Richard Ngo
Richard Ngo
Biographical profile of Richard Ngo covering his career across DeepMind, OpenAI, and back to DeepMind, with focus on his written contributions to the AI safety literature. Documents his influential 'AGI safety from first principles' essay series (2020), his 'The Alignment Problem from a Deep Learning Perspective' paper (2022), and his role as one of the clearest contemporary articulators of the case for taking AI risk seriously.
Quick Assessment
| Dimension | Assessment |
|---|---|
| Current Role | Independent researcher; previously at OpenAI Governance team (2022–2024) and Google DeepMind (2020–2022) |
| Education | MPhil in Philosophy of Physics, Cambridge; MSc in Machine Learning, UCL; undergraduate at Oxford |
| Key Contributions | "AGI safety from first principles" essay series (LessWrong, 2020); lead author on "The Alignment Problem from a Deep Learning Perspective" (ICLR 2024); essays on power-seeking and inner alignment that have shaped community discourse |
| Communication Role | Among the field's most cited essayists; widely regarded as one of the clearest written-articulation of the case for AI risk from a contemporary deep-learning perspective |
| Public Profile | Active on LessWrong, the Alignment Forum, and Twitter/X; engages with both safety-skeptical and safety-focused audiences |
Overview
Richard Ngo is an AI safety and governance researcher whose career has spanned Google DeepMind (2020–2022), OpenAI's Governance team (2022–2024), and independent research. He is widely cited as one of the clearest contemporary writers articulating the case for taking AI risk seriously from a deep-learning-grounded perspective — distinct from the earlier Yudkowsky-era arguments that pre-dated the modern transformer era.
His "AGI safety from first principles" essay sequence — published on LessWrong in 2020 — was an attempt to restate the safety argument in language accessible to mainstream ML researchers, and has been one of the most-linked introductory resources for the field. His subsequent paper "The Alignment Problem from a Deep Learning Perspective" (ICLR 2024, with Lawrence Chan and Sören Mindermann) was an effort to bring those arguments into the academic peer-reviewed literature.
Background
Education
Ngo studied at Oxford as an undergraduate, then completed an MPhil in the Philosophy of Physics at Cambridge before pursuing an MSc in Machine Learning at University College London. The Cambridge-to-UCL trajectory reflects an early interest in foundational questions in physics and philosophy that gradually shifted toward AI as the field he judged most important for the long term.
Early Career
After his MSc, Ngo worked briefly in AI policy research at the Future of Humanity Institute at Oxford before joining Google DeepMind in 2020 as a research engineer / safety researcher.
DeepMind (2020–2022)
At DeepMind, Ngo worked within the AGI safety team, contributing to internal research on alignment and writing publicly on the LessWrong and Alignment Forum communities. His most influential publication during this period was the "AGI safety from first principles" sequence — a five-part LessWrong essay series published in late 2020 that:
- Established why AGI is a coherent research target
- Argued that AGIs may have goals, and goals may not be well-aligned
- Argued for plausibility of recursive self-improvement and rapid capability gains
- Discussed why alignment is hard
- Articulated implications for research priorities
The sequence took as its motivating audience mainstream ML researchers who might be skeptical of AI risk and tried to restate the case in their conceptual vocabulary rather than the inferential-distance-heavy framing of earlier MIRI-era arguments.
OpenAI Governance (2022–2024)
In 2022, Ngo joined OpenAI's Governance team, focusing on AI policy research, including the relationship between AI capabilities and policy interventions. His OpenAI-era publications continued his focus on essay-form articulation of safety arguments. He left OpenAI in 2024.
"The Alignment Problem from a Deep Learning Perspective" (ICLR 2024)
In 2022, Ngo (with Lawrence Chan and Sören Mindermann) published "The Alignment Problem from a Deep Learning Perspective" — first as a preprint, then accepted at ICLR 2024. The paper:
- Articulates four properties of advanced AI systems that produce alignment difficulties: situational awareness, broadly-scoped goals, mesa-optimization (learned subgoals), and reward gaming
- Argues that these properties are likely to emerge from current deep-learning training paradigms scaled up
- Provides what the authors characterize as the first major peer-reviewed articulation of the alignment problem grounded in current ML techniques
The paper has been cited in policy reports, government documents (including UK AI Safety Institute briefings), and academic alignment research as a canonical reference for what the alignment problem is.
Post-OpenAI Independent Research (2024–Present)
After leaving OpenAI in 2024, Ngo has been doing independent research, including continued writing on the EA Forum, Alignment Forum, and his personal blog. Public interest in his views has remained high; his Twitter/X account is widely followed within the AI safety community.
His post-OpenAI writing has continued to articulate evolutions in how he thinks about alignment, including engaging with takeoff-dynamics debates and the question of whether the alignment problem looks different in light of post-2023 LLM behavior.
Style and Influence
Ngo's writing is characterized by:
- A willingness to write at length, often in essay-sequence form, with substantial inferential-distance bridging
- A focus on bringing safety arguments into language compatible with mainstream ML practice
- An emphasis on first-principles reasoning rather than appeals to authority or community consensus
- Willingness to engage with skeptical and critical perspectives
He is widely regarded — alongside Rohin Shah, Jan Leike, and a handful of others — as one of the field's most accessible and credible written-articulation voices on alignment.