completed the real learning process.

Real learning is closer to reinforcement learning:

First, propose your own idea.

Try a possible solution.

Observe the result.

Find the mistake.

Adjust the strategy.

Try again.

Through this process, students gradually build judgment and problem-solving ability.

This is why I do not recommend that children ask ChatGPT for complete answers right away.

AI can be used, but it should not do the thinking for the child.

Programming Trains This Ability to Learn from Feedback

Why do we still encourage children to learn Python, algorithms, and programming in the AI era?

One important reason is that programming naturally gives students the experience of "action, feedback, and revision."

When a child writes code, that is an action.

When the program runs, the result is feedback.

An error message, a wrong output, or a failed sample case is a signal from the environment.

The student then needs to ask:

• Did I misunderstand the problem?
• Did I update a variable incorrectly?
• Is there an off-by-one error in the loop?
• Is the algorithm too slow?
• Did I miss an edge case?

This process is similar in spirit to how intelligent systems learn.

It is not about memorizing answers. It is about becoming smarter through repeated attempts, testing, and revision.

So the value of a programming class is not only teaching Python syntax.

More importantly, it trains students to:

• Describe problems clearly
• Design step-by-step solutions
• Test results
• Locate errors
• Improve based on feedback

These are foundational abilities in the AI era.

Sutton's "Bitter Lesson" Is Also Important for Parents

Sutton also wrote a well-known essay called The Bitter Lesson.

Original article: http://www.incompleteideas.net/IncIdeas/BitterLesson.html

The main idea is that, over the long run, general methods that scale with computation tend to outperform human-designed domain-specific approaches in AI.

This idea is often used to explain why deep learning, large-scale training, and general algorithms have achieved so much.

But for education, I think it has another implication:

Children cannot rely on memorizing fixed answers to face a changing future.

They need more general abilities:

• Learning how to learn
• Learning how to experiment
• Learning how to verify
• Learning how to abstract
• Learning how to use tools for real problems

These abilities matter more than simply memorizing how to use a particular tool.

So How Should Children Use ChatGPT?

I do not believe children should be banned from using ChatGPT.

On the contrary, I believe they should learn how to use AI correctly.

But the way they use it should change.

Do not ask:

"Please give me the answer directly."

Ask instead:

"Here is my thinking. Can you help me check whether there are any gaps?"

Do not ask:

"Write the complete code for me."

Ask instead:

"Please give me hints only. Do not give the full code. I want to try it myself first."

Do not ask:

"How do I solve this problem?"

Ask instead:

"I tried this approach, but sample case 2 does not pass. Which part should I inspect first?"

In other words, AI should not be the final destination. It should be a feedback tool.

Children should not hand over the learning process. They should use AI to complete the learning process more effectively.

A Simple Standard for Parents

After a child uses AI, parents can ask four questions:

First, did the child write down their own idea first?

If the child asks for a complete answer immediately, the learning value is low.

Second, can the child explain where they are stuck?

"I don't know" is too vague. "I don't know why my loop runs one time too few" is a useful help request.

Third, did the child ask AI for hints instead of asking it to do the work?

That shows the child still owns the problem-solving process.

Fourth, can the child restate and rewrite AI's suggestion?

If the child only copies and pastes, it is hard to develop real understanding.

These questions are more important than simply asking whether the child used AI.

What Kind of Learners Do We Really Want to Build?

ChatGPT is powerful, but it is not a replacement for learning.

Sutton's critique of the LLM path reminds us of something important:

Real intelligence is not only imitation. It is the ability to act, receive feedback, revise, and grow in the real world.

Children are the same.

A child who only copies AI answers may look efficient, but the ability has not truly developed.

A child who can ask questions, design attempts, analyze feedback, and keep improving is genuinely learning.

This is why Accel Thinking emphasizes:

In the AI era, children need programming, algorithmic thinking, and AI literacy more than ever.

Not because they need to compete with AI on who writes code faster.

But because they need to understand the boundaries of AI, ask better questions, judge whether results are reliable, and turn tools into their own abilities.

Conclusion

If we apply Sutton's view to youth education, I would summarize it this way:

Do not treat AI as an answer machine. Treat AI as a feedback tool.

Do not let children skip thinking. Help them grow through attempts, feedback, and revision.

Do not only teach children how to ask ChatGPT. Teach them how to verify, improve, and solve real problems.

That is the kind of learning children need most in the AI era.

If you want to understand whether your child is ready to start learning Python, or what stage their programming ability is currently at, visit Accel Thinking:

https://accelthinking.com/

You can also start with our Python Programming Assessment:

https://accelthinking.com/python-assessment/

We will continue sharing content on Python learning, CCC competitions, algorithmic thinking, and AI literacy to help students move from "using tools" to "solving problems."