SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Vals AI is seeking an exceptional researcher to design and build the next generation of AI benchmarks for detecting misaligned behavior in frontier models and agents. You will create high-impact evaluations that measure reward hacking, scheming, deception, loss of control, and multi-agent safety risks. Your benchmarks will directly influence how foundation models are developed and generative AI applications are built.
The company works with every major foundation model lab, leading financial institutions, and frontier application-layer companies. Their work has been featured in the Wall Street Journal, New York Times, Washington Post, and Bloomberg. Vals AI recently closed a $40M Series A at a $400M valuation, led by Andreessen Horowitz.
In this role, you will:
- Design and develop novel benchmarks and environments that assess real-world misaligned behavior in frontier models and multi-agent systems
- Conduct research to ensure benchmarks are valid, reliable, and meaningful
- Collaborate with foundation model labs and enterprises to understand evolving safety evaluation needs
- Analyze model behavior across benchmarks and communicate results
- Publish research findings and contribute to the broader safety and evaluation research communities
- Stay current with latest developments in LLM capabilities and alignment evaluation methodologies
The company offers highly competitive salary with meaningful ownership, relocation and transportation support, health/dental insurance, lunch and dinner provided, unlimited PTO, a $1,500 housing stipend (within one-mile radius), and opportunities to publish and present your work.
The founding team includes Stanford PhDs, ex-Jane Street quants, and the first designer at Snorkel, with prior experience at NVIDIA, Meta, Microsoft, Palantir, and HRT. The tech stack is primarily Python, built on Django with React frontend, deployed on AWS using CDK.
REQUIREMENTS:
- Substantial, hands-on work in safety evaluations, monitorability, scheming, deception, or multi-agent safety. Experience on an AI safety/alignment team or PhD focused on AI safety. Published safety research, including papers at leading conferences or high-quality write-ups.
- Deep ML/AI background with understanding of neural networks and model training fundamentals.
- Strong engineering skills with excellent Python coding ability and capacity to design and build complex evaluation environments.
- Ability to clearly communicate complex research ideas to both technical and non-technical audiences.
- Must be in-person in San Francisco (company will support relocation or transportation).
NICE TO HAVES:
- Mechanistic interpretability experience
- Contributions to open-source safety or evaluation tools and datasets
