Comment by Veniversum Vivus Vici (veniversum-vivus-vici) on What's your viewpoint on the likelihood of GPT-5 being able to autonomously create, train, and implement an AI superior to GPT-5? · 2023-05-26T07:42:45.490Z · LW · GW

Creating an AI that could autonomously design, train, and implement a superior version of itself is a concept referred to as recursive self-improvement or AI bootstrapping. While this is a fascinating idea and a topic of much discussion in AI research, it is a difficult task with many challenges and risks.

Let's consider the capabilities you mentioned:

Enhanced Short-Term Memory (Extended Token Window): This is an issue of architecture. In principle, GPT-5 could include such improvements, and they could be beneficial. However, a larger token window would significantly increase computational requirements, and it's unclear how much benefit this would actually provide for the specific task of developing a superior AI.

Improved Long-Term Memory (Modifying Its Own Weights and Training Dataset): AI models such as GPT-4 or hypothetical GPT-5 do not have the ability to modify their own weights or training dataset. This ability would require a very different architecture. For the task of designing a superior AI, the model would need to understand the complex relationship between the model's weights and its performance, which is a task that's currently beyond the capabilities of AI. Even if the AI had this capability, training AI models is a resource-intensive task that requires specific hardware resources and infrastructure.

Specific Resources (Access to a DGX data center): Even if an AI had access to such resources, it would still need to understand how to use them effectively, which would require capabilities beyond what GPT-4 or a hypothetical GPT-5 have.

Code Composition: While GPT-4 can indeed generate code, the task of generating code to train a superior AI is far more complex. It involves a deep understanding of AI architectures, algorithms, and principles, as well as the ability to invent new ones. Even for human AI researchers, creating a superior AI model is a significant challenge that requires years of study and expertise.

In addition to these points, there's also the problem of evaluation. Even if an AI could generate a new AI architecture and train it, it would still need to evaluate the new AI's performance and make decisions about how to improve it. This requires an understanding of AI performance metrics and the ability to interpret them, which is another complex task that current AI models are not capable of.

Furthermore, it's worth noting that creating an AI that can improve itself poses significant ethical and safety concerns. Without careful safeguards and oversight, such an AI could potentially lead to unwanted or even dangerous outcomes.

In conclusion, while the idea of an AI improving itself is theoretically possible and an interesting research direction, it's currently beyond the capabilities of current AI technology, including GPT-4 and a hypothetical GPT-5. Achieving this goal would likely require significant advances in AI architectures, algorithms, and understanding of AI principles, as well as careful consideration of ethical and safety issues.