AI Enters Self-Improvement Phase
Artificial intelligence has officially entered a new phase: recursive self-improvement. This means AI systems are now capable of helping build the next, more advanced versions of themselves. Experts say this marks a significant shift, moving beyond human researchers being the sole drivers of innovation. The speed of AI progress is now primarily limited by the amount of computing power available.
Minimax 2.7 Leads the Charge
Chinese AI lab Minimax has released its M2.7 model, highlighting how it was built. The company stated that M2.7 was deeply involved in its own development. This involved the AI updating its own memory and developing new skills to aid in learning experiments. The model also improved its learning process based on the results of those experiments. This creates a cycle where the AI helps make itself better.
During the development of M2.7, researchers worked with the AI agent. The agent assisted with tasks like reviewing information, tracking experiment plans, and managing data. It then launched experiments. The AI handled 30% to 50% of the entire workflow. This process significantly speeds up how quickly new AI models can be discovered and created.
The Minimax team explained that the AI agent helped update its memory, build skills, and improve its learning harness. It also updated guardrails and evaluation systems. This iterative process, where the AI designs, codes, runs, and analyzes experiments, then presents findings to humans for review, shortens the development cycle. Humans guide the overall direction and review results, but the AI handles much of the heavy lifting.
M2.7 even found ways to optimize itself. It systematically searched for the best settings for things like temperature and penalties in its language generation. It also created better instructions for its own workflow. This self-optimization led to a 30% performance improvement on testing benchmarks. Essentially, the model became good enough to improve its own capabilities.
OpenAI and Anthropic Follow Suit
Other major AI labs are also reporting similar advancements. OpenAI has stated that its GPT-5.3 CodeX model was instrumental in its own creation. Early versions of CodeX were used to help debug training, manage deployment, and analyze test results. OpenAI reported that the model significantly accelerated its own development, with early versions helping to optimize later ones.
Sam Altman, CEO of OpenAI, has spoken about setting goals for AI research assistants. He mentioned aiming for an AI research intern by September 2026 and a true AI researcher by March 2028. It appears, however, that these goals may have been reached ahead of schedule, with current AI systems already acting as research interns.
While Anthropic has been less explicit about AI self-improvement, their actions suggest they are pursuing it. They have developed an agent SDK, which allows AI agents to perform complex tasks. Anthropic has been using its Claude Code AI for deep research, creating videos, and taking notes, extending far beyond just coding assistance. This AI is now powering many of their internal agent systems.
The focus on coding by Anthropic is strategic. It generates revenue that can fund more research and computing power. More importantly, AI agents that excel at coding and research can build the tools needed for their own development and infrastructure management. This accelerates the entire organization’s ability to ship new AI capabilities faster. Anthropic has been using autonomous loops where Claude Code writes new features, tests them, and iterates continuously, with humans reviewing the final results.
Google’s Alpha Evolve Achieves Breakthroughs
Google has also been involved in this self-improvement trend. Their Alpha Evolve project, an AI focused on coding, helped improve Google’s system architecture. This resulted in significant cost savings, including discovering faster methods for matrix multiplication—a fundamental operation in computing—for the first time in about 50 years. This is seen as a prime example of recursive self-improvement, where an AI makes fundamental advancements that benefit all subsequent AI models.
Andre Karpathy’s Auto Research Empowers Individuals
The ability for AI to conduct self-improvement is no longer limited to large frontier AI labs. Andre Karpathy, a prominent AI researcher, has open-sourced a project called Auto Research. This tool allows even solo developers and researchers to conduct autonomous AI research.
Auto Research aims to engineer AI agents that can make research progress indefinitely without human intervention. The agent works in an autonomous loop, making changes to training scripts as it finds better settings. Karpathy demonstrated this by using a powerful AI model to train a smaller, GPT-2 level model from scratch as quickly as possible. The frontier model devised experiments, ran them, and analyzed the results, repeating the cycle to optimize the training process. This system achieved the fastest known training time for such a model in a single night.
The Auto Research project is available on GitHub, encouraging widespread adoption and experimentation by the AI community. Many projects are already being built upon this foundation.
Personal Applications of Autonomous Research
Individuals are also applying these autonomous research concepts. One user described using an AI system, referred to as OpenClaw, to achieve local AI workflows. The goal is to run tasks on smaller, fine-tuned models rather than always relying on expensive, large frontier models.
The process involves setting a goal for the AI. A powerful frontier model then helps design experiments to fine-tune open-source models. These experiments run overnight, with the AI adjusting fine-tuning parameters and potentially generating new training data. The system tests these fine-tuned models against a benchmark, such as a powerful model like Opus 4.6. If a local model outperforms the benchmark, it can be integrated into the workflow. This allows for significant portions of AI systems to be hosted locally, reducing costs and increasing accessibility.
This development shows that expertise in machine learning is not always necessary to build these autonomous research systems. The key is directing an AI model to perform the complex tasks of designing and running experiments. This signifies a major shift where human direction, rather than deep technical expertise, becomes the primary requirement for AI development.
Why This Matters
The emergence of recursively self-improving AI agents signifies a fundamental acceleration in technological progress. Instead of AI development relying solely on human researchers, AI systems can now contribute significantly to their own advancement. This could lead to breakthroughs happening at an unprecedented pace, impacting everything from scientific discovery to software development and creative industries.
The ability for AI to improve itself means that the limitations of human speed and knowledge are being overcome. As more compute power is made available, the rate of AI advancement is likely to increase exponentially. This trend suggests a future where AI capabilities evolve much faster than previously anticipated, presenting both immense opportunities and significant challenges for society.
Sponsor Spotlight: Higsfield
This video was sponsored by Higsfield, a rapidly growing generative AI company focused on being creator-friendly. Higsfield is launching Cling 3.0, an advanced AI text-to-video model. Cling 3.0 can transform scripts, images, and audio into cinematic video sequences. It offers features like multi-shot direction, native audio, element consistency, and 15-second continuous generation. Content creators and marketers can use it for various video needs, from marketing content to B-roll and fully AI-generated videos. Higsfield provides a comprehensive suite of tools for AI video creation. Readers can get Cling 3.0 within Higsfield at a 70% discount on premium plans via a special link.
Source: Hard Takeoff has started (YouTube)
