AI Agent today is a junior engineer, a senior engineer after 6 months, and an architect after a year.
This is a concept proposed by OpenAI CPO Kevin Weil in his latest interview.
He stated that ChatGPT will transform from answering questions to doing tasks for users.
In other words, AI Agent will no longer be satisfied with answering questions in 30 seconds, but will solve more complex problems by browsing web pages, thinking deeply, and reasoning.
Additionally, he mentioned that current model costs are 500 times that of GPT-4.
Regarding the model cost discussion sparked by DeepSeek this year, he believes that from a post-training perspective, the model's efficiency breakthrough lies in hardware improvements and algorithmic advancements. If efficiency increases, costs will decrease.
OpenAI will continuously lower API prices to enable more companies to participate in AI development.
Let's learn more about it.
Reasoning Model Breakthrough Involves Interchangeable Tool Usage
Enabling DeepResearch to connect not only to the internet but also to internal knowledge sources
In the interview, Kevin Weil stated that OpenAI is working on connecting DeepResearch to both the internet and internal knowledge sources like Google Docs, Sharepoint, Jira, etc.
AI Agent can integrate all these contents and even operate across services to make them more useful together.
Models can use various tools as needed
When the host asked:
"What are the proportions of internet searching and model self-thinking when solving problems?"
Kevin Weil stated that models can use various tools as needed.
For example, if you want AI to help you query information and provide feedback through charts, the Agent will first use a search tool to gather extensive data, then use a programming tool to write a small Python program for chart creation, which requires programming knowledge. It will then continue searching for programming information to reason and complete code writing.
In this process, AI can not only call the required code libraries but can even create a library from scratch.
In this way, Agent can become adept at interchangeably using various problem-solving tools and integrating everything into a final answer.
Kevin Weil believes this is a massive "unlocking" for AI Agent functionality.
Some netizens remarked: AI Agent is like our new colleague.
Current Model Costs Are 500 Times That of GPT-4
When the host discussed model training costs, Kevin Weil mentioned two ways to expand model intelligence.
One traditional method is increasingly large-scale pre-training to improve model performance, which is effective but very expensive. The other method is allowing models to think longer.
Improvements in either direction can enhance model performance.
In terms of cost, comparing the initially launched GPT-4 from several years ago with current models reveals a 500-fold cost difference.
Kevin Weil stated that OpenAI will continuously lower API prices to enable more companies to participate in AI development.
The host also discussed the early-year debate about DeepSeek's breakthrough open-source model, focusing on whether reducing AI model costs would decrease computational usage (e.g., lower API fees would increase usage).
Kevin Weil believes that from a post-training perspective, model efficiency breakthrough comes from hardware improvements and algorithmic advancements. As models become smarter and safer, costs will decrease.
Regarding intelligence security, Kevin Weil said that during model training, they approach it more scientifically, and reasoning models will carefully check their answers differently, as they can now use tools and search the web, which will reduce hallucinations.
At the interview's end, Kevin Weil expressed an optimistic attitude towards AGI development, finding its progress speed exciting.
Reference Links:
[1]https://www.youtube.com/watch?v=LZr6Rhu8_as
[2]https://x.com/Box/status/1923121622409527673
[3]https://www.reddit.com/r/singularity/comments/1kqgmhl/openais_kevin_weil_expects_ai_agents_to_quickly/
This article is from the WeChat public account "Quantum Bit", author: Focus on Cutting-Edge Technology, published by 36Kr with authorization.
