Summary of Key Developments in the AI Field
Recently, there have been two significant developments in the AI community: First, former OpenAI core member Wen Li has returned to lead a team that is exploring the concept of "AI creating AI" (recursive self-improvement), involving AI in the development of the next generation of models. Second, Ilya, another prominent AI expert, founded a company called SSI, which has received a substantial investment from NVIDIA, focusing on a new approach to ensure model evolution. However, thousands of professionals, including Ilya, have signed a petition requesting control over the pace of this automated development process. While they recognize that this is key to AI breakthroughs, they are concerned that rapid progress could lead to uncontrollable security risks. This highlights the core dilemma faced by the AI industry: balancing technological advancement with risk management.
1. Wen Li Returns to OpenAI: Making AI a "Super Research Assistant" to Accelerate the Creation of Next-Generation AI
Wen Li's goal is not for AI to directly modify its own neural networks (as depicted in science fiction), but rather to train it to assist humans in the research and development process. Previously, AI development involved a slow cycle of human researchers writing code, conducting experiments, and analyzing results. Wen Li's team aims to gradually hand over these tasks to AI: for example, allowing AI to read papers, replicate experiments, write test scripts, organize data sets, and even analyze parameter issues. What used to take a team several days to complete could now be done by multiple AI systems working in parallel, with humans only making the final decisions.
Her approach is more practical: she focuses on optimizing the "working environment" for AI, such as automatically analyzing failure records and adjusting tool usage, creating a cycle where better AI tools lead to higher efficiency, stronger models, and even smarter tools. In other words, she aims to turn AI into a super assistant that can help humans quickly iterate through model development.
2. NVIDIA's Investment in Ilya: Focusing on Model Continuous Learning
Ilya, former chief scientist at OpenAI, founded SSI, which specializes in "secret research." NVIDIA has seen the potential of their work and invested heavily, providing them with the latest computing resources, doubling their computational power. Unlike Wen Li's focus on optimizing processes, Ilya's approach is more fundamental: he believes that the current trend of relying on massive amounts of data and computing power is reaching its limits. He wants models to learn from limited experience and continuously evolve through interaction with their environment (rather than becoming static after training). NVIDIA's investment indicates that they see promising results in this direction.
3. AI Creating AI: Not Science Fiction, but the Automation of the Development Process
When people hear about "AI creating AI," they often imagine AI modifying its own code, but the reality is more mundane. Modern large models are already capable of handling many tedious tasks involved in development, such as managing code repositories, writing test scripts, and analyzing data. Both Wen Li and Ilya's approaches aim to involve AI in its own evolution: Wen Li through engineering processes, and Ilya by allowing the models to learn on their own. The goal is to significantly reduce the time required to develop the next generation of AI. For example, what used to take a year could now be accomplished in just a few months; in the future, models might even design experiments and evaluate results on their own.
4. Thousands of Professionals Sign a Petition: Concerns about Uncontrolled Development Pace
Thousands of experts, including Ilya, have signed a petition expressing concerns that automated development could accelerate too quickly, leaving humans unable to keep up. Their concern is not with the idea of AI creating AI itself, but with the potential for out-of-control progress. Automation speeds up the process: stronger models lead to faster development, which in turn creates even stronger models, creating a self-reinforcing cycle. They raise questions such as:
- Who will set the development goals for AI? What if AI’s optimization choices conflict with human interests?
- Will competing companies skip safety assessments to catch up?
- If model iteration times shrink from years to months, security teams won’t have time to identify and fix vulnerabilities.
5. The Core Dilemma: As AI Becomes a Collaborator, How Much Control Should Humans Retain?
The research by Wen Li and Ilya aims to transform AI from a tool into a partner in the development process. The main question is not whether AI can perform these tasks, but how to control them:
- At what points should humans make decisions? For instance, should human approval be required for AI-generated development proposals?
- Who will define the safety standards? How can we ensure that changes made by AI do not pose risks?
- In a competitive landscape, which company will be willing to pause development for security checks?
In summary, AI creating AI is not just science fiction; it’s a reality in the making. However, as we advance, we must address security and control issues. After all, the goal of using AI to develop more powerful models is to benefit humanity, not to lose control over them.
AI creation is not an unattainable concept; it’s a practical engineering challenge that requires careful consideration of its implications for both technology and society.