Thesis

Three claims define the project. Each is stated with the evidence for it and the strongest objection to it.

1. Capability is compounding

The length of task that frontier models complete without help has grown exponentially. METR measures the human-time length of tasks a model finishes at 50% reliability. Across 2019–2025 that length doubled about every seven months, and the doubling time measured after 2023 is shorter, about 131 days with a 95% interval of 107 to 161 days.

Two inputs drive this. Training compute at the frontier has grown roughly four to five times per year. The compute needed to reach a fixed level of language-model performance has halved about every eight months.

Strongest objection. Exponential trends end. Power, chips and data are finite, and part of the measured algorithmic gain depends on scale rather than adding to it. The benchmarks behind the trend are mostly software tasks. See The capability frontier.

2. Safety is an engineering discipline, not a slogan

A system that is more capable than its overseers cannot be made safe by intention alone. It needs methods that can be tested: training that targets human intent, tools that inspect what a model computes, evaluations run before release, and controls that hold even if a model is misaligned.

Each of these exists today in early form and each has known gaps. The Safety section describes them without overstating what they can certify.

Strongest objection. Current methods may not scale to systems much more capable than their evaluators. This is an open research problem, and the project treats it as one.

3. Abundance has to be shared by design

Greater production does not distribute itself. Extreme poverty fell from 2.3 billion people in 1990 to 831 million in 2025, yet the richest 1% hold 37% of global wealth. The data centres, models and robots that would produce abundance are themselves owned assets.

Who owns and governs those assets decides who benefits. The project's position is that the rule for sharing should be set in the open, before the output exists. See Sharing abundance.

Strongest objection. Distribution is a question of law and politics, and a network cannot settle it alone. The project agrees. Its aim is narrower: a transparent, member-governed mechanism that can be examined and improved.

What follows from the thesis