0 m — a few metres down, still full of sun · 40 sources drifting, none of them checked yet
Research you can
prove and control.
Deep Water reads thousands of sources, checks every fact against the page it came from, and hands back a report you can trust — with the receipts attached.
−300 m — the problem
AI sounds sure of itself.
That doesn't make it right.
Most AI reads a few web pages and stops. Five or ten pages, in English, whatever came up first — then it tells you the answer like it's certain. Everything else stays unread: the papers, the PDFs, and anything not written in English.
−700 m — the sweep
We check every source. Then we throw some away.
We read each page. We check it really says what we say it says. If it doesn't hold up, it doesn't go in the report. Keep scrolling — the light is passing over them now.
−2,400 m — the proof
Every sentence tells you where it came from.
Nothing is in the report without a source. Hover a sentence and its evidence lights up on the right. Never “studies show”.
Browser-native compute matured sharply this year, with workgroup memory limits raised across all three major engines and a common subset finally shipping by default.12
Independent benchmarks recorded a median 3.1× throughput gain on fused kernels against the 2024 baseline, holding across discrete and integrated hardware.07
Adoption stayed gated by driver parity on mobile, where two vendors still trail the specification by a full release and expose only a reduced feature set.23 31
Japanese conformance testing surfaced a regression that no English-language source reported — モバイルGPUドライバの適合性に回帰が確認された — filed against two shipping driver builds.23
German trade press put the driver gap at two release cycles, which the vendors have not disputed publicly, though neither has committed to a date.44
Chinese developer forums report the same regression reproducing on three mid-range chipsets, with a workaround that costs roughly 18% of the gain.52
Patent filings from one vendor describe a scheduler change that would close the gap, but the filing carries no ship signal.61
Where the record is thin, the report says so instead of filling the gap: no source we could verify reports shipped hardware using the new limits in production.08
Some teams report 5× gains. could not be verified · dropped
hover a line. its evidence lights up.
−1,500 m — how deep do you want to go?
104 sources in this run
You choose how deep we go.
You pay for the work you order, not per person. Deeper means more sources, more time, and more cost. Pick a tier — the sediment on the left is every source that run would read.
−3,300 m — for people building AI agents
Your agents sit on top. We are the layer underneath.
An agent that ships work needs to know how it's done today.
Most guess from training data. What works now, what failed last year, what's different in Germany than in Japan — none of that is in there.
We do the reading first. Your agent gets facts with sources attached, before it starts working.
For your work
For developers and researchers
Three ways in.
Pick the one you already use. One line and it's running.
In your own code01
Send us a question. We send the finished report back when it's ready, or you can watch it come in live.
On the command line02
npm install -g @unlikeotherai/deepwater
deepwater research "EU battery rules 2026"
In your AI assistant03
npx -y @unlikeotherai/deepwater-mcp
−4,100 m — the lamp turns on us / for researchers and scientists
Nothing here is hidden from you.
−4,700 m — public or private
You decide who sees it.
Reports are published to our open library by default. That library is free for anyone to read — it's how you can look at hundreds of finished reports right now without paying anything.
If your work has to stay in your team, private research is $50 a month. Then nothing leaves your workspace.
−5,100 m — price
You pay for the work you order.
Not per person. Your team shares one pot of credits, and 1,000 credits is one dollar. A short report costs a little. A deep one costs more, because it reads more.
Before you start, we tell you how long it will take. You are charged for what the run actually used.
Everyone on your team can see what's left.
What one report costs
Indicative — you are charged for what the run actually used, and we tell you how long it will take before you start.
−5,600 m — questions people ask
Is it really checking, or just guessing better?
Every citation gets read back against the page it came from. What doesn't match gets dropped before you see it.
How long does one take?
A quick one takes minutes. The deepest one takes a lot longer, because it's reading hundreds of pages. We show you the estimate before you start.
What if I don't speak the language of the source?
We translate it for you and keep the original next to it, so you can always check.
Can I see it before I pay?
Yes. Our open library is full of finished reports, free to read. Open one and click through the sources yourself.
What can't it do?
It doesn't watch videos or listen to podcasts. It reads text — web pages, papers, and scanned PDFs.