Forecast Categorical Outcomes for Two Stealth Labs

Categorical mode gives one probability per outcome for a mutually exclusive set. The outcomes are researched together and forecast jointly with a single rationale, so the probabilities are coherent and sum to 100. Each row names an input column holding its own option list.

Make the set exhaustive. Both rows below carry an explicit catch-all, so nothing falls outside the listed options. The rule and the two grouped modes are covered in Forecast categorical and threshold questions.

The questions

Two labs that have shipped nothing public, each asked what it ships first by the end of 2028:

• Safe Superintelligence: what form its first public artifact takes
• Ineffable Intelligence: the domain of its first headline result

Run it

import asyncio

from pandas import DataFrame
from futuresearch.ops import forecast

questions = DataFrame([
    {
        "lab": "Safe Superintelligence",
        "question": "What form will Safe Superintelligence's first public artifact take, by the end of 2028?",
        "outcomes": [
            "Research publication or demo",
            "Hosted model or API",
            "Consumer product",
            "Enterprise or government",
            "Nothing public through 2028",
        ],
    },
    {
        "lab": "Ineffable Intelligence",
        "question": "In which domain will Ineffable Intelligence's first headline result land, by the end of 2028?",
        "outcomes": [
            "Games or simulated environments",
            "Mathematics or theory",
            "Software or computer agents",
            "Robotics or physical world",
            "Other domain",
            "No public result by 2028",
        ],
    },
])


async def main():
    result = await forecast(
        input=questions,
        forecast_type="categorical",
        categories_field="outcomes",
        effort_level="HIGH",
    )
    print(result.data[["lab", "probabilities", "rationale"]])


asyncio.run(main())

probabilities comes back as a JSON object mapping each outcome to its probability. Categorical is HIGH effort only, and each row's option list must hold 2 to 50 unique entries.

Results

Form of Safe Superintelligence's first public artifact, by end of 2028:

Domain of Ineffable Intelligence's first headline result, by end of 2028:

Both catch-all options carry more than a fifth of the mass, which a set without them would have redistributed across the named outcomes.

Going deeper

• Long-form analysis: When Will Safe Superintelligence Release a Model? and How Long Will David Silver's Ineffable Intelligence Stay Dark?
• Published output: SSI's first artifact form and Ineffable's first result domain
• Reference: forecast