AI Futures Model

AI software R&D uplift vs timeConsortium usedDefaultmilestones

Consortium ETSP gains (OOMs since pause)

HardwareWindfallMin SW (mandatory)Used SWForgone SW

monthly

Software progress slowdown from 10× less experiment computei

2.8×

Current rate of software progress (OOMs/yr)i

1.00 OOM/yr

Median→top-human jumps above SAR to reach TED-AIi

1.54

Ratio of top to median researchers' value per experimenti

5.28x

How quickly AIs improve at research tastei

3.5 SDs/2025-effective-FLOP-growth

Input: training compute (H100e)

Input: total R&D compute (H100e)

Safety tax paid (OOMs)

Safety tax paid (dumb version)

Detection survival — P(covert undetected)

Effective compute (OOMs)

AI software R&D uplift

Coding automation fraction

ETSP (OOMs/yr)

Software efficiency (OOMs)

Training system perf (OOMs)

Research effort (realized)

Research effort (covert indigenous)

SW progress rate (covert indigenous, OOMs/yr)

AI coding labor mult (ref present-day)

AI research taste