Programs

Fellowship

Overview

TAS Intro Fellowship

The TAS Intro Fellowship is an eight-week reading group on AI safety, run once a semester. We cover why AI might go badly, what the core risk areas are, and what you can do about it.

Applications

Closed

Join the Discord

This semester's applications closed on October 16th. The next round opens next semester.

8weeks of sessions
90minutes, once a week
0hours of work outside the session

The two tracks

Technical Track

Training and RLHF, adversarial attacks and jailbreaking, mechanistic interpretability, and a survey of control and evaluations as research agendas. No prerequisites Read the overview Hide the overview Full curriculum outline

You meet once a week for a 90-minute discussion. The reading happens in the session, so there is nothing to do outside it. Food is served every week.

No prerequisites and no previous experience. Open to every major.

How it runs

Week 0

  • Week 0
    • Introduction to machine learning
    • Introduction to LLMs

    Optional and self-paced, about 28 minutes. Not a session.

Weeks 1–4

  • Week 1
    • Trends in AI
    • Recursive self-improvement
    • Timelines to AGI
  • Week 2
    • The LLM training pipeline
    • Specification gaming
    • Alignment faking
  • Week 3
    • Orthogonality thesis
    • Instrumental convergence
    • Shutdown resistance
    • Agentic misalignment
  • Week 4
    • Malicious use
    • AI race dynamics
    • Organizational risks
    • Rogue AI
    • Threat models & response plans

Weeks 5–8

  • Week 5
    • RLHF
    • Reward modeling
    • Constitutional AI
  • Week 6
    • Adversarial attacks
    • Jailbreaking
    • Prompt injection
    • Shallow safety alignment
  • Week 7
    • Mechanistic interpretability
    • Features & steering
    • Limits of interpretability
  • Week 8
    • AI control
    • Evaluations
    • Scalable oversight
    • Careers in AI safety

    Surveyed as research agendas you could join, not taught as skills.

Full curriculum outline

Governance Track

How AI policy gets made, compute governance and export controls, frontier safety frameworks, and liability and evaluation standards. No prerequisites Read the overview Hide the overview Full curriculum outline

You meet once a week for a 90-minute discussion. The reading happens in the session, so there is nothing to do outside it. Food is served every week.

No prerequisites and no previous experience. Open to every major.

How it runs

Week 0

  • Week 0
    • Introduction to machine learning
    • Introduction to LLMs

    Optional and self-paced, about 28 minutes. Not a session.

Weeks 1–4

  • Week 1
    • Trends in AI
    • Recursive self-improvement
    • Timelines to AGI
  • Week 2
    • The LLM training pipeline
    • Specification gaming
    • Alignment faking
  • Week 3
    • Orthogonality thesis
    • Instrumental convergence
    • Shutdown resistance
    • Agentic misalignment
  • Week 4
    • Malicious use
    • AI race dynamics
    • Organizational risks
    • Rogue AI
    • Threat models & response plans

Weeks 5–8

  • Week 5
    • The AI policy toolkit
    • Frontier AI regulation
    • Standard-setting & compliance
  • Week 6
    • Compute governance
    • Export controls
    • US–China competition
    • International coordination: guest speaker
  • Week 7
    • Frontier safety frameworks
    • Evaluation standards
    • Corporate governance
    • Liability & transparency
  • Week 8
    • Contested questions in AI governance
    • Open weights
    • Careers in AI policy
Full curriculum outline

Frequently asked questions

Who can apply?

Any Texas A&M undergraduate, in any major. There are no prerequisites and no previous experience is assumed.

What is the time commitment?

90 minutes a week, for 8 weeks. The reading happens inside the session, so there is nothing to prepare and nothing to finish afterwards.

Do I choose a track when I apply?

Yes. The application asks which track you want. Weeks 1 to 4 are the same for everyone and both tracks sit in the same room. The tracks separate for weeks 5 to 8.

How many sessions can I miss?

You need to make at least 7 of the 8. The fellowship is a discussion, so a week you miss is a week the rest of the group misses you.

Is there a capstone or a final project?

Not inside the eight weeks. The only thing you produce there is a one-page statement of what you believe about AI risk and why, written during the final session. After the fellowship we help fellows who want to keep going find a research direction and start a project in it. That is not a formal program yet and it does not have a name.

What is Week 0?

An optional, self-paced introduction to machine learning and LLMs, about 28 minutes. It is not a session and you can skip it. It is there so someone who has never trained a model is not behind in Week 1.