run #412 · started 02:14 · finished 03:46 · 4 sources

02:14retrieverpulled 1,847 raw signals — tavily, hn, arxiv, rss

02:31validatordropped 1,203 off-topic / clickbait

02:39rankerscored 644, boosted your 10 topics, kept top 14

03:02evaluatortoo thin → back to retriever, scaled up

03:40analyzer14 grounded summaries, 2 with prior-briefing context

03:44evaluatorpruned 2 weak, balanced by source

03:46composerbriefing filed · SQI 61

1,847→12signals in / signals out

You slept. It read everything. Here are the twelve.

free while in preview · topics take a minute · unsubscribe anytime

03:46 — filed

…and this is what it hands you.

Not a link dump. A dated document: each signal is a summary, the concrete details, and why it matters — against its precedent.

top story · tavily

Nvidia buys Hugging Face in a $12.9B bet on open-source AI

Consolidates the model-hosting layer under one vendor; the “open” commons now has a landlord.

research · arxiv cs.CL

Multi-step tool-calling over public APIs: a benchmark and a data recipe

A new eval for agents that chain real API calls, plus a synthesis method for training data.

community pulse · hacker news

“Open-source AI is the path forward” — but nobody agrees what open means

The thread is optimistic on open weights; the friction is definitional, not ideological.

Why it matters —sets up the licensing fight that decides whether the next products are built on a commons or on rented APIs. Same tension as June’s Meta-weights debate.

12 signals · 4 sourcesgrounding 100% · retry ×1

In your inbox at 8:00, or on the web with the full archive behind it.

reads, every night

Hacker NewsarXivTavilyHugging FaceVentureBeatMIT Technology Review

the machine

Six agents in a graph. It can send itself back.

Each briefing is one traversal of a LangGraph state machine. If the evaluator judges the run too thin or too one-sided, it routes back to the retriever and widens the net — once — before it will compose.

graph · run #412

recovery — the only edge that runs backwards.

retrieveTavily, Hacker News, arXiv and lab RSS, deduped into one set.

validateFast yes/no relevance pass. Drops clickbait. Fails open.

rankTrend score, then your topic boosts and hard exclusions.

analyzeGrounded summary, details, “why it matters” — with vector-memory context.

evaluatePrunes weak generations, checks summary vs. headline, balances sources.

composeAssembles the digest, writes the intro, files it to your archive and inbox.

your inputs

You set the beat. It mutes the rest.

Ten topics, picked at onboarding, changed whenever.

The ranker adds a fixed boost to anything matching your topics, so your beats float to the top of the run.

Your exclusion list is stronger: matching signals are dropped in the ranker, before the analyzer ever spends a token on them.

ai agentslarge language modelsfoundation modelsopen source aillm servingrag & vector dbai hardwarealignment & safetyllmopsmultimodalregulation & policycoding assistantsevalsstartups & fundingreasoning & planning

excluded — dropped before analysis

cryptonftsgamingsports

questions

The short answers.

Is it actually free?

Yes, while it’s in preview — no card, no trial timer. If that changes you’ll get plenty of notice.

How is this different from an AI newsletter?

Those are one editor’s picks for everyone. This is a pipeline that ranks the day against yourtopics, cites precedent from your past briefings, and drops anything it can’t ground against a source.

What does it read?

Tavily news, Hacker News, arXiv (cs.AI / cs.CL / cs.LG) and lab RSS — Hugging Face, VentureBeat, MIT Technology Review. One deduped set per run.

Can I change my topics later?

Any time, from Preferences. The next night’s run uses the new set.

get started

Schedule your first run.

next run: tonight, 02:00 · your briefing lands 8:00 AM

already have an account?