Capabilities are growing faster than our ability to control them.¹

Austin AI Alignment runs fellowships in technical AI safety and AI governance for UT students — no prior background required.

JOIN OUR SLACK →

AI AGENT TASK LENGTH, 2023–2026 — LINEAR SCALE

SOURCE: METR TASK-HORIZON DATA (v1.1)

NOTES

¹ METR, Measuring AI Ability to Complete Long Tasks, 2025. Every model released since 2023 is plotted (labeled points trace the state of the art; muted points are contemporaries that didn’t lead). The curve is a single fitted exponential — one growth rate from regression, not forced through the points — continued past the real data to show where the trend leads if it holds. Task-time-horizon is the length of task a model can complete with 50% reliability, shown on a linear axis (hours) rather than the usual log scale.

THIS SEMESTER

- We believe AI can go very well or very poorly, we want to make sure it goes well.

- Weekly GMs are open to everyone, no dues or application required.

- We run two intro fellowships open to all majors, applications open at the start of each semester.

FELLOWSHIPS

Technical AI Safety

INTRO FELLOWSHIP

Introduces the subfields of technical AI safety: alignment, dangerous-capability evaluations, mechanistic interpretability, and AI control. Weekly meetings, then a capstone project of your choosing.

Previous ML experience is appreciated but not required.

WEEKLY · 1 SEMESTER

CAPSTONE PROJECT

AI Governance

INTRO FELLOWSHIP

Introduces AI Policy & Governance. With topics in compute, evaluation & standards regimes, liability, and international coordination etc. This group will have active discussion-led meetings with a final practical policy deliverable.

No previous experience required.

WEEKLY · 1 SEMESTER

PRACTICAL POLICY DELIVERABLE

Talk to an organizer

Fifteen minutes, no commitment, no application required.

Aarushi Lakhi

PRESIDENT

TECHNICAL CO-LEAD

Book a time