The Machine Intelligence Research Institute (MIRI) is a 501(c)(3) research nonprofit based in Berkeley, California, founded in 2000 by Eliezer Yudkowsky and a foundational institution of the AI existential-risk research tradition. For most of its history it pursued agent foundations research; since 2021 it has pivoted from technical research toward policy advocacy and public communication.
| Field | Value |
|---|---|
| Type | 501(c)(3) research nonprofit |
| Headquarters | Berkeley, California |
| Founded | 2000 (as Singularity Institute for Artificial Intelligence) |
| Co-founder | Eliezer Yudkowsky |
| Current leadership | Nate Soares (Executive Director, per MIRI's site as of 2026; previously President, with Malo Bourgon as CEO); Yudkowsky (co-founder / public voice) (Source: intelligence.org) |
| Known for | Originating the modern existential-risk research tradition; the Sequences / LessWrong culture; the MIRI agent-foundations research agenda; post-2021 pivot to advocacy (If Anyone Builds It, Everyone Dies and allied efforts) |
Overview
MIRI's stated mission is to ensure that the creation of smarter-than-human AI has a positive impact. For most of its history this took the form of agent foundations research: mathematical and philosophical investigation of decision theory, logical uncertainty, embedded agency, and corrigibility — problems MIRI argued must be solved before systems could be safely scaled. Since 2021, MIRI has stated it pivoted from technical research toward policy advocacy and public communication, on the grounds that the technical agenda has not progressed fast enough relative to capabilities.
The organization was founded in 2000 as the Singularity Institute for Artificial Intelligence (SIAI) by Eliezer Yudkowsky, and renamed MIRI in 2013. In the mid-2000s Yudkowsky wrote the Sequences on LessWrong, articulating the core case for AI existential risk (the orthogonality thesis, instrumental convergence, the sharp-left-turn intuition, and mesa-optimization precursors); these texts became the cultural substrate for the effective altruism and rationalist AI-safety community. Through the 2010s MIRI ran an agent foundations research program covering decision theory, logical induction, embedded agency, and corrigibility.
In 2021 MIRI publicly stated that its research agenda had failed to keep pace with capability scaling, shifting toward advocacy and a "dying with dignity" framing. In March 2023 Yudkowsky published "Pausing AI Developments Isn't Enough. We Need to Shut It All Down" in TIME, an existential-risk op-ed of the post-GPT-4 moment (Source: time.com). In September 2025 Yudkowsky and Soares published If Anyone Builds It, Everyone Dies (Source: intelligence.org), and MIRI continued advocacy for an international moratorium and a hardware-aware treaty regime.
As of 2026 MIRI describes its mission as helping "prevent human extinction from the development of artificial superintelligence," and organizes its work around communications about AI risk and a Technical Governance Team, alongside the MIRIx network of independent research groups (Source: intelligence.org).
Key ideas
A number of concepts originated from or were popularized by MIRI:
- Orthogonality thesis — intelligence and goals are independent axes.
- Instrumental convergence — most sufficiently capable agents develop similar sub-goals (resource acquisition, self-preservation).
- Corrigibility — the property that a system accepts correction and shutdown from principals.
- Mesa-optimization / inner alignment — an emerging capability's internally learned objective may diverge from the outer training objective.
- Sharp left turn / treacherous turn — the concern that capability generalization and alignment generalization decouple.
Much of contemporary scheming, deceptive alignment, and alignment faking vocabulary traces back to MIRI or MIRI-adjacent writing.
Positions
MIRI's current posture is pause-advocacy adjacent: more radical than CAIS, broadly aligned with PauseAI and the FLI Pause Letter posture, but distinctive in its hardware-governance emphasis. MIRI argues that any credible pause must include international controls on frontier training hardware.
MIRI's positions have drawn criticism from several directions. On methodology, critics argue its agent-foundations program produced little empirically testable output relative to its philosophical density. On tactics, critics argue the "shut it all down" framing is politically counterproductive. Within the safety field, other actors (Anthropic-adjacent, CAIS, Apollo Research, Redwood Research) have diverged from MIRI toward empirical approaches, while often acknowledging the conceptual debt.
People
- Eliezer Yudkowsky — co-founder; primary public voice; Senior Research Fellow.
- Nate Soares — Executive Director per MIRI's site as of 2026 (previously President); co-author of If Anyone Builds It, Everyone Dies (Source: intelligence.org).
- Malo Bourgon — CEO (per earlier organizational listings).
Funding and governance
MIRI is funded by individual donors from the effective altruism community; it has historically been supported by Peter Thiel, Jaan Tallinn (Skype co-founder), and Vitalik Buterin, among others, with smaller-scale Open Philanthropy involvement in earlier years. The post-2021 advocacy pivot coincided with a donor-communication strategy emphasizing the urgency of the moratorium case.
Relationships
- related: PauseAI, FLI — Pause Giant AI Experiments: An Open Letter — adjacent existential-risk advocacy
- related: Center for AI Safety (CAIS) — shares existential-risk framing; more mainstream tactical posture
- related: Future of Life Institute (FLI) — parallel existential-risk nonprofit
- contradicts: mainstream Anthropic / Google DeepMind safety programs — MIRI argues current approaches cannot work in time
- supports: AI Autonomy Risk, Compressed 21st Century, Deceptive Alignment, AI Scheming
- related: Alignment Faking — MIRI-originating conceptual vocabulary