How close is this to 2000, really?

What is this?

I googled "when will the AI bubble pop," got a wall of alarmist click-bait, and eventually landed on a Polymarket bet, "AI bubble burst in 2026?". I liked that it pinned the question down: six specific conditions, TRUE or FALSE, and a number for the odds. There's a proper answer! (We'll see if it's the right one.) That gave me a kernel to build around. I'm not a gambling man — I just want to know how much longer I can blissfully sleep before everything goes to shit.

The best analogy we have is the dot-com bust. So the Oracle lines the two cycles up on the same clock and asks, metric by metric: where does today sit on the 1995→2000 timeline?

That's the Then & Now page — the front door. Each metric — price appreciation, valuation multiples, market concentration, capex, and monetary/sentiment signals — becomes a 0–1 intensity, matched against where the dot-com path was at that same level. Roll them up and you get a single projected top, and the crash that historically follows it. It's a bold hook over a transparent method, not a forecast: change the weights, the smoothing, or the anchors and the date moves. That's the point. You can see exactly why it says what it says.

The Polymarket piece

The market's rulebook is still tracked in full on the Polymarket page: NVIDIA and semiconductor drawdowns, a major supplier collapse, an OpenAI/Anthropic bankruptcy or acquisition, H100 rental prices cratering — each a live chart, with the "3 of 6 within 90 days" rule evaluated in front of you. Then & Now grew out of wanting a longer-range read than one market's resolution window.

Where it's going

The first pass wired one metric per pillar; the honest ones disagree with each other, which is the interesting part. The goal is a living read on the indicators that actually help anticipate a turn.

• Deepen each branch — more leaves per pillar (breadth and cap-weight for concentration, a dollar-semis and turnover read for capex) beside the ones already wired.
• Add leading indicators — compute prices, datacenter buildout, funding rounds, model economics — that may turn before the headline conditions do.
• Keep every number inspectable and every assumption a toggle, so you can see how it reaches each answer.

How it's built

Full disclosure: I don't actually know how to do anything. I'm figuring out this vibe-coding thing like the rest of you. I had a long quiet weekend while my wife and kid were out of town and — bam — here we are. Built almost entirely with Claude Code, switching between Opus for the heavy lifting and Fable for fast iteration.

Data comes from free public feeds — Yahoo Finance, Polymarket's CLOB API, CourtListener/RECAP, GDELT, Vast.ai (an H100 rental proxy), Shiller CAPE via multpl.com, FRED macro series, and Jay Ritter's IPO data. Claude Haiku classifies which news actually concerns a company's distress, keeping the buzz markers honest, and writes a plain-language read on each Then & Now metric. The full pipeline is documented on Data Sources.

Open source

The whole thing — collectors, the evaluation engine, the site generators — is public. Fork it, check the methodology, or open an issue. I'm new to open source but I welcome ideas and collaboration. It's a fun topic, assuming this is also your idea of fun.