AI Safety

AI 2040's Total Research Transparency Plan proposes public AI research

Almost all algorithmic secrets and training runs would be publicly visible immediately.

Deep Dive

In a detailed LessWrong post, Thomas Larsen outlines the AI 2040 project's transparency plan for AGI development. The core recommendation is 'Total Research Transparency', where nearly all AI research — algorithms, code, documentation, and experiments — would be immediately visible to the public. Only model weights, significant training data fractions, and a small set of other sensitive information would be kept inside secure datacenters. This approach is designed to satisfy three baseline desiderata: trustless US/China verification of slowdown agreements (each nation can see the other's compute usage), robust model weight security against nation-state adversaries, and ongoing consumer access to frontier AI models (the most informative form of transparency).

Larsen compares Total Research Transparency with three alternatives. Filtered Transparency (Proposal #2) allows only redacted reports, model weights, and researchers to leave a secure boundary. Algorithmic Security (Proposal #3) tightens exfiltration further, while a fourth regime works only for a more aggressive Plan S. The main argument for Total Research Transparency is that it would dramatically improve government and corporate decisionmaking during the critical Plan A phase, and make it harder for either side to abuse power. This is identified as the most concerning failure mode. The post invites discussion on which supplements to crosspost next.

Key Points
  • Total Research Transparency would make all algorithms, code, and training runs publicly visible, with only model weights and some data kept private.
  • The plan enables trustless US/China verification of compute usage without relying on mutual trust.
  • Consumer access to frontier AI models is considered the most important form of transparency, more valuable than evaluations or high-level descriptions.

Why It Matters

Open AI research could transform safety verification and prevent power abuse by governments and corporations.

📬 Get the top 10 AI stories daily