AI reads the papers overnight.
You listen to the briefing.
Sample briefing
A real briefing · played in your browser
Two hosts. A few papers. Eight minutes. Nothing scripted by hand — every word was written and voiced overnight from the actual papers that landed in this stream.
Live transcript
What it does
Four pieces, built to compound. The briefings carry continuity. The anchor answers your follow-ups. The graph traces the context behind every claim.
On air
One host introduces each paper by author and institution. The other probes. Written and narrated overnight, before you sit down.
Mech Interp · Today · 5 papers
So this one’s out of Anthropic — Henighan and team. They’re showing sparse autoencoders actually scale to GPT-4 size.
Wait, GPT-4 size? That’s the result everyone was waiting for.
Interrupt
Mid-briefing, ask anything. The anchor pauses, answers from the papers it just read, and resumes.
Youhow does this build on what we covered last week?
AnchorGood question — Aoki’s May paper proposed dictionary learning at this scale was infeasible. This refutes that directly.
Continuity
Each briefing keeps takeaways. Tomorrow’s episode references them when today’s papers extend or refute earlier work.
3 weeks ago
Aoki et al. argued dictionary learning didn't scale past 1B params.
Last week
DeepMind's interim result on sparse coding at 7B sized.
Today
Anthropic refutes Aoki — full SAE at GPT-4 scale, all features stable.
Trace
Open any paper to see what it cites, what cites it, and the labs publishing at the frontier.
Seed
Anthropic SAE @ GPT-4
How it works
Write what you care about like you'd brief a new research assistant. Sciport runs a literature review on your topic, surfaces sub-topics, seed papers, and key researchers — you pick what to follow.
Step 1 of 4 · New stream
Each night it scans new papers across your sources, ranks for relevance, synthesizes the ones that matter, and writes a two-host dialog. Catchy headline. Key takeaways. Audio rendered before you wake up.
2:14 AM · Overnight
Hit play whenever — commute, walk, between meetings. Interrupt your anchor by voice for follow-ups. Tap any paper for its claim, its references, and the citation neighborhood behind it.
▶ Now playing
Mech Interp · today’s briefing
“Pause — who else is publishing in this area right now?”
For working researchers
No fixed taxonomy. Write your stream like a brief — topics, authors, what to skip — and Sciport runs the lit-review, picks the threads, and brings you a briefing every morning.
PhD student
Your streams
Postdoc
Your streams
Principal Investigator
Your streams
Early access
Sciport is in private beta. Drop your email and we’ll open seats to researchers as fast as we can keep up.