AI Model Achieves Self-Improvement Milestone
A Chinese AI company called Minimax has announced a major step forward in artificial intelligence with its new model, M2.7. The company claims this model can “evolve itself,” a process they describe as the “early echoes of self-evolution.” This development suggests AI models might soon be able to improve their own capabilities without constant human help.
Minimax, founded in 2022, has already gained a large global user base and backing from tech giants like Alibaba and Tencent. While the idea of AI improving itself has been explored before, Minimax’s M2.7 appears to be pushing the boundaries further.
How AI Self-Improvement Works
Think of an AI model like a pilot and its “harness” as the airplane it flies. The harness includes all the tools, data, and code the AI needs to do its job. Minimax used an early version of M2.7 to build this harness for itself. This means the AI acted like a pilot who is also the head engineer, improving the airplane while flying it.
Initially, the AI acted as a research assistant. It helped with tasks like reviewing research papers, analyzing experiments, preparing data, running tests, and fixing bugs. This alone is impressive, as AI agents are already capable of many complex tasks. However, Minimax states that M2.7 is now handling 30-50% of its reinforcement learning team’s workflow, meaning it’s doing a significant portion of the work needed to train and improve the AI itself.
Autonomous Optimization: The Next Level
The truly groundbreaking part comes next. M2.7 was tasked with improving its own harness through a process called autonomous scaffold optimization. Over 100 rounds, the AI did the following:
- Formulated a hypothesis on how to improve itself.
- Designed experiments to test the hypothesis.
- Modified its own code.
- Ran benchmark tests to see if the changes improved performance.
- Compared results to its previous performance.
- Kept changes that improved performance and reverted changes that didn’t.
All of this happened without any human intervention. It’s like the AI is performing its own scientific method, looping through hypothesis, experiment, and analysis over and over.
Adjusting AI Creativity
One of the specific adjustments M2.7 made involved changing its own “temperature” setting. In AI terms, temperature controls how creative or predictable the model’s output is. A high temperature encourages wilder, more unexpected responses, while a low temperature leads to more focused and standard answers. Minimax found that adjusting this setting improved its performance.
Learning from Mistakes
The model also improved its own guidelines. For example, if it found a bug in its code, it learned to search for similar bugs in other parts of its system. This shows the AI is learning from its errors and applying that knowledge to prevent future mistakes.
Performance Benchmarks: A Shocking Result
Minimax reported a 30% improvement on internal benchmarks after M2.7’s self-improvement process. While impressive, the exact nature of these benchmarks is not public.
More telling are the results from OpenAI’s Machine Learning Engineer (MLE) benchmark. This test measures how well AI models can perform tasks typically done by human machine learning researchers. M2.7, running on a single, relatively affordable A30 GPU, achieved a score of 66.6%. This placed it among top-tier models:
- It tied with Google’s Gemini 3.1.
- It performed close to OpenAI’s GPT 5.4 (71.2%) and Opus 4.6 (75.7%).
These results are significant because M2.7 achieved them on accessible hardware, while the top-performing models likely required much larger, more expensive computing resources. Minimax also reported strong performance on other benchmarks like SWE-Pro and Vibe Pro, showing its capabilities extend beyond research tasks to areas like software development and project delivery.
Why This Matters
The ability of an AI model to autonomously improve itself has profound implications. It suggests a future where AI development could accelerate dramatically, with models becoming more capable and efficient without direct human input for every step. This could lead to:
- Faster Innovation: AI could solve complex problems and create new technologies at an unprecedented pace.
- Increased Accessibility: Models like M2.7 performing well on less expensive hardware democratize advanced AI capabilities.
- New Business Models: The concept of an “AI-native organization” or a “zero-person company,” run entirely by AI, moves closer to reality.
Minimax itself sees M2.7 as a key part of its future, stating it’s accelerating their transformation into an “AI native organization.”
Open Room: An Interactive AI Interface
Alongside M2.7, Minimax launched Open Room, an open-source project available on GitHub. Open Room provides a graphical user interface where users can interact with AI agents. These agents can interact with the user’s files, calendar, and other digital tools, making them proactive and personalized assistants.
Interestingly, most of the code for Open Room was written by AI itself, highlighting the growing trend of AI creating AI. The interface aims to make AI interactions more personable, allowing users to develop a rapport with the AI assistant, similar to how people might prefer one human colleague over another based on personality.
The development of M2.7 and projects like Open Room signal a new era in AI, where self-improvement and personalized interaction are becoming central themes. The industry will be watching closely to see how these advancements shape the future of technology and work.
Source: M2.7 just BROKE the Entire Industry… (YouTube)
