The fastest way to lose my attention in a pitch is the phrase “AI-powered.” It's meant to end the conversation — a magic word that makes further questions feel rude. But it's exactly where my questions start. Powered how? Trained on what? And when it's confidently wrong, how would anyone even know?
I'm not anti-AI. I build with it constantly. I'm against a specific bad habit: treating a model as an answer instead of as a component. An output I can't interrogate isn't intelligence. It's a vibe with good PR.
The seduction of the oracle
Here's the trap, and it's a good one. A modern model is a phenomenal pattern machine. Ask it almost anything and it produces something fluent, specific, and plausible. That fluency feels like understanding. It scratches the exact itch a real answer would.
Which is precisely why it's dangerous to trust blindly. A confident wrong answer and a confident right answer are indistinguishable from the inside. The more articulate the model, the more it tempts you to stop checking. An oracle doesn't have to be correct to be persuasive — it just has to be smooth. And smooth is cheap.
What I actually trust
So what would make me believe a prediction? Two things, and neither of them is the model sounding sure of itself.
First, a backtest: replay the claim against history. If the model says “this pattern leads to that outcome,” show me every time that pattern appeared in the past and what actually happened next. Second, a scenario test: replay it against conditions that didn't happen. Perturb the inputs — what if volatility doubled, what if demand halved — and watch whether the prediction degrades gracefully or simply falls apart. A claim that survives both has earned a little trust. A claim that can't be put through either hasn't earned any.
AI is a stage, not the engine
This is why, in everything I design, AI is never the whole system. It's one stage in a loop: ingest → analyse → simulate → predict → act. The model lives at one node of that ring, not at the centre of it. The simulation wrapped around it is what converts a fluent guess into something you can actually stand behind.
That loop is the same engine I keep coming back to — it's literally where the whole platform started, in markets, where an untested prediction costs you money in a very direct and educational way.
Where AI is genuinely brilliant
None of this is a demotion. Used as a component, AI is extraordinary — at exactly the thing the rest of the system is bad at: widening the search. It surfaces candidates a human would never have time to enumerate. It compresses oceans of noise into a few hypotheses worth checking. It proposes. And then a deterministic, testable system disposes — simulating each candidate and keeping only what survives.
That's the division of labour behind ALICE, the research copilot in the ecosystem: the AI suggests, the engine verifies. Generation explores the space; simulation decides what's real.
Propose, then verify
// AI widens the search. Simulation decides what survives.const candidates = await alice.propose(context) // fluent, fast, unverifiedconst survivors = candidates.filter((c) => backtest(c).passes && scenarioTest(c).robust // earn the trust)
Use AI to widen the search. Use simulation to earn the trust.
The honest version of intelligence
What I'm really chasing is intelligence that can show its work. Not a box that emits verdicts, but a system whose reasoning you can replay, stress, and disprove. That isn't a limitation I tolerate; it's the whole point. The same instinct runs under everything I do — the refusal to accept an output without understanding the mechanism underneath it.
Intelligence I can't audit is decoration. Beautiful, sometimes useful decoration — but decoration. The real product was never the model. It's the system I built around it to keep it honest.
