ScrollSci is a vertical, full-screen feed of science — the format short-video
apps use, pointed at real research. Every card starts from a named source:
arXiv, NASA, Hacker News, Wikipedia or Project Gutenberg. You pick from seven
fields and get a 30-second summary plus a plain line on why it matters. Free,
in any browser.
The first image of a black hole took eight telescopes
On 10 April 2019 the Event Horizon Telescope released the first
direct image of a black hole — the one at the centre of galaxy M87, 55 million
light-years away. Eight radio observatories on four continents observed together,
turning Earth itself into a single planet-sized lens.
Why it mattersIt turned a prediction of general relativity
into a photograph, and the same method later imaged the black hole at the centre of
our own galaxy.
Event Horizon Telescope Collaboration · Apr 2019
Aerospace
Parker Solar Probe passed the Sun at 692,000 km/h
On 24 December 2024, NASA's Parker Solar Probe passed 6.1 million
kilometres from the Sun's surface — closer than any spacecraft has been — at roughly
692,000 km/h, the fastest a human-made object has travelled. Its 11.4 cm carbon
shield reached about 1,400 °C while the instruments behind it stayed near room
temperature.
Why it mattersMeasuring the solar wind where it is born,
rather than where it arrives, is the difference between describing space weather and
forecasting it.
NASA · Dec 2024
Biotech
A 50-year-old problem in biology now has 200 million answers
In July 2022, DeepMind and EMBL's European Bioinformatics
Institute published predicted 3D structures for over 200 million proteins — very
nearly every one known to science — in the AlphaFold Protein Structure Database.
Before that, resolving a single structure in the lab could take a doctoral student
years.
Why it mattersStructure constrains function, so this
shortens the first step of almost every drug-discovery and enzyme-design project.
DeepMind & EMBL-EBI · Jul 2022
Mathematics
The four-colour theorem was the first proof no human could check
In 1976 Kenneth Appel and Wolfgang Haken proved that four colours
are enough to fill in any map without two bordering regions matching. Their proof
reduced the problem to roughly 1,900 configurations, then checked each by computer —
about 1,200 hours of machine time.
Why it mattersIt forced the field to decide what a proof is
for — certainty, or understanding — a question formal proof assistants are still
working through.
Appel & Haken, University of Illinois · 1976
Four sample cards. Scroll them. In the app each one also carries a generated illustration.
The short version
What is ScrollSci?
ScrollSci is a science feed built in the shape of a short-video app: one full-screen
card at a time, swipe up for the next. The difference is what fills the cards. Nothing
is posted by users and nothing is invented — each card is a summary of one real item
pulled from a public research feed or archive, with the source named on the card.
Grounded in a real source
Each card begins with a fetched item — an arXiv abstract, a NASA release, a Wikipedia
entry — and the source name and link travel with it.
No engagement algorithm
Your feed is decided by the fields you tick, not by what keeps you watching. There is
no recommender model in the loop at all.
It has an end
Cards come from one shared, finite pool. You can reach the bottom of it — which is the
point. There is no infinite tail engineered to hold you.
Provenance
Where does each scroll come from?
Seven fields, seven public feeds. ScrollSci does not have a newsroom and does not accept
submissions — every card traces back to one of the sources below, and each source is free
and publicly readable, so you can always check the original yourself.
The complete source list. Every ScrollSci card originates from exactly one of these seven feeds.
Field
Source feed
What that source actually is
Biotech
arXiv q-bio
Preprint abstracts in quantitative biology
Aerospace
NASA newsroom
NASA's own press releases, via its public RSS feed
Computer Science
Hacker News
Front-page posts, plus the linked page's own description
Mathematics
arXiv math
Preprint abstracts across mathematics
Physics
arXiv physics
Preprint abstracts across physics
History
Wikipedia “On this day”
Wikipedia's dated events feed
Literature
Project Gutenberg
The daily feed of newly released public-domain books
A note on preprints. arXiv hosts preprints — papers shared before peer
review. That is why they are fast, and also why a ScrollSci card is a starting point for
reading, not a citation. The card links the paper so you can judge it yourself.
Accuracy
How does ScrollSci keep AI summaries honest?
The summaries are written by a language model, and that is worth being direct about. The
model's job is deliberately narrow: compress one fetched source into 30 seconds of reading.
Four things keep it inside that job.
1
It summarises, it does not research
Every card is built from text that was fetched first. The prompts instruct the model
to stay inside that source and not add facts from anywhere else — it is a compression
step, not a knowledge lookup.
2
Thin sources get hedged, not filled in
Most Hacker News front-page posts are bare links with no text of their own. ScrollSci
fetches the linked page's description as real grounding, and where there still isn't
enough to say, the summary says so rather than inventing specifics.
3
The source stays on the card
Every card names where it came from and links out. Nothing is presented as ScrollSci's
own reporting, and the original is always one tap away.
4
Anyone can flag a card
Every card has a flag button that opens a comment box. Flagged cards go to a review
queue and are corrected or removed. Readers are the last line of the check.
What ScrollSci is not. These are machine-written summaries of real
sources. That is not peer review, and it is not a substitute for reading the paper.
Treat a card as a pointer worth following — and if one looks wrong, flag it.
The scroll gesture is not the problem — what it is attached to is. ScrollSci keeps the
format and changes the machinery underneath it: no recommender, no user posts, no
bottomless tail.
ScrollSci compared with a conventional short-video feed, on the five things that decide how a session goes.
An endless short-video feed
ScrollSci
What decides what you see
A model optimising for watch time
The fields you ticked, and nothing else
Where content comes from
Anyone with an account
Seven named research and archive feeds
What a session leaves you
Usually nothing you could restate
A specific claim, its source, and why it matters
When it ends
It doesn't
A finite pool you can reach the end of
How it's generated
Personalised per viewer, endlessly
One shared pool — everyone sees the same cards
One pool, not one per person
Cards are generated once and served to everyone. There is no per-viewer generation,
which means there is no profile of you steering what you see — and no reason to build
one.
Nothing is watching the session
No ad network, no third-party analytics, no tracking pixels. The app stores what it
needs to run your account and nothing beyond it.
Who is ScrollSci for?
People who reach for their phone out of habit and would rather come away with something.
Students keeping a loose eye on fields next to their own; anyone who has tried to read
more papers and lost to the activation energy; and anyone who has closed a video app
after forty minutes and could not name one thing they saw.
It works best as a replacement habit rather than an extra one — same gesture, same idle
five minutes, different residue.
What does it cost?
ScrollSci is free to use. There are no ads and nothing is sold on — the running cost is
low by design, because cards are generated once for everyone rather than per visitor.
You need an account so the app can remember which fields you picked and what you saved.
Email and password, or sign in with Google.
Questions
Questions people ask
Do I need to install an app?
No. ScrollSci runs in any modern mobile or desktop browser. On a phone you can use your
browser's “Add to Home Screen” option to get an app-style icon that opens straight into
the feed.
Which fields can I choose?
Seven: biotech, aerospace, computer science, mathematics, physics, history and
literature. You pick any combination when you sign up and can change it at any time in
settings — the feed filters to your selection immediately.
Are the summaries written by AI?
Yes. Each card's summary and its “why it matters” line are generated by Google's Gemini
models from the text of one fetched source, and the illustration is generated separately.
The model is instructed not to introduce facts beyond that source. The source is named on
every card so you can check it, and a flag button on each card sends corrections to a
review queue.
How often are new cards added?
In batches, drawn from the source feeds as they update. Because the pool is shared rather
than generated per visitor, everyone sees the same new cards at the same time.
Can I save something for later?
Yes. Every card can be liked, saved or flagged, and everything you have liked or saved is
collected in your library.
Is my reading tracked or sold?
No. There is no ad network, no third-party analytics and no tracking pixels on the app.
It stores your account, the fields you picked, and the cards you liked, saved or flagged —
which is what makes those features work. See the
privacy notice for the full list.
Same gesture. Something to show for it.
Pick your fields and start scrolling. It takes about a minute to set up, and it's free.