- Creating diverse evaluations that range in difficulty. This involves finding naturally occurring interesting and undesirable behaviors exhibited by open-source models.
- Developing novel architectures and objectives for training interpretability assistants.
- Scaling up the training and inference pipelines to support up to 1T-scale models.
- Experience with fine-tuning language models, designing new architectures, and creating evaluations.
- Reliable results: good experimental design, epistemic self-awareness and transparency
- Generativeness: coming up with original, productive ideas for unblocking progress
- Curiosity: a desire to understand ML systems and how they work
- Strong programming ability, including navigating trade-offs between prototyping speed and maintainability
- Strong communication skills, low ego, openness to giving and receiving feedback
Skills Required
- Experience fine-tuning language models and designing new model architectures
- Experience creating diverse evaluations to find undesirable behaviors in open-source models
- Experience scaling training and inference pipelines to support very large (up to 1T-scale) models
- Reliable experimental design, epistemic self-awareness, and transparency in results
- Strong programming ability and pragmatic trade-offs between prototyping speed and maintainability
- Generative problem-solving and original idea generation to unblock progress
- Curiosity about ML systems and strong communication skills, openness to feedback
- Willingness to work in-person in San Francisco (organization enthusiastic about in-person collaboration)
What We Do
Transluce is an independent research lab that builds open, scalable technology for understanding AI systems and steering them in the public interest. Transluce means to shine light through something to reveal its structure. Today’s complex AI systems are difficult to understand—not even experts can reliably predict their behavior once deployed. Given AI's extraordinary consequences on society, we need scalable and open analyses of the capabilities and risks of AI systems. We are building open source, AI-driven tools to understand and analyze AI systems. We will apply these tools to open-weight models, so the world can vet our analyses and improve their reliability. Once our technology has been vetted, we will work with frontier AI labs and governments to ensure that internal assessments reach the same standards as our publicly vetted procedures. Email: [email protected]
