AI master Karpathy personally teaches the use guide of ChatGPT model: o3 is suitable for handling complex tasks, 4o is competent for daily questions, and GPT-4.1 is used for code writing. Eric Hayes reveals the core memory system behind ChatGPT, which makes ChatGPT understand you more and more.
You must have used ChatGPT and enjoyed it very much, but there is a high probability that you are using it incorrectly!
Don’t worry, this is what Andrej Karpathy, the founder of OpenAI, said. He is "passionate" about popularizing AI and has made another move.
This time Andrej Karpathy uses a picture to teach you how to choose the ChatGPT model!
As we all know, OpenAI has given the model a lot of names: GPT-4o, o3, o4-mini, o4-mini-high, o1 pro mode...
It's so dazzling that people can't tell the difference between these models.
Even because of the inability to choose the right model, daily use is not worth the $20 Plus "price" per month, not to mention the $200 Pro membership.
Many people don’t know that these models of ChatGPT are very different from each other, for example:
o3 is the best choice for dealing with important or difficult problems.
It is a model with very strong reasoning ability, more powerful than 4o. If you use ChatGPT professionally and don't use o3, you will probably suffer a loss.
4o and o4 are completely different things.
Even the great Karpathy had to joke that OpenAI's naming is really confusing.
4o is a good choice for daily workhorse to cope with simple to medium difficulty problems. However, o4 currently only has a mini version, which is not as good as o3.
Don't use these models!
Karpathy also emphasized that models like o4-mini, o4-mini-high and o1-pro should not be used, as there is no benefit at all!
Karpathy said he couldn’t even understand why OpenAI released o4-mini now.
AI experts teach you how to play ChatGPT
Based on his experience and the differences between different models of ChatGPT, Karpathy provides the following "User Guide".
Any simple question
For example, "Which foods are rich in dietary fiber?" => Use 4o (about 40% of what Karpathy uses)
Any difficult or important problem, as long as I am willing to wait a little longer
Like "Help me understand this tax issue...") => Use o3 (also about 40% of Karpathy's usage)
When writing code casually or changing the code as you think
For example, “Change this code to…” => Use GPT-4.1 (about 10%)
Want to understand a topic in depth?
I hope GPT spends 10 minutes, checks a lot of links and summarizes a complete set of information.
For example, “Help me understand the rise and fall of Luminar” => Deep Research (about 10%)
Note: Deep Research is not a version in the model selector (!!!), it is a function turned on in the Tools tool.
At the bottom level it is based on o3, but Karpathy believes that it is not exactly the same as using o3 directly (although Karpathy is not sure).
If you master the correct usage of the model given by Karpathy, you will find that ChatGPT is getting better and better.
And the more you use it, you will find that ChatGPT understands you more and more, as if it has become your "Jarvis" and remembers everything about your daily life.
The reason behind this is actually because ChatGPT has a super memory, which is why everyone still loves to use ChatGPT despite its "confusing" naming, and it is the AI tool with the highest weekly activity, without a doubt.
Unveiling ChatGPT’s “Super Memory”
ChatGPT gets to know you better and better, seemingly remembering your past preferences and even connecting context across multiple conversations.
There is more to this than just a simple context window; there is a complex and sophisticated memory system at work.
Recently, Eric Hayes, an engineer from a startup company, revealed the complex memory system behind ChatGPT through a blog post.
Today, follow Eric Hayes to analyze in depth the "memory magic" that makes ChatGPT stand out!
ChatGPT’s memory system makes its experience far superior to other large language models (LLMs).
This is mainly due to the fact that it divides the memory system into two categories: "Saved Memory" and "Chat History".
Saved Memory
Imagine that you tell ChatGPT your preferences and it will remember them. This is the function of "saving memory".
It’s a simple, user-controllable system where you can update these memories with explicit instructions like “Remember that I’m a software engineer” or “When I recommend a recipe, remember that I’m a vegetarian.”
These memories are injected into system prompts as facts that influence subsequent conversations. For example, you can let ChatGPT remember that you like concise responses or that you are an expert in Rust programming, and it will adjust its responses based on these preferences.
ChatGPT provides a simple user interface that allows you to view and delete these memories. You can also ask ChatGPT to delete the saved memories through commands.
This feature is also intelligent. It performs minimal duplication and inconsistency checks.
For example, if you tell it “I’m a software engineer,” it might save it, but if you say “I’m not a software engineer,” it will reject it and ask you to clarify.
However, for highly related but different information (such as "software engineer", "front-end engineer"), it is allowed to coexist.
This system is implemented through a tool called "bio tool". Eric Hayes reversed the memory system of ChatGPT and discovered this method called bio.
Chat History
The "chat history" system is much more complicated than "saving memory", and it may be the key to improving ChatGPT's responsiveness. It can be divided into three subsystems:
1. Current Session History
This is like a short-term "short-term memory" for the bot. It records the user's most recent messages in the current session.
This log is small, containing only messages from the last day, usually less than 10.
It can directly reference the message sent by the user in the current session.
2. Conversation History
This system is responsible for remembering relevant context from your past conversations with ChatGPT.
And it can perform cross-conversation memory, ChatGPT can reference the direct messages you sent in other conversations.
Eric Hayes found in testing that ChatGPT could accurately quote messages within two weeks.
After two weeks, it will usually only provide a summary of your message rather than a direct quote.
Message retrieval is indexed by "conversation summary" and "message content".
This means that even if you can’t remember the exact words, ChatGPT can find relevant information as long as you describe the content or topic of the conversation.
For older conversations, the system might store inferred information, providing smaller, less specific context.
3. User Insights
This may be the most powerful and "smart" form of memory for ChatGPT!
It's a more advanced, more opaque version of "preserving memory."
These insights are automatically generated by the system by analyzing user behavior and questions across multiple conversations. They will include a time range and confidence level.
User insights can capture in-depth information such as your professional areas, interest preferences, and questioning style.
For example, it might record that you have "extensive experience and knowledge of Rust programming" or "prefer concise answers."
With these detailed insights, ChatGPT is able to minimize frustrating interactions and present information in a way that users can easily understand.
Eric Hayes believes that the user insight system may be responsible for more than 80% of ChatGPT’s perceived intelligence improvement!
Eric Hayes also found that these insights might not be generated in real time, but updated through batch processing (such as a cron job that runs once a week).
It extracts unique and useful insights from a large amount of historical messages by modeling user queries as a clustering optimization problem.
What does this mean for user experience?
It is these complex memory systems that make ChatGPT feel so "easy to use"!
Save memory allows you to set preferences directly and let ChatGPT tailor responses for you.
User insights automate this process. Even if you don't tell it explicitly, ChatGPT can automatically adjust its response based on your historical behavior, provide more appropriate explanations, and avoid repeated questions.
The chat history system ensures the continuity of the conversation, allowing you to review old topics while ChatGPT maintains shared knowledge of previous interactions, avoiding repeated, circular, or contradictory interactions.
It can be said that ChatGPT's memory system goes beyond a simple context window and greatly improves the user experience by storing and retrieving information in a hierarchical and intelligent manner, making it feel more like an intelligent partner that can learn and grow.
References
https://x.com/karpathy/status/1929597620969951434
https://macro.com/app/md/54115a42-3409-4f5b-9120-f144d3ecd23a
This article comes from the WeChat public account "Xinzhiyuan" , author: Dinghui, and is authorized to be published by 36Kr.
