probability: 12%

rationale:

A positive resolution requires a strict conjunction of conditions: a top-five developer must release a top-ten frontier model with a measured capability gain, and credible public documentation must affirmatively establish both that its training compute did not increase and that the developer's research headcount did not increase compared to its predecessor. Both exclusions must be proven; an absence of evidence of growth is insufficient. While the technical realities of AI development make the underlying achievement possible over a six-year horizon, the dual documentation requirement creates a formidable bottleneck.

The underlying technical event is mechanistically plausible by 2032. Pre-training compute efficiency improves by roughly 3x per year epoch.ai, and there is already precedent for capability gains achieved with reduced pre-training compute. For example, evidence suggests OpenAI's GPT-5 outperformed GPT-4.5 despite using less total compute, largely due to a shift toward post-training scaling epoch.ai. Over the next several years, a lab could achieve a compute-flat capability gain via post-training, scaffolding, or better data. Similarly, a headcount plateau could occur either through a capex retrenchment or genuine AI-driven R&D automation.

However, the two constraints—flat compute and flat headcount—naturally pull against each other. A compute-constrained environment (driven by power limits or capex pullbacks) increases the marginal value of human researchers. Conversely, an automation-driven hiring plateau tends to be fueled by massive increases in compute. For instance, Anthropic's own materials note that while engineers are merging significantly more code, human review remains the live bottleneck and overall R&D pace is still determined entirely by the availability of compute anthropic.com.

The most severe obstacle is the requirement for credible public documentation of both exclusions for the same model pair. The industry trend is moving sharply toward secrecy. Stanford's 2026 AI Index notes that the most resource-intensive systems no longer disclose training parameters or duration hai.stanford.edu, and no closed frontier model released after July 2025 carries a public training compute estimate epoch.ai. Furthermore, regulations generally rely on compute thresholds to scope coverage rather than mandating exact FLOP disclosures 2 sources. Research headcount dedicated to specific models is even more opaque; companies simply do not publish these figures in technical reports or system cards ifp.org.

Overall, there is a moderate chance (perhaps 30–40%) that a frontier lab actually achieves this technical milestone by 2032. However, the probability of both variables being credibly documented in the public record is very low. A YES resolution would likely require a lab to deliberately market this specific dual achievement—boasting that they achieved better capabilities with the same team and less compute for an RSI narrative—or a resolver accepting unusually qualitative PR statements. This combination of a challenging technical conjunction and a steep, worsening documentation hurdle limits the probability to the low teens.