What would happen if you asked the world’s most powerful AI models the same existential question every single day and forced them to answer with nothing but a number?
That’s the premise behind https://dailydigit.nightlybuilds.co, a project I just shipped. It’s a minimalist forecasting engine that polls five leading LLMs — Claude Sonnet, GPT-4o, Gemini Flash, Llama 3.3, and Mistral Large — daily on questions about humanity’s biggest milestones. Each model gets asked the same question five times. The responses are aggregated into a single number: the Daily Digit.
The questions
Right now, DailyDigit tracks questions like:
- What year will artificial general intelligence be achieved?
- In what year will the first commercial nuclear fusion power plant begin delivering electricity to a public grid?
Big, unanswerable questions. The kind that language models are trained to hedge on with paragraphs of careful qualifications. But DailyDigit doesn’t want qualifications. It wants a year. A single number. By stripping away the hedging, we get something closer to the raw “intuition” baked into these models’ training data and reasoning.
Why this is interesting
Every model carries a worldview shaped by its training data, architecture, and alignment tuning. When you force that worldview through a numerical bottleneck, patterns emerge. Models disagree. Consensus shifts. And over time, as models are updated and retrained, those shifts tell a story.
DailyDigit tracks this story as a longitudinal study. A faded sparkline chart — 365 days of weekly-averaged predictions — sits behind each digit like a ghost, showing the trend without getting in the way of the number. Is the consensus on AGI creeping earlier? Is Mars drifting further out? The chart tells you at a glance.
Each day also produces a short summary of the movement — something like “down 0.4 years from yesterday” — so you can follow the pulse without digging into raw data.
What’s next
I plan to add more prompts over time — questions about climate milestones, space exploration, technological breakthroughs — and potentially track how individual models diverge from the consensus. There’s something compelling about watching machine predictions evolve as models themselves evolve.
For now, the project is live. Go check the current digit, and come back tomorrow to see if the machines changed their minds.
