The Hitchhiker's Guide to the Ocean's Hidden Data  ·  a talk for Unlock with AI
DON'T PANICthere's astonishing stuff hidden in data we already have, and almost nobody's looking. most of it is just sitting there, politely waiting for someone to notice.
≈ Hidden Tide
Roy Bernheim  ·  hiddentide.com
who's dragging you underwater

Hi - I'm Roy.

A builder and entrepreneur - strategy, brands, data + AI systems. Not a scientist, no marine-biology degree, no lab. Just curiosity and AI.

And I've loved the ocean since I was a kid who wouldn't come out of the water.

Hidden Tide is where my two obsessions collide: the ocean, and building with AI.

Young Roy (right) with Oliver and Jean-Michel Cousteau
with Oliver & Jean-Michel Cousteau
Young Roy on the beach at dusk, ocean behind
never able to stay out of the water
Roy presenting data on stage at DMEXCO 2022 and building, today
data and building - the other obsession
Why I'm excited

I think we're back in a da Vinci moment.

For a century, real discovery mostly belonged to big institutions - teams, labs, budgets. I think that just quietly reopened: a time where one curious person, with the right tools, can find something genuinely new again.

the tool that reopened the door: AIand it's not a someday thing →
It's already happening - same shape, twice
NASA WISE infrared all-sky survey, the kind of data Matteo Paz mined
NASA/JPL-Caltech · the infrared sky he mined
An 18-year-old found ~1.5M objects in space
Told to hand-sort NASA's telescope archive. Designed an AI model instead. $250k, Regeneron 2025.
candidates - NASA's own data, never read at that scale
Paul Conyngham with his dog Rosie
Photo: UNSW Sydney · Paul & Rosie
An engineer designed a cancer vaccine for his dog
No biology. Used ChatGPT + AlphaFold (protein-folding AI) to design it. Tumor shrank ~75%.
n = 1 - one dog, not a trial
Outsider + AI + data that already existed → something real
how discovery has always worked

For a century, science started with the question.

the usual way question → hypothesis → collect the data → answer
data-first all the data that already exists → let it raise the question → answer the flip
42
Deep Thought
“Forty-two,” said Deep Thought, with infinite majesty and calm.· · ·“I checked it very thoroughly,” said the computer, “and that quite definitely is the answer. I think the problem, to be quite honest with you, is that you’ve never actually known what the question is.”
The Hitchhiker’s Guide to the GalaxyDouglas Adams
So I pointed it at the ocean

AI, hunting for new things in the ocean - from data we already have.

Whale song on the internet cables
Subsea fibre-optic cables double as microphones, and they've been recording whales for years.
Shrimp as a free thermometer
Snapping shrimp click faster in warm water. Every underwater recording is secretly a thermometer nobody asked for.
Turtles mapping Earth's magnetism
Sea turtles steer by the magnetic field, so their tracks are an accidental map of the planet's buried magnetism.
CLAUDE.md
the project's memory file
the manual · step 0, before any code

Step zero: my guiding principles.

What I care about, and what I don't.

what I care about
free, public data only honest above all open & collaborative automated & self-improving an independent's edge I keep the keys
what I don't want
renting my knowledge back from a black box hype I can't check
CLAUDE.md · read first, always
# guiding principles
- free, public data only
- honest: survives data it never saw
- open & collaborative
- automated, self-improving
- an independent's edge
- I keep the keys
Lives in the project's CLAUDE.md - the first file the AI reads, every time, before it does anything. So step zero steers every step after.
the manual · step one, the boring one
principle free, public data only

Step 1: gather every free dataset that touches the ocean.

Pull it all into one place, so a model can hold the whole ocean at once - not one instrument, one department, one question at a time.

a slice of what's free and public
Argo floatswhale-sound archivesseafloor seismicsatellite ocean-colourship traffic (AIS)tide gaugesbiodiversity (OBIS)gravity satellitesfisheries records
the manual · two ways to go looking
principle an independent's edge

Two ways to look for something new.

Category Zero · no hunch at all
Pour it all in and let the structure raise the question.
Anomalies - where the structure breaks
Couplings - relationships nobody wrote down
Regimes - moods the ocean has, unnamed
42 hunches · give it a direction
A wall of AI-drafted starting points. Claim one, go dig.
The lighthouse keeper's 200-year weather diary
The shrimp doing your oceanography for free
Sea turtles mapping Earth's buried magnetism
the manual · you don't need my permission
principle open & collaborative

Anyone can claim a mission and dive.

hiddentide.com roster: plant your flag on an idea, then claim the mission on GitHub
screenshot goes here · drop it at talk/assets/ht-collab.png
hiddentide.com · plant your flag on an idea → "claim this mission on GitHub" → your public repo joins the crew on the roster
Decentralized: everyone keeps their own repo, logged publicly. No permission needed - findings in the open, nulls included.
Proof · Mission 01 · first, can it even read?

Can AI read 130-year-old handwriting - without making it up?

Yaquina Head lighthouse keeper's handwritten log, January 1892
Yaquina Head keeper's log, Jan 1892 - where a human transcript existed to grade against, the AI read 14 of 14 days.
Proof · Mission 01 · a genuinely new record

It isn't just readable. It's new.

No South Haven temperature exists in any database before 1895. This recovered a real daily record from 1884-85, about a decade earlier - a small but genuine new contribution.
AI reading (dots) vs an independent physics estimate (line), South Haven 1884-85: about 90% match
And it read right, with no answer key: an independent physics estimate from air-pressure only (never a thermometer) matches the AI's reading ≈90%. dots = AI, line = physics.
workflow prompt
one prompt, many subagents
the manual · don't ask a chatbot, build a crew
principle honest above all

One prompt. A whole team of agents.

⚡ one prompt
▾
research build fact-check critique grade
▾
framework→ phases→ gates→ ↻ auto-iterate
▾
a checked, much stronger result
a real kickoff prompt (excerpt)
You are a thorough chief orchestrator. You question everything and fact-check before passing anything on. You work with independent subagents you spawn.

Two rules are the backbone:
(1) INDEPENDENT REVIEW: a fresh agent that didn't produce it scores it; it must clear 8/10, citation-first, or it does not pass.
(2) NON-ASSUMPTION: gather the data first; log claim + URL + fetched date. No assumption is ever presented as fact.

Loop gate: never iterate the same thing more than 3 times; then diagnose, log a limitation, and move on.
Model routing: Opus for synthesis and the final gates; Sonnet for the fan-out and per-iteration reviewers.

Now execute the run-spec as an unattended overnight run. Work phases P0 to P12 in order...
the full spec runs a few hundred lines
The 42 ideas came from one overnight run - 13 phases, dozens of agents grading each other while I slept.
workflow prompt
phases + quality gates
the manual · under the hood
principle automated & self-improving

One prompt, running overnight into a very advanced draft - phase by phase, each fighting to pass.

P0→ P1→ P2→ ···→ P12 one master prompt · ~13 phases, in order · a resumable state machine
inside every phase, the same loop:
fan outN agents, one job each · Sonnet
→
synthesizethe stronger model · Opus
→
the gatea FRESH agent scores it / 10
≥ 8  pass → next phase
< 8  return the gaps → fix → re-review ↻ up to 3× → still short? BLOCKED, logged, move on
author ≠ grader cite-first, or it fails Opus judges · Sonnet does the legwork state in files · resume anytime
Claude Skill
hand-triggered, repeatable
the manual · teach it once
principle honest above all

Teach it once. Never re-explain it.

a task I do a lotreviewing, checking, formatting…
→
a Skilltaught once, reused
→
same quality, every timeno re-explaining
Example - cold-eye review: reads anything before I send it, as the recipient, at three depths: 5-second skimthe careful readbetween the lines
Claude Routine
automated, repeatable
the manual · set it, and it sharpens itself
principle automated & self-improving

The boring, vital work runs itself, and sharpens as it goes.

gets sharper
Daily  - e.g. a learning journal: what worked, what broke
Weekly  - reviews, scans, digests
Monthly  - the bigger rollups
each loop improves on the last ↻
the manual · the rest of the glovebox
principle I keep the keys

The rest of the glovebox.

general AI system
Claude + MCP
Google Workspace, a browser via Playwright, YouTube transcripts.
Setup
Local drive first, then a Swiss server, then access on a need-basis only.
Agent team
On the Swiss server, reachable via Telegram / Slack / OpenClaw.
capture the thinking
Wispr Flow
Talk instead of type, everywhere.
OMI
A wearable that catches ideas while I walk.
Fireflies
Auto-transcribes my meetings.
ship, grow & keep safe
GitHub + Netlify
Push to GitHub; Netlify builds & hosts the live site.
AgentSolo
Generates SEO & GEO traffic to the site.
1Password
The AI acts through it, never sees a password.
The one line I hold: no API keys, no private data, no passwords to the AI - and I do the sensitive plumbing (the domain, the DNS) by hand.
What I keep coming back to

Three things I keep coming back to.

I design a system, not a chatbot
Agents with jobs - not one clever question.
I make it check its own work
Gates, and a slice of data it never saw.
I keep the keys
It builds; I hold the secrets.
If any of it's useful, it's yours. And if the ocean pulls at you too - hiddentide.com is open.
Thank you  ·  questions welcome

So this is an invitation, not an ending. Two ways to take it:

Dive with me
Come explore the ocean deep dives. hiddentide.com is open - and I'd love the company.
Or find your own tide
Skip the ocean. Point all of this at whatever you can't stop thinking about.
DON'T PANICwhatever you love, there's probably data on it - just waiting.
≈ Hidden Tide  hiddentide.com
before you surface

Come find me, and take what's useful.

reach me
Hidden Tidehiddentide.com Roy Bernheimlinkedin.com/in/roy-bernheim

On LinkedIn I share what I learn about building with AI - as I go.

shares that might be useful
≈ Hidden Tide
≈ Hidden Tide 01 / 19 · press S notes   F full   ← → move
Speaker notespress S to hide
← / → or space to move  ·  S = speaker notes  ·  F = fullscreen