Google DeepMind is looking to solve cybersecurity misuse on Gemini by designing and implementing defense systems.
Requirements
- Experiences in fine-tuning and adaptation of LLMs (e.g. advanced prompting, supervised fine-tuning, RLHF)
- Deep understanding of machine learning and statistics
- Strong knowledge of systems design and data structures
- Proven experience with TensorFlow, JAX, PyTorch, or similar leading deep learning frameworks
- Recent experience conducting applied research to improve the quality and training/serving efficiency of large transformer-based models
Responsibilities
- Develop and deploy advanced AI/ML solutions to identify and mitigate potential cybersecurity misuse, leveraging prompt engineering, machine learning, and post-training.
- Implement advanced post-training algorithms to optimize Gemini for cybersecurity misuse prevention.
- Diagnose and interpret training outcomes (regressions, gains), and propose and implement solutions to improve model capabilities.
- Actively monitor and evolve the overall defense system's mitigation capability through system metric design and improvement.
- Develop reliable automated evaluation pipeline for cybersecurity misuse mitigation metrics that are strongly correlated with human expert judgment of threat severity.
- Construct adversarial evaluation benchmarks to probe the limits and failure modes of cybersecurity misuse mitigation performance.
Other
- We are seeking individuals who excel in fast-pacing environments and are eager to contribute to the advancement of AI.
- We highly value the ability to invent novel solutions to complex problems, embracing a can-do and fail-fast mindset.
- We are looking for someone who genuinely believes in the future of AI and is committed to devoting their energy in this field.
- Cybersecurity knowledge and background is a great advantage.
- A passion for Artificial Intelligence and Cybersecurity.
- Excellent communication skills and proven interpersonal skills, with a track record of effective collaboration with cross-functional teams
