Synthetic data for animal-aligned value generalization
- commitment
- 12 hrs/week
- flexibility
- Open to proposals
The project
- Value generalization from animal-aligned training. Does training models to reason compassionately about animals and behave sympathetically towards them also improve treatment towards humans, other AIs, ecosystems, or sentient beings not explicitly mentioned in the training data?
- Trade-offs of animal-aligned training primarily on synthetic data. Can animal-aligned synthetic data degrade human-facing behavior, create unwanted behavior, or shift model personas into undesirable directions?
- How do linguistic diversity and deeper linguistic structures/characteristics (fundamental building blocks of languages, for instance, how the language itself treats the sense of self) affect synthetic data quality in value generalization? And how is this phenomenon represented in English and currently used training data? - I am happy to narrow this direction down, for instance, to the mentioned sense of self if we agree so with the mentee.
Across all three directions, I would aim to conduct research with the mentee that contributes to improving synthetic data quality/generalization in animal-aligned perspectives.
The mentee's role
The mentees would carry out the research end-to-end as primary contributors, with regular co-working, feedback, and review from me. Their tasks will vary based on the final direction.
Who I'm looking for
Must-haves:
- Open-mindedness and genuine care for sentient beings
- Skills necessary for the relevant project - or willing to fast-develop them
- Enthusiasm, willingness to question core assumptions and methodologies
Nice-to-haves:
- Generalist mindset
- Interdisciplinary knowledge and curiosity
Questions for applicants
- Which research direction are you applying for - or are you proposing your own project? What concrete results would make the project successful?
- What excites you about this project, and how do you expect it to contribute to your future impact?
Support offered
- Shaping direction
- Providing connections
- Limited co-working
- Accountability
- Interdisciplinary experience
- Continuous feedback and brain-storming
- Interesting discussions
- format
- Research & writing · Product & entrepreneurship · Field-building & outreach
- topic
- Near-term impact · Post-AGI transition · Tools for advocates · Wild animal welfare · Artificial sentience · AI safety & alignment
Balázs László
AI Safety Generalist, LINAK A/S
I spent most of my time in EA learning about AI safety from technical, governance, advocacy, and field-building perspectives. With a primary background in robotics, technical AI safety research fascinated me; however, I came to realize the importance of a generalist view not just in AI safety but effective altruism as a whole. That is why I became interested in animal welfare and its relationship to AI development and AI safety. My main interest in animal welfare is towards wild animals - primarily through Disease Free Futures - and alignment towards sentient beings.
Some of my background:
- Secretary General at JEF Hungary
- Leader and Founder of Next Generation Hungary (NGH) Informal Group - Erasmus+ (ID:E10372683)
- Founder of SIRAIS - student AI Safety research group at the University of Southern Denmark
- Algoverse AI Research Program - research on steganography detection (mechanistic interpretability)
- Sentient Futures Project Incubator - zoonotic disease detection research, Disease Free Futures
- AI Integrations - LINAK Denmark Operations Excellence Department and R&D