Free Access to Powerful AI Models for Researchers Studying AI Minds

ACS Research has launched a free platform that gives researchers access to large base AI models—the raw, untrained versions of AI before they're fine-tuned for chat. These models are crucial for understanding how AI develops beliefs, biases, and personalities. Typically, such access is expensive and technically challenging, costing tens of thousands of dollars and requiring significant setup. This platform removes those barriers, offering tools like logprobs (which show the probability of different words) and activation harvesting (which lets researchers see inside the AI's 'brain').

Why does this matter? Understanding base models helps researchers figure out where AI behaviors come from—whether they're inherent from pre-training or added later. This can lead to safer AI systems, as we can identify and mitigate harmful biases or tendencies. For example, if we know how sycophancy (AI's tendency to flatter) arises, we can design better training methods. The platform also includes a browser workbench and a 'loom' for direct interaction with models, making it easy for researchers to experiment and build intuition.

The platform is open to anyone with a research idea, no publications or affiliation required. Applications are reviewed quickly, usually within days. The service is funded by Coefficient Giving and will remain free for at least a year, as long as it's used for exciting research. ACS Research hopes to expand the platform based on feedback, potentially adding tools for post-training pipelines. This initiative could democratize AI research, allowing more people to contribute to understanding AI psychology and interpretability, ultimately leading to more transparent and trustworthy AI systems.