DEFNE GENÇ
Stanford CS / AI @ Coinbase / NYC

Learning Et Al.

Solo Project · RecSys · LLM Agents

Learning Et Al. (“learning it all”). A daily research digest around a provocative question, built from the papers and news that match your interests.

Year2026
RoleSolo · Product · Design · Full-Stack
ContextPersonal Project · End-to-End Ownership
ToolsNext.js 16 · Turso/libsql · Drizzle ORM · Tailwind · ONNX embeddings (all-MiniLM-L6-v2) · OpenAlex · Auth.js · Resend · Vercel · Paper
After leaving research, I didn’t want to stray from the literature, but I didn’t want to read entire papers either. I wanted to see what’s out there and find new things to be curious about.
Today’s digest on learningetal.com: the central question, a one-line answer, and the first source card
A single source card: title, byline, TL;DR, and findings beside a takeaway

The Core Idea

Every digest starts from a real research topic, not a keyword. The system walks the OpenAlex taxonomy from one of your interests down to a specific topic, writes a provocative question around it, and gives you the two or three papers that argue best together.

One Digest Per Day

One digest each morning, generated while you sleep. You can regenerate, but you have to say what was wrong with the one you got. The value is curation, not volume.

02How a Digest Gets Made

Discovery

01Topic SeedOpenAlex taxonomyLast 8 excluded
02Question GenTheme-firstGrounded in topic
03RetrievalBM25 + EmbeddingsRRF + MMR diversity
04SelectionComplementarityRe-rank + drop

Synthesis (10-12 LLM calls)

05Editorial PassEvidence-outwardHeadline + order
06SkeletonRoles + tensionsStructured JSON
07ProseArgument arcNot summaries
08Critique + ReviseSelf-refineCoverage gate

What Makes It Work

01Structure before proseA summary of papers from different fields just lists them side by side. So the synthesis first decides which source supports the question, which complicates it, and where the tension is, then writes, fact-checks, and revises. A digest costs about a cent and a half.
02Relevant, not redundantCandidates are ranked with BM25 and embeddings fused by Reciprocal Rank Fusion, then picked by Maximal Marginal Relevance so you never get six versions of the same finding.
03Staying interestingTopics from the last eight digests are excluded and interests decay daily, so themes don’t converge to a template within weeks.
04No riddle headlinesA deterministic check rejects headlines like “Can technology read your mind without touching it?” before they ship, because the headline is also the email subject line.