DeepMind AI Solves Real Research Problems
Scientists at DeepMind have developed a new artificial intelligence agent named Aletheia. This AI can perform actual research, solve new problems, and even help write parts of scientific papers. This marks a significant step forward in AI’s ability to contribute to human knowledge.
From Math Puzzles to Scientific Discovery
DeepMind’s research team, led by Quoc Le, previously created an AI that could achieve a gold medal performance on the Mathematical Olympiad. This new AI, Aletheia, goes much further. It tackles open-ended research questions, which are far more complex than contest problems.
Unlike math contest problems that have known solutions and tools, real-life research questions are often uncertain. We don’t always know if they can be solved or what methods might work. This is where Aletheia’s unique approach comes into play.
How Aletheia Works
When given a problem, Aletheia generates a potential solution. A crucial part of its process is a ‘verifier’ that acts like a filter.
This verifier checks the generated solution, discarding junk and asking the AI to try again if it’s not good enough. Solutions that show promise are then polished and reviewed further.
This system is designed to overcome common AI challenges like making things up (hallucinations) and a lack of training data for unknown topics. These issues make it very hard for AI to produce genuinely new and useful research.
Key Innovations in Aletheia
Aletheia uses three main innovations to achieve its capabilities. First, it uses natural English language to check its own proofs, not just rigid mathematical language. This helps prevent the AI from simply agreeing with its own flawed reasoning.
The researchers cleverly separated the AI’s thinking process from the verification step. This way, the messy thought process is hidden from the verifier, stopping the AI from tricking itself. This separation allows for more honest self-assessment.
Second, Aletheia has been optimized to think longer and more efficiently. While the AI is as smart as previous versions, it now uses 100 times less computing power. This is because they trained a stronger base model that is better at reasoning.
This more efficient model can now easily beat the previous gold-medal math AI, improving its score from around 65% to 95%. It achieved this significant improvement in just a few months. This shows a remarkable leap in AI reasoning capabilities.
Third, Aletheia can now search for information and combine techniques from many research papers. This ability to read and synthesize information from dozens of cutting-edge works without getting lost or making errors is what stops it from generating nonsense.
Real-World Impact and Results
Aletheia has already shown impressive results. It autonomously solved four open math problems left by the mathematician Paul Erdős. While these problems were considered somewhat easy and had been ignored for years, solving them is still a notable achievement.
More significantly, Aletheia has helped write parts of actual research papers. One paper focused on calculating constants in arithmetic geometry. The AI also assisted human scientists in writing four other papers, including work on finding new limits for interacting particles.
These research works have been submitted for peer review. Independent mathematicians have checked them for correctness and novelty, confirming that the AI’s contributions are sound and new. This is the first time an AI has created core parts of research that is genuinely new, impactful, and useful.
The Future of AI in Research
The AI’s progress is measured in levels of novelty. Aletheia can now produce publishable-level research and even do so autonomously. While groundbreaking work (levels 3 and 4) is still out of reach, the rapid pace of AI development suggests this may not be the case for long.
This allows a wider audience to experience and utilize this advanced AI capability.
The development of Aletheia shows a clear path towards AI becoming a powerful partner in scientific discovery. This collaboration between human intellect and artificial intelligence promises to accelerate progress and improve lives.
Source: DeepMind’s New AI Just Changed Science Forever (YouTube)
