Building an early warning system for LLM-aided biological threat creation
OpenAI is investing in the development of improved evaluation methods for AI-enabled safety risks. They are particularly concerned about the potential for AI systems to assist malicious actors in creating biological threats. To assess this risk, OpenAI conducted a study with 100 human participants, including biology experts and students. Half of the participants had access to GPT-4, a language model, in addition to the internet, while the other half only had access to the internet. The participants were asked to complete tasks related to the end-to-end process of biological threat creation. The study found mild uplifts in accuracy and completeness for participants with access to GPT-4. However, the effect sizes were not statistically significant, highlighting the need for further research to determine meaningful increases in risk. OpenAI also emphasizes that information access alone is insufficient to create a biological threat, and the study did not test for success in physically constructing threats. The evaluation procedure and results, as well as methodological insights and limitations, are discussed in detail. OpenAI welcomes community feedback on their research.
New embedding models and API updates
The new highly efficient embedding model released marks a significant upgrade over its predecessor, offering stronger performance and reduced pricing. In terms of performance, the average score on the commonly used MIRACL benchmark for multi-language retrieval has increased from 31.4% to 44.0%, and on the MTEB benchmark for English tasks, it has increased from 61.0% to 62.3%. Moreover, the new model is significantly more efficient than the previous generation, leading to a 5X reduction in pricing. The price per 1k tokens has dropped from $0.0001 to $0.00002. While the newer model is recommended, the previous generation model is not being deprecated, allowing customers to continue using it if they prefer. The new next generation larger embedding model, with up to 3072 dimensions, represents the best performing model yet. On the MIRACL benchmark, the average score has increased from 31.4% to 54.9%, and on the MTEB benchmark, it has increased from 61.0% to 64.6%.
Democratic inputs to AI grant program: lessons learned and implementation plans
OpenAI has selected ten teams out of nearly 1,000 applications from 113 countries for its Fellowship for Safety and Engineering program. The chosen teams have members from 12 different countries, offering a diverse range of expertise in fields such as law, journalism, peace-building, machine learning, and social science research. Throughout the program, the teams received support and guidance to facilitate collaboration. They were encouraged to document and describe their processes, enabling faster iteration and identification of opportunities to integrate with other teams' prototypes. OpenAI organized a Demo Day in September for the teams to showcase their concepts to each other, OpenAI staff, and researchers from other AI labs and academia. The projects covered various aspects of participatory engagement, including video deliberation interfaces, crowdsourced audits of AI models, representation guarantees, and mapping beliefs for fine-tuning model behavior. AI technology played a useful role in many projects, providing customized chat interfaces, voice-to-text transcription, and data synthesis. OpenAI has released the code created by these teams and provided brief summaries of their work.