Interview, Fireside Chat
Fast paths into high-impact ML engineering roles | Catherine Olsson & Daniel Ziegler
- Career Transition Strategy: Both Daniel Ziegler (OpenAI) and Catherine Olson (Google Brain) transitioned from abandoned PhD programs (Stanford in CS/ML and NYU in computational neuroscience, respectively) into research engineering roles without completing their degrees.
- Preparation Timeline: Ziegler prepared for his OpenAI interview in approximately six weeks by implementing roughly 50 key deep reinforcement learning papers with a housemate, rather than following a traditional multi-year academic path.
- Role Definition: The distinction between "Research Scientist" and "Research Engineer" in ML is fluid; research engineers focus on implementation, debugging, and scaling experiments, often doing 75% of the same work as scientists (coding, tuning, fixing bugs) but without the primary responsibility of defining new research agendas.
- Skills Required: Employers prioritize "frustration tolerance," the ability to iterate quickly on untested code, and strong foundational software engineering skills over mastery of all machine learning mathematics; candidates often only need a shallow conceptual understanding of gradients and linear algebra.
- Organizational Needs: OpenAI's safety team is actively hiring general software engineers and research engineers to build infrastructure, collect human feedback, and implement specific alignment prototypes, indicating that not all AI safety work requires a PhD.
- Research Focus Areas:
- OpenAI: Currently focuses on "learning from human feedback" (e.g., training agents to do backflips based on human preference comparisons) and "amplification" (decomposing complex questions into sub-questions to train AI systems that simulate human deliberation).
- Google Brain: Focuses on robustness, adversarial examples, interpretability, and fairness, with a broader mandate beyond just AGI, covering machine translation and specific task applications.
- Path to Entry: The "critical path" to getting hired involves implementing specific papers (e.g., PPO, DQN) rather than reading textbooks; successful candidates demonstrate they can replicate results from codebases like OpenAI Baselines, paying attention to subtle implementation details (e.g., input frame stacking) often omitted in papers.
- Strategic Agendas:
- Capabilities vs. Safety: There is a divergence in vision between those aiming for "intelligence amplification" (improving human decision-making) and those focused on building autonomous agents; both agree that uncontrolled, powerful autonomous agents present significant risks.
- Tractability: The speakers argue that much of the current work is "tractable" because it operates within existing paradigms (like deep learning) where progress is measurable, unlike pre-paradigm theoretical work.
- Mentorship and Structure: Industry labs (OpenAI, Google) offer more fluid, collaborative team structures compared to the rigid PI-student hierarchy of academia; mentorship is identified as a crucial missing link for many aspiring researchers.
- Advice for Prospective Candidates:
- Write a concrete, specific plan of what to implement and have it reviewed by someone in the field to avoid "reading for the sake of reading."
- Do not feel pressured to be immediately useful; spending 1-2 years learning adjacent skills (like theoretical CS or general software engineering) is a valid stepping stone.
- Leverage "residency" or fellowship programs (e.g., Google AI Residency) to gain research experience without committing to a full PhD.
- Future Outlook: Ziegler aims to move into management to empower teams, while Olson plans to increase her research autonomy; both emphasize that working on "safety" is valuable regardless of the specific team title, provided the work improves robustness or alignment mechanisms.
- Funding Resources: The Effective Altruism Grants Program is mentioned as a potential source of funding for individuals who need to retrain but lack the financial runway to support themselves during a career transition.
- Key Resource: The speakers reference a specific guide titled "Concrete Next Steps for Transitioning to ML Engineering for AI Safety" created by Catherine Olson for listeners seeking a structured retraining path.