What is Vibe Science?

Vibe Science is the applied study of how weak human signals and machine inference become validated decisions under uncertainty. It formalizes the discipline between intuition and metrics using the VSV (Vibe → Signal → Validation) loop.

What is the VSV Loop?

The VSV Loop is a three-phase methodology: Vibe (human perception, intuition, discomfort, excitement), Signal (weak data amplified by machines), and Validation (structured experiments, often co-designed with AI). The novelty is that machines enter before certainty, not after.

How is a Vibe Scientist different from a Vibe Coder?

Vibe Coding is tactical prototyping. Vibe Science is systems-level inquiry. A Vibe Scientist designs and validates hypotheses at the boundary between human intuition and machine inference — creating probes, instruments, simulations, and falsifiable narratives.

Who is already doing vibe science work?

Many existing roles already practice aspects of vibe science: Chief Product Officers synthesizing weak signals, CTOs making architectural bets under uncertainty, Heads of Design holding ambiguity, Chiefs of Staff detecting organizational drift, UX Researchers clustering qualitative data, and early-stage founders operating on pattern recognition daily.

Field Notes

Observations, patterns, and reflections from work in ambiguity. Lab notes, not content marketing.

Editor's note: the current lead piece is a practice guide distilled from one week of measuring generative video on a single edge device — five practices ranked by the time they saved, a literature-check prompt skeleton, and a staffing observation about what was actually scarce. The notes below it are the underlying measurements: 30 paired hours of locked-versus-auto camera response, and the running tally of everything that had to be caught first.

2026-08-22

The law held on a second model. The number attached to it did not.

Feeding a video model its own output collapses it. That reproduces on a completely different architecture — but the point where it dies moved by a factor of five, and the number I had been quoting turned out to belong to one model.

video-generationreplicationmethodologydiffusion-modelsai-research

2026-08-19

The correction was the error

My stabiliser removed a third of a pixel of camera shake and charged 28% of the image sharpness to do it. Then the quality gate threw away the result for being blurry.

video-generationimage-processingmethodologymeasurement

2026-08-10

A tripod is the hardest test I can give a video model

Real camera: 28.8 minutes. Frontier model: 20 seconds. The gap is 86x, and the ground truth is free.

video-generationbenchmarksmethodologyevaluation

2026-08-09

The stillest reference made the camera move the most

A week of trying to make a small video model hold still — and the part of the finding that did not survive a second test.

video-generationdiffusion-modelsmethodologyedge-compute

2026-08-03

How not to believe your own instruments

A field guide from one week of measuring generative video on a single edge device. Not a guide to going faster — a guide to not being fooled by measurements you built yourself.

methodologyinstrumentationpre-registrationevaluation

2026-08-02

What a monitoring camera writes down is not what the light did

Two cameras, same room, 30 paired hours — one locked, one on auto. The auto camera removes 83% of the diurnal luminance range and 86% of the colour temperature, and its response is inverted.

video-generationmeasurementinstrumentationearly-signals

2026-08-02

Every instrument I built this week was wrong at least once

Three of seven sealed predictions wrong. Three instruments wrong before they were right, every one caught by a control rather than by looking at the output.

methodologyinstrumentationpre-registrationevaluation

2026-08-01

Generated video has no stillness

A real fixed camera holds perfectly still for 28.8 minutes. No generated footage we measured stays still longer than 4 seconds.

video-generationai-researchmethodologyearly-signalsevaluation

2026-08-01

Latent autoencoders quietly throw away colour detail — and the metric that certifies them can't see it

Measured on a single Jetson AGX Orin across three autoencoders. Destroy 80% of an image's chroma detail and SSIM reports 0.9998.

autoencodersai-researchmethodologyearly-signalsevaluation

2026-07-31

I tried to make an AI generate four hours of a hospital room. It made ninety seconds, then quietly lost the patient.

What actually breaks when you push a video model past its horizon — measured, on one small computer.

video-generationai-researchmethodologydriftearly-signals

2026-07-29

I ran a two-day AI research sprint. Seven of my results died. The eighth was already on arXiv.

What actually happens when you let an AI agent run experiments — and hold it to real controls.

ai-researchmethodologyearly-signalsagentsgrokking

2026-07-29

Write the number, not the adjective

The gap between 'fairly confident' and '70%' is the gap between a team that gets calibrated and a team that just gets older.

calibrationforecasting

2026-07-28

The probe that refused to die

Every unbounded experiment becomes infrastructure with no owner. The teardown date is the cheapest part of the spec and the first thing dropped.

probesexperiment design

2026-07-27

Audit the synthesis, not the summary

A machine will cluster three hundred conversations into six confident themes in a minute. The themes are checkable in about the same time, and almost nobody checks them.

AIevidence

2026-07-27

Decide the reach before you read the result

Any number of responses is encouraging if you never said how many you needed.

experiment designself-deception

2026-07-26

Keep a kill ledger or lose the only data you have

A team's abandoned bets are the highest-information artefact it produces, and almost every team throws them away within a quarter.

kill ledgerdecision hygiene

2026-07-24

A signal without provenance is just an anecdote with a timestamp

The difference between research and lore is whether anyone can say where the observation came from six weeks later.

provenanceweak signals

2026-07-21

Put an expiry date on every conclusion

Beliefs do not announce that they have gone stale. They just keep getting used, long after the world they described stopped existing.

signal half-lifeepistemics

2026-07-19

Naming is infrastructure

Acronyms are usually a smell. Occasionally they are load-bearing — when the alternative is that a practice cannot be taught, delegated, or refused.

namingframeworks

2026-07-18

Killing an idea is a result

The most under-reported outcome in product work is the idea that was killed on purpose, on time, for a written reason.

falsifiabilitydecision hygiene

2026-07-10

Not every problem is agent-shaped

Agents are the current default answer. The vibe scientist question is what problem is genuinely agent-shaped — and what problem just wants a good form.

agents2026

2026-06-24

The quietest users carry the loudest signal

Loud feedback is easy to collect and biased toward the extremes. The middle — the ones who tried it once and drifted — is where the real hypothesis lives.

weak signalsresearch

2026-06-05

The cost of being early is that you look wrong for a while

Being early to a real signal is indistinguishable, from the outside, from being wrong. Vibe science is the discipline of tolerating that ambiguity honestly.

timingepistemics