Programs

A lecture series and formalization program to establishAI risk thresholds.

01Lecture Series ↓Researchers present the historical and current research to engineers, standards bodies, and policy institutions.02Formalization Program ↓Tackling formalizations of key uncontainability results, definitions of generality, and risk thresholds that hardware, software, and policy can be built from.

01 · Lecture Series

The curriculum

Researchers present the historical and current research to engineers, standards bodies, and policy institutions. Each lecture is a recorded 30 to 45 minute technical talk that ends with a proposal for what can be done in the near term and with Q&A, published with a companion essay and a profile page. Open a module to see its lectures.

Outcomes we are working toward

  1. 1A stronger basis for distinguishing safe, narrow AI from dangerous, general AI.
  2. 2Measurable risk thresholds along that continuum to guide hardware and software monitoring, verification, and related policy.
  3. 3A technical definition of artificial superintelligence that can be implemented in regulatory mechanisms to prevent its development.
  4. 4More researchers applying formal methods to estimating the danger of AI systems.

Workshops

02 · Formalization Program

Results

The Research Program verifies existing uncontainability results and organizes new work on operationalizing generality risk thresholds.

Result 01 · Machine-checked proof

Machines That Halt

A proof, checked by computer, that no algorithm can verify whether an arbitrary AI model is aligned, and that making models always halt does not change this.

Read the result →Field Survey · Nov 6 2026

Can Advanced AI Be Contained?

A survey of uncontainability research, from formal limits of AI control to deployable risk thresholds.

Read the survey →
In progress

Next results

Further proofs, policy proposals, and recommendations will be published here.

Outcomes

Everything the program builds serves the creation of effective estimates of how likely a particular AI system is to be dangerous, in a form that regulation and technical monitoring can apply.

Four purposes

  1. 01Work with the authors of promising papers on formalizing their results.
  2. 02Make progress on defining the gradient from narrow to general systems and associate risk thresholds.
  3. 03Develop a technical definition of artificial super intelligence that can be implemented in policy to prevent its development.
  4. 04Build the field of risk threshold researchers.

Formalization selection criteria

  1. 1. Gradient construction

    A formal or informal finding is prioritized if it enables a system to be located and compared on a narrow-to-general gradient.

  2. 2. Deployability

    A theorem or question is prioritized if it helps to set a threshold that can be operationalized.

Additional ranking consideration

Automated monitoring

A theorem or research question is prioritized if they define measures that can be detected automatically by hardware and software to support automated monitoring and control.