arXiv is too big. I cannot read all of it, and I do not know enough papers in advance to write the perfect search query or recommendation profile.
What I can do is read one short edition every day with voice mode running. I comment as I go: this is interesting, this explanation is too long, I do not understand this term, show me more of this, show me less of that.
Those comments shape the next edition. The reader gradually learns which research interests me, what I already understand, which explanations work for me, and what kind of evidence changes my mind.
The daily summary makes the corpus approachable. Voice feedback lets the summary improve with me.
The next edition begins by showing what yesterday’s reading changed.
Talk while the reaction is still useful
I did not want to annotate every paper or fill in a preferences form after reading. I wanted to open the edition on my iPad, tap Record once, read, speak whenever I had a reaction, tap Finish, and walk away.
A single recording can contain confusion about a term, a connection to something I am building, a disagreement with the summary, a request for more of the long tail, and a question to follow up tomorrow. That is much richer than thumbs up or thumbs down, and it captures the reaction at the moment it occurs without breaking the reading flow.
So far the system has recovered and processed four verified arXiv feedback sessions: about 51 minutes of audio across 247 uploaded chunks.
Each paper gives me the finding, a plain-language explanation, a connection to my existing interests, a question, and a caveat. The format is predictable enough to scan, but each part gives me something specific to respond to.
The card connects a finding to ideas the reader has already developed.
This is not “read arXiv without reading arXiv.” The summaries are scaffolding. If the system works, the boundary of what I can understand should move over time.
Tomorrow has to prove it learned
The reader does not get credit for storing a preference. The next edition has to show the effect: a changed selection, a shorter explanation, a stored connection, or a follow-up grounded in something I actually said.
It should also learn without becoming narrow. A daily edition cannot collapse into more of the topic I mention most. It needs enough breadth to surface adjacent work, while giving the best slots to papers that might change or extend an existing belief. When there is no strong paper for one of my interests, it should say so rather than pad the category with a weak match.
The presentation learns too. Early cards explained how to skim and repeated long headings, dates, and product names. My feedback compressed them to short labels and direct findings. An adaptive reader should learn how much explanation its owner needs about the interface itself.
The feedback loop has to survive ordinary use
Voice is only low-friction if the recording is trustworthy. I learned that by starting a recording, navigating elsewhere inside the reader, and later seeing a page claim it was not recording. The visible component had disappeared, but the browser recorder was still alive and uploading invisible ten-second chunks.
The system eventually recovered 172 chunks, 48.6 MB, and roughly 34 minutes of audio. The failure forced recording state to persist across navigation, incomplete sessions to remain recoverable, and uploaded chunks to be preserved and processed repeat-safely.
Freshness needs the same discipline. arXiv’s announcement date, mailing date, listing heading, and my Perth delivery date do not always line up. A daily edition now identifies a source batch by a fingerprint of its paper IDs instead of trusting the date. If the fingerprint has not changed, the feed is not fresh.
The reusable lesson is that personalization depends on state before it depends on model intelligence. The product has to know what was read, what was said, what finished processing, whether the source changed, and whether yesterday’s feedback actually reached today.
The same loop works outside arXiv
I adapted the reader for my dietitian sister as a small portability test. Her version uses PubMed and a separate evidence policy tuned for nutrition research.
Her first edition swept 2,159 PubMed records, screened 554 nutrition-adjacent candidates, and selected 12 across seven sampler lanes. It found no strong ADHD-specific paper that day and said so.
The PubMed edition makes the selection funnel visible, then asks the reader to correct it.
The interaction is reused; the evidence policy and personal state are not.
Her profile, recordings, knowledge state, storage, publishing history, and evidence policy remain separate from mine. The reusable part is the loop: narrow a fresh corpus to one readable edition, let the reader react naturally while reading it, and make the next edition show what those reactions changed.
Codex made it cheap to turn those reactions into rules, tests, and another working version. ChatGPT Sites gave the reader a private, signed-in surface on my iPad early enough for real daily use to shape the product. That speed matters because this kind of reader cannot be specified once. It has to learn through use.