Microsoft CFP: "Accelerate Foundation Models Research"
Note: "Foundation model" is another term for large language model (or LLM).
Accelerate Foundation Models Research
...as industry-led advances in AI continue to reach new heights, we believe that a vibrant and diverse research ecosystem remains essential to realizing the promise of AI to benefit people and society while mitigating risks. Accelerate Foundation Models Research (AFMR) is a research grant program through which we will make leading foundation models hosted by Microsoft Azure more accessible to the academic research community via Microsoft Azure AI services.
Potential research topics
Align AI systems with human goals and preferences
(e.g., enable robustness, sustainability, transparency, trustfulness, develop evaluation approaches)
- How should we evaluate foundation models?
- How might we mitigate the risks and potential harms of foundation models such as bias, unfairness, manipulation, and misinformation?
- How might we enable continual learning and adaptation, informed by human feedback?
- How might we ensure that the outputs of foundation models are faithful to real-world evidence, experimental findings, and other explicit knowledge?
Advance beneficial applications of AI
(e.g., increase human ingenuity, creativity and productivity, decrease AI digital divide)
- How might we advance the study of the social and environmental impacts of foundation models?
- How might we foster ethical, responsible, and transparent use of foundation models across domains and applications?
- How might we study and address the social and psychological effects of large language models on human behavior, cognition, and emotion?
- How can we develop AI technologies that are inclusive of everyone on the planet?
- How might foundation models be used to enhance the creative process?
Accelerate scientific discovery in the natural and life sciences
(e.g., advanced knowledge discovery, causal understanding, generation of multi-scale multi-modal scientific data)
- How might foundation models accelerate knowledge discovery, hypothesis generation and analysis workflows in natural and life sciences?
- How might foundation models be used to transform scientific data interpretation and experimental data synthesis?
- Which new scientific datasets are needed to train, fine-tune, and evaluate foundation models in natural and life sciences?
- How might foundation models be used to make scientific data more discoverable, interoperable, and reusable?
Some serious and some playful emendations:
How should we together evaluate foundation models? How might we mitigate the risks and potential harms of foundation models such as bias, unfairness, manipulation, and misinformation as well as those yet-to-be discovered? How might we enable continual learning and adaptation, informed by collective human feedback? How might we ensure that the outputs of foundation models are open to and open about real-world evidence, experimental findings, and other explicit knowledge? How might we advance the study of the social and environmental impacts (and collectively guided impacts) of foundation models? How might we foster ethical, responsible, and transparent use of foundation models across the boundless horizons of imagination and the fetters of our dreams? How might we study and address the social and psychological effects of large language models on human and machine behavior, cognition, and emotion? How can we develop AI technologies that are inclusive of everyone on the planet while also being sure not to "diffuse the radical potential of difference and normalize otherwise oppressive structural conditions" [@hoffmann2020terms]? How might foundation models be used to enhance creative play? How might foundation models accelerate knowledge discovery, hypothesis generation, analysis workflows, and joy in natural and life sciences? How might foundation models be used to transform scientific data interpretation, experimental data synthesis, and research funding? Which new scientific datasets are needed to train, fine-tune, evaluate, and support responsible diffusion of foundation models in natural and life sciences? How might foundation models be used to make scientific data more discoverable, interoperable, reusable, and widely comprehensible?