Mathematicians build custom AI models to reclaim control from tech giants
Frustrated by OpenAI's opaque methods, researchers at Caltech and other institutions are developing their own open-source models to solve complex math problems.
The short version
- Caltech professor Sergei Gukov leads a lab building open-source AI models specifically for math problems.
- Critics argue OpenAI's solutions are often 'slop' that require weeks of human effort to decipher.
- New initiatives like Open Math OS aim to give researchers control over their data and model outputs.
- Current custom models are highly specialized, often solving only one specific type of problem at a time.
Why it matters
This shift represents a move toward 'owning the means of production' in academic research. It allows scientists to ensure that AI-generated results are verifiable and that their private data remains secure.
A move toward independence
Sergei Gukov, a professor at Caltech, is leading a laboratory effort to build AI models from the ground up. His team focuses on creating open-source systems specifically engineered to tackle difficult mathematical questions. This initiative comes as a reaction to the way major tech companies have introduced AI into the field.
The academic community has expressed significant frustration with companies like OpenAI. Many mathematicians feel that these corporations are using the field to generate hype for commercial goals, such as an IPO. They argue that the solutions provided by these large models are often published without being vetted by experts in the field.
By building their own models, researchers aim to put the ownership of AI-generated solutions back into their own hands. This approach allows them to understand the inner workings of the systems they use. Gukov believes that the more researchers work directly with these systems, the better they can control the outcomes and the logic behind them.
The problem with commercial AI
Many mathematicians have criticized the quality of solutions produced by mainstream AI. Some describe the outputs as 'slop,' a term implying that the results are meaningless without a human researcher spending weeks trying to decipher the steps. This lack of transparency makes it difficult for the scientific community to trust or verify the work.
There are also significant concerns regarding data privacy and intellectual property. For example, two mathematicians recently accused OpenAI of using their private chat logs to help solve the Navier-Stokes problem. This sparked a wider debate about how much data is being harvested from academic interactions to train commercial products.
In response, some researchers are advocating for platforms that provide open-source frontier models and cloud computing. These tools are designed to let mathematicians control exactly how their data is used. This ensures that private research remains private and is not used to train models for outside commercial interests.
Specialized vs. General models
The models currently being built by Gukov's lab are highly specialized and perform very specific tasks. He compares these systems to Ferraris, describing them as luxury cars that offer high speed and craftsmanship but are not intended for general use. Because they are so focused, a model built for one specific math problem might fail if asked to branch out into a different area.
In contrast, commercial models from companies like Anthropic are more generalized. These systems can handle a wide variety of math prompts across many different disciplines. While they are more versatile, they lack the specific engineering and transparency that the Caltech team is prioritizing in their custom builds.
Gukov admits that a more generalized model is currently in development but is not yet ready for release. For now, the focus remains on creating high-quality, specialized tools. The goal is to provide the math community with reliable instruments that they can fully understand and manipulate.
A divided academic landscape
The rise of AI in mathematics has created a divide among researchers. Some mathematicians are diametrically opposed to using any form of AI in their work, preferring to keep the field human-centric. These individuals worry about the long-term impact of automated reasoning on the integrity of mathematical discovery.
On the other side of the spectrum, many students are embracing AI as a necessary tool for the future of the field. Some believe that integrating AI is the only way to save mathematics, while others are simply exploring the new territory. This tension is visible in lecture halls and online forums across the country.
Despite these differences, there is a growing movement toward creating independent infrastructure. Figures like Terence Tao have advocated for mathematicians to build their own systems. This movement seeks to provide a middle ground where AI can be used as a tool without sacrificing the control and transparency required for serious research.
The future of math research
The development of open-source, mathematician-led AI represents a significant shift in how research might be conducted. It moves the power away from large corporations and back into the hands of the people who understand the underlying principles. This could lead to a new standard where AI is a collaborative partner rather than a black-box service.
As Gukov noted, the research aspect of AI is an exciting opportunity for growth. He views the current era as a period of exploration, comparing the work to explorers discovering new territory. The focus will likely remain on balancing the power of these models with the need for human-verifiable results.
Moving forward, the success of these independent models will depend on their ability to scale. While the current specialized models are powerful, the next step is creating systems that can handle broader complexity while maintaining the same level of transparency. This balance will determine how AI ultimately reshapes the landscape of mathematical discovery.
Sources and links
First reported by Business Insider. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.