AI safety
This board is still in beta. The reason below is our judgement — tell us if you think we have it wrong. Give feedback
Among the best in the world at what they do20 roles
This tier is about the solution itself. We are fairly confident these organisations are among the best in the world in their field — vetted by GiveWell, Animal Charity Evaluators, Founders Pledge or The Life You Can Save, or featured by 80,000 Hours. Somebody else has already judged that what they do works, so we list every role there that is open to someone in the Netherlands, including support and operations work. Expect fierce competition, and expect remote work or a move.
Tech Lead Manager, GPU Cluster Infrastructure
FAR AIRemote (International)Remote
FAR.AI's researchers test frontier models for OpenAI and the EU AI Office, but that work grinds to a halt as soon as the GPU cluster falters.
AI safetySoftware engineeringManagementSeniorSenior Software Engineer, GPU Cluster Infrastructure
FAR AIRemote (International)Remote
FAR AI tests AI systems before they are released, for OpenAI and the EU AI Office — and that research runs on a GPU cluster that grew this year from dozens to hundreds of cards.
AI safetySoftware engineeringInformation securitySeniorHead of Special Projects
KairosRemoteRemote
In just over two years, Kairos has guided more than 1,200 people into work in AI safety.
AI safetyManagementOperationsSeniorGeneralist (Special Projects)
KairosRemoteRemote
Kairos does not run research itself, but builds the routes towards it: over 1,200 people went through programmes such as SPAR and Pathfinder in two years.
AI safetyOperationsManagementEntry levelEngineering Manager (Red Team)
FAR AIRemote (International)Remote
When FAR.AI finds a weak spot in a new AI model, developers such as OpenAI and Anthropic adjust their safeguards and regulators in the EU follow.
AI safetySoftware engineeringManagementSeniorGCP Program Manager
KairosRemoteRemote
Whoever invites the speakers and selects the participants for a three-day workshop like this decides, in practice, which students and professionals end up in AI safety and biosecurity — and how well prepared they arrive.
AI safetyOperationsManagementMid levelExpression of Interest: Special Projects Associate
Epoch AIRemoteRemote
Epoch AI's figures on compute and scaling are cited by policymakers and researchers worldwide; you do not do that research yourself, but make sure a small team can publish more by coordinating experts, licensing data…
AI safetyOperationsResearchEntry levelGeneralist (Workshops)
KairosRemoteRemote
Kairos runs three-day workshops that steer students towards work in AI safety and biosecurity; over 400 participants so far.
AI safetyOperationsManagementMid levelHead of Groups
KairosRemoteRemote
University groups are the first step towards work on AI safety for many people, and Kairos provides most of the support those groups currently get.
AI safetyManagementOperationsSeniorAssociate Program Officer, Transformative AI Fund
Centre for Effective AltruismRemote - GlobalRemote
This fund finances projects on AI risks at the stage where a few tens of thousands of euros is the difference between existing and not existing: amounts between 10,000 and 150,000 dollars, too small for most other…
AI safetyFinanceResearchEntry levelHead of Kairos Labs
KairosRemoteRemote
Whoever sets up this incubator determines which new AI safety organisations will exist at all in a few years' time: you choose the founders, the questions and the design of the programme.
AI safetyManagementOperationsExecutiveGeneralist (Groups)
KairosRemoteRemote
Many of the people now working on AI safety started out in a student group.
AI safetyManagementOperationsMid levelIT and Security Specialist
Epoch AIRemoteRemote
Epoch AI's figures on compute and scaling are taken up by policymakers and researchers worldwide; that only works as long as the data and the systems around it cannot be manipulated.
AI safetyInformation securityOperationsMid levelTask Development Engineer
METR (Model Evaluation and Threat Research)RemoteRemote
METR's measurements — how long an AI model can work on a task on its own — are taken up in the system cards of AI labs and used by governments for policy.
AI safetyEngineeringResearchMid levelHead of Communications
FAR AIRemote (International)Remote
FAR.AI tests AI models for OpenAI and carries out independent evaluations for, among others, the EU AI Office — but that research only changes something once policymakers and journalists understand it.
AI safetyCommunications and outreachPolicyExecutiveRecruiter / Senior Recruiter
Epoch AIRemoteRemote
Epoch AI publishes the figures on computing power and AI trends that policymakers and researchers worldwide fall back on.
AI safetyOperationsCommunications and outreachMid levelSoftware Engineer, Benchmarking
Epoch AIRemoteRemote
Debates about AI policy often run aground on the question of what models can really do.
AI safetySoftware engineeringDataMid levelTechnical Project Manager, Red Team
FAR AIRemote (International)Remote
FAR.AI tests advanced AI models for weak spots before they are rolled out widely, and does so on behalf of labs such as OpenAI and governments such as the EU AI Office.
AI safetyManagementOperationsSeniorJailbreaking Lead, Red Team
FAR AIRemote (International)Remote
Whoever finds a universal exploit here changes not one system but the behaviour of models used by hundreds of millions of people: FAR.AI's findings demonstrably led to changes at OpenAI and Anthropic, and feed into…
AI safetyResearchSoftware engineeringSeniorResearcher / Senior Researcher
Epoch AIRemoteRemote
Policymakers and investors base their decisions about AI on figures that someone has to gather — on compute, data centres and what models can actually do.
AI safetyResearchDataMid level
What do you make of this?
The board is still in beta, so this is exactly the moment to say what you think. A vacancy we have missed, a judgement you do not share, something that does not work — all of it helps.
Curious?
If you want to know how we arrive at this list, we run a short intro course and a newsletter. Neither costs anything.