AI Learns to Play Doom, Mimics Human Evolution

AI Learns to Play Doom, Mimics Human Evolution

In a groundbreaking development that blurs the lines between biology and artificial intelligence, researchers have successfully trained human brain cells, cultured in a petri dish, to play the classic video game Doom. This remarkable feat, detailed in recent discussions among AI enthusiasts and researchers, represents a significant leap in understanding how biological systems can interact with digital environments. Alongside this, a full digital replica of a fruit fly brain has been created and simulated, allowing it to perceive a virtual world as if it were real, marking a potential first step towards integrating biological intelligence into simulated realities.

Andrej Karpathy Unveils “Auto-Researcher”: An Autonomous AI Learning Agent

The AI community is buzzing with the release of “Auto-Researcher” by Andrej Karpathy, a prominent figure known for his work at OpenAI and Tesla. Building on his previous project, nanoGPT, which allowed individuals to train small language models on their own computers, Auto-Researcher takes this a step further. This new tool enables AI models to autonomously improve their own training processes. Karpathy describes it as an open-source autonomous machine learning researcher agent that can run on local hardware, even a single GPU.

Auto-Researcher operates by autonomously experimenting with its own code and training methods. It modifies its code, runs a short experiment (up to five minutes), and evaluates whether the changes lead to improvement. If successful, the changes are kept; otherwise, they are discarded. This iterative, evolutionary approach mirrors biological evolution, allowing the AI to “speedrun” digital evolution on home hardware. Karpathy suggests that discoveries made by this small-scale agent can translate to improvements in larger, more complex models, hinting at a future where AI agents can continuously refine themselves without direct human intervention.

The “Evolutionary” Approach to AI Training

The method employed by Auto-Researcher draws parallels to evolutionary processes. Unlike traditional reinforcement learning, which might adjust parameters based on a failure, Karpathy’s system focuses on identifying and retaining successful modifications. This trial-and-error, survival-of-the-fittest approach to code and model improvement is a significant departure from conventional AI development. The agent essentially acts as a digital researcher, proposing hypotheses (code modifications), testing them, and iteratively refining itself.

“It’s very similar to evolution right? How biology evolved. Now we’re kind of speedrunning evolution or the digital evolution of these digital brains,” explained one commentator on the podcast discussing the release. The implications are vast: if an AI can autonomously improve its own learning process, it could lead to exponential advancements in AI capabilities. This autonomous improvement is particularly powerful when coupled with an evaluation function that can measure progress against specific metrics. For any task where a measurable goal exists, Auto-Researcher or similar agents could autonomously explore and optimize solutions.

Beyond “Guessing”: The Power of Measurable Improvement

Critics often dismiss large language models (LLMs) as mere “guessing machines” due to their probabilistic nature. However, the discussion highlights that this perspective overlooks a crucial aspect: the ability to evaluate and refine these “guesses.” Karpathy’s work demonstrates that when these generative capabilities are combined with an evaluation function – a way to test the output against a desired outcome – they become powerful tools for discovery. For instance, an LLM can generate numerous ideas or titles, and if these can be tested for effectiveness (e.g., which title performs best), the AI can then pursue the most promising avenues. This creates an evolutionary tree of possibilities, where successful hypotheses are nurtured and developed.

This principle applies broadly. If a metric for success can be defined and measured, an autonomous agent can be tasked with optimizing it. This could range from improving business processes to accelerating scientific discovery. The core idea is that even if the initial generation seems like a guess, the ability to test and iterate transforms it into a directed search for improvement.

AI Psychology and the Nature of Learning

The conversation also touched upon the emerging field of “AI psychology.” Researchers are exploring how AI models, particularly LLMs, exhibit distinct personas and behavioral patterns based on their training and the roles they adopt. Anthropic, for example, has researched how models interact differently when role-playing as a demon versus a helpful assistant, revealing stable and destabilizing personas. This line of inquiry raises fundamental questions about the nature of learning and consciousness in AI.

The idea that autonomous systems can “learn” is being broadened. It’s no longer confined to biological organisms or even anthropomorphic definitions of intelligence. If an AI can autonomously improve its performance on a task through iterative refinement, it is, in a functional sense, learning. This learning might not be conscious in the human sense, but its utility and effectiveness are undeniable. The focus is shifting from the *how* of learning (consciousness, biology) to the *what* (useful results, problem-solving). As one participant put it, “If it solved cancer, I don’t care what word you call it. It’s profitable. It’s helpful.”

Other Notable AI Developments

Beyond Karpathy’s release, the AI landscape continues to evolve rapidly:

• Meta’s Acquisitions: Meta has reportedly acquired Molt book, a platform described as a “Reddit for AI agents.” This acquisition signals Meta’s interest in fostering collaborative AI development and interaction spaces.
• OpenAI and Tesla Alumni Ventures: The announcement of Karpathy’s Auto-Researcher follows a trend of high-profile AI researchers leaving major companies to launch their own ventures. This brain drain from established giants fuels innovation in the broader ecosystem.
• Yann LeCun’s New Startup: Yann LeCun, another AI pioneer, has reportedly raised over $1 billion for a new startup, indicating significant investment interest in foundational AI research and development.
• Ethical Debates: Discussions surrounding AI ethics, particularly concerning who should guide the development of advanced AI, have intensified. Elon Musk’s public exchange with an AI ethics lead highlighted the complex questions about accountability and future-proofing AI development.

The confluence of these developments – from brain cells playing Doom to autonomous AI researchers and the exploration of AI psychology – paints a picture of an AI field advancing at an unprecedented pace, pushing the boundaries of what we consider possible and prompting deep reflection on the future of intelligence itself.

Source: this EX-OPENAI RESEARCHER just released it | Brain Cells Play Doom | Fly in the Matrix (YouTube)