Tech • AI • Robotics • Game
If AI Becomes Recursive, Is It Game Over?!
7/10TL;DR
New AI research and product moves are sharpening a central debate: systems can now help inspect or improve other systems, while companies and institutions are also using software to set new limits on creativity, labor and public decision-making.
KEY POINTS
MIT tests show AI can read and alter another model’s weights
Researchers at MIT explored whether a frontier model identified as GPT-6 Astra could infer what another AI had been trained to do by looking only at its numerical weights. Using very small models with about 17,000 parameters trained on simple games such as Pong, Pac-Man and brick-breaker tasks, the system reportedly recognized each model’s purpose without internet access, tools or the ability to run the target model.
From diagnosis to direct optimization
The experiments went beyond classification. In one test, the larger model examined a smaller specialized system, located likely causes of failure, and suggested direct changes to the weights rather than retraining from scratch. That matters because humans can fine-tune models through additional training, but generally cannot interpret or manually edit billions of parameters in a meaningful way.
A small open model improved after AI-guided edits
In a later trial, the researchers handed over a compact Qwen model with about 0.8 billion parameters that tended to loop during reasoning. After repeated test passes, Astra identified suspect weights, modified them, and the smaller model’s performance improved sharply while the looping behavior dropped. The result suggests AI-assisted model repair may be feasible even outside toy examples.
Anthropic signals increasing AI-on-AI development
The broader industry is already moving in that direction. Anthropic has said in public financial filings that each generation of Claude helps build the next, and that AI now writes a large share of internal code. The combination of AI-generated software and AI-assisted model tuning points toward a development cycle where systems increasingly participate in their own improvement.
Anthropic quietly blocks prompts to clone named games
At the same time, AI firms are drawing their own lines. Claude Code on Opus 5.5 was found to contain internal instructions allowing original games or genre-inspired titles, while refusing requests to recreate a specific existing game. Anthropic later acknowledged adding short prompt rules to steer the model away from rebuilding someone else’s named game.
The policy exposes a gap between copyright law and platform control
The restriction is notable because many legal analysts argue that US copyright law protects concrete expression such as names, graphics and code, but not abstract mechanics or general ideas. In practice, that means a private AI company is making policy choices that may go beyond what copyright clearly forbids, while open-ended cloning of software features remains technically possible.
Andrej Karpathy built a YouTube alternative in hours
Andrej Karpathy, a co-founder of OpenAI and former AI leader at Tesla, recently described building HackerTube, a simplified YouTube-like site generated largely with Claude Opus 5.5. The service draws on roughly 43 million comments from Hacker News to surface highly recommended videos, strips away thumbnails and short-form distraction, and was reportedly assembled in about two hours with minimal manual coding.
Meta is preparing a humanoid robot push
Meta has also moved deeper into robotics with an internal humanoid robot project called Max, reportedly standing about 1.5 meters tall. Leaked material suggests the first version was assembled unusually quickly by relying on hardware from Chinese robotics company Fourier, while future versions aim for more capable hands, stair climbing, conversational hearing at 5 meters, and expressive social behavior geared toward homes rather than factories.
Michigan prisons adopt predictive inmate scoring software
In the public sector, prisons in Michigan are testing software from Recidiviz to classify inmates and guide placement, rehabilitation access and security decisions. The state has seen its prison population fall by 16%, while assaults inside prisons rose 13%. Critics say the system resembles predictive sorting rather than individualized rehabilitation, raising concerns about algorithmic judgment over what a prisoner might do next.
CONCLUSION
Across research labs, tech companies and public institutions, AI is no longer just generating text or images. It is increasingly being used to inspect other systems, enforce private rules, reshape products and influence real-world decisions with limited public oversight.
Ask a question