Photo by JacLou- DL on Pexels

Artificial intelligence is undergoing a paradigm shift, moving beyond simple programming to self-improvement. While human talent, exemplified by figures like Meta’s Mark Zuckerberg, remains crucial, AI systems are now increasingly capable of ‘bootstrapping’ themselves to achieve higher performance levels. This self-improvement aspect distinguishes AI from other groundbreaking technologies.

Zuckerberg envisions AI liberating humans from tedious tasks, becoming a valuable companion in pursuing individual goals. However, experts like Chris Painter from METR caution that rapid AI self-improvement could accelerate advancements in areas like hacking, weapon development, and manipulative tactics. Some researchers even speculate about a potential ‘intelligence explosion’ where AI swiftly surpasses human intellect.

Jeff Clune, a professor at the University of British Columbia and advisor at Google DeepMind, asserts that automated AI research represents the quickest path to creating powerful AI. He emphasizes AI’s potential in tackling critical issues like cancer and climate change. While human innovation remains essential, AI is playing an ever-increasing role in its own evolution.

Here are five ways AI is actively improving itself:

1. **Boosting Productivity:** Large language models (LLMs) are assisting in code generation, thereby increasing the productivity of software engineers. Google CEO Sundar Pichai reports that 25% of the company’s new code is now AI-generated. Popular tools include Claude Code and Cursor.
2. **Optimizing Infrastructure:** AI is being used to enhance the design of AI chips, thus accelerating the training process. Azalia Mirhoseini from Stanford and Google DeepMind is leveraging AI to refine chip designs. Google’s AlphaEvolve system is improving various elements of its LLM infrastructure.
3. **Automating Training:** LLMs are generating synthetic data and providing feedback for reinforcement learning, which addresses the problem of data scarcity. Mirhoseini at Stanford has piloted a method where an LLM agent formulates potential problem-solving steps, and another LLM acts as a judge to evaluate these steps.
4. **Refining Agent Design:** LLM agents are being designed to optimize their tools and instructions to improve task performance. Clune and Sakana AI have developed a Darwin Gödel Machine, an LLM agent capable of modifying its prompts and tools to enhance its task performance.
5. **Advancing Research:** AI is contributing to scientific literature by determining research questions, conducting experiments, and drafting results. Clune and Sakana AI’s ‘AI Scientist’ has even submitted a paper to the International Conference on Machine Learning (ICML).

The possibility of superintelligence is real. While AI self-improvement could lead to models outstripping human capabilities, the ultimate impact remains uncertain. Google’s AlphaEvolve has accelerated Gemini’s training, but the 1% speedup may not fundamentally alter the trajectory of AI development. METR is closely tracking AI capabilities and has observed that the length of tasks that AI systems can complete independently has doubled every four months since 2024, indicating an acceleration in AI progress. Whether this acceleration will continue remains to be seen.