3signals3signals turns a noisy stream of AI news, posts, essays, and technical updates into a smaller set of signals you can actually use. The product is designed for people who want to understand what matters without reading every source themselves.
The daily email gives you depth on the most important signals. The Karpathy style second brain wiki gives you breadth, context, and a way to explore how ideas connect over time. Before anything is presented, the system checks that the headline, tone, and supporting evidence line up. Strong claims need strong sources.
Past briefings are preserved in a newsletter archive, and new subscribers receive the latest archived issue after confirming their email so they can start with the current signal set immediately.
Readers can choose daily signals, a weekly digest, both, or wiki-only access. The signup form starts with both email streams selected, but the wiki remains available to verified readers even if they turn off both email subscriptions. Each email has separate daily and weekly unsubscribe links so changing one stream does not accidentally change the other.
What Is A Signal?
A signal is a meaningful pattern supported by source material.
It is not just one link. It is a topic, claim, release, workflow, debate, or market shift that appears important enough to track. A signal may come from one excellent source, but it becomes stronger when multiple sources point toward the same idea.
For example, separate items about a new model, benchmark behavior, infrastructure change, and developer workflow may all strengthen the same broader signal around model releases or agent workflows.
Each signal includes:
- a plain-language title
- topic keywords
- tags that describe the angle
- a date
- a short summary
- supporting articles or posts
- an italicized "Why is this signal important?" sentence explaining why it surfaced
How Signals Are Derived
The system starts with a curated influence list of builders, researchers, companies, investors, educators, and technical publications. It watches the channels that are most useful for each source, then normalizes the material into a consistent shape.
For each item, the system asks a few practical questions:
- What is this about?
- Which concepts does it support?
- Is it original or just repeating familiar commentary?
- Is the author likely to know what they are talking about?
- Does it contain useful evidence, or only hype?
- Does it connect to other items we have already seen?
The answers become structured metadata: topics, tags, entities, summaries, excerpts, source context, and scores. Related items are then grouped together into signals. The grouping is claim-aware: the system fingerprints the actual evidence, entities, and wording behind a candidate signal instead of treating a broad topic label as the whole story. That means two different developments under "agent workflows" or "AI products" can rotate independently when they are supported by different source material.
The extraction step now uses a stricter scoring rubric before anything reaches the ranking layer. New releases, benchmarks, papers, datasets, policy decisions, product launches, implementation details, and first-hand operational reports can earn high novelty scores. Recaps, commentary, predictions, reposts, promotional items, memes, merchandise, and other noisy material are pushed down before they can become a featured signal.
How Scoring Works
Scoring is a ranking system, not a claim of absolute truth. It helps decide what should rise into the daily briefing and what should stay in the background.
The main ingredients are:
- Signal quality: how directly the item contributes useful AI intelligence.
- Authority: how much trust the source has based on the seed list and past role.
- Novelty: whether the item adds something new.
- Attention: lightweight evidence that the item is being noticed.
- Noise reduction: a penalty for sources or items that are likely to be hype,
repetition, or low-context commentary.
Novelty is calibrated against evidence, not excitement. A genuinely new release, benchmark, paper, dataset, policy move, product launch, or first-hand operating lesson can score high. A recap or opinion piece usually lands in the middle. A promotional post, meme, merchandise mention, administrative note, or engagement bait should score low even when it comes from a famous source.
The result is a score for each supporting article. Signal groups are then ranked by their strongest evidence, the amount of support behind them, and how recent that support is. Scoring gets a candidate signal into consideration; it does not automatically make that signal publishable.
The presentation layer also looks for more specific signal concepts inside broad topic families. For example, a general "AI products" topic may split into separate concepts for a browser agent, a clinical workflow model, or an evaluation tool. Recently featured topics and concepts receive a repeat penalty, so the daily briefing is less likely to show the same broad bucket every day when fresher angles are available.
The newsletter focuses on the top few signals at depth. The wiki keeps broader coverage so useful but lower-priority material is still available for exploration.
How Signal QA Works
Before a signal becomes a newsletter item, homepage feature, wiki highlight, or social post, it goes through an editorial QA pass. This checks whether the headline, sentiment, and top supporting article all tell the same story.
3signals treats presentation as part of the intelligence product. The Editor layer turns derived signals into punchy, factual headlines that make the claim directly instead of pointing back at the source. Good headlines read like "Company X launches..." or "Technique Y reduces..." rather than "The article discusses..." or "The author claims...".
Once the Editor finalizes a signal headline, that same headline is used across the newsletter, wiki, homepage feature cards, and X posts. The category label, keywords, and supporting articles may appear as context, but the main claim stays consistent everywhere.
The QA step asks:
- Does the premier article actually support the signal summary?
- Do the supporting articles point at the same concept, or are they only loosely
related?
- Is the headline stronger than the evidence?
- Are we using hype words that the source does not justify?
- Is the source material substantive, or is it promotional/noisy?
If the lead article does not really support the headline, the system looks for a better article in the same signal group. If the wording is too strong, it is softened to match the evidence. If the evidence is weak, promotional, or misleading, the signal is held back.
This is especially important for strong claims. The system should not call something a breakthrough unless the source itself gives concrete evidence, such as a release, benchmark, paper, model, dataset, system card, or measurable capability.
The same QA-approved signal set is used across the email, website, wiki, and X thread so the product does not say one thing in the newsletter and a different thing elsewhere.
Vibe Check
Alongside our tracked authoritative authors, each daily brief may include a separate Vibe Check section. It reflects what the broader community is discussing on Reddit, Hacker News, and GitHub over the last 30 days, summarized in our voice, labeled clearly, and kept strictly separate from our authoritative signals. It is loud, not yet authoritative, and it never replaces or mixes with the three tracked signals.
How The System Self-Checks
The system also checks its own output before publication. It looks at whether the day's featured signals are fresh, whether the set includes newly supported claim fingerprints, and whether the evidence is strong enough to be worth your attention. The current goal is for at least two of the three featured signals to be new or newly supported, while still allowing truly important long-running themes to remain visible.
Why A Signal Surfaced
Each signal includes a short "Why is this signal important?" explanation. This tells you why the system believes the signal deserves attention.
A signal may surface because it has a high-scoring source, multiple supporting articles, strong novelty, high source authority, relevant tags, or fresh evidence that changes the context around an existing topic. New authors and new entity mixes can also lift a candidate when they add a distinct angle rather than simply repeating yesterday's story.
The goal is not to make the ranking mysterious. You should be able to see both the summary and the reason it was elevated.
How Reader Feedback Helps
Each newsletter includes a simple thumbs-up or thumbs-down prompt. Feedback is attached to the specific issue, so it can be compared with what the system chose to feature that day.
Over time, this helps tune what "useful" means in practice. A high-scoring signal that readers consistently ignore may need a better explanation or a lower presentation priority. A signal that earns strong feedback may deserve deeper wiki coverage or more aggressive resurfacing when new evidence appears.
How Signals Decay Over Time
AI moves quickly, so signal ranking fades older evidence over time. Fresh evidence gets more weight in the daily briefing. Older evidence remains useful as context, but it becomes less likely to dominate the top of the feed unless it is reinforced by new support.
Decay applies at more than one level. Old articles lose ranking power, recently featured source links are suppressed, and recently featured signal concepts are discounted. That lets a strong continuing theme stay visible when there is real new evidence, while making room for genuinely new developments inside the same larger topic area.
This helps the system avoid getting stuck on yesterday's story while still preserving older material in the wiki.
Decay is meant to shift attention, not erase history. Fresh evidence should win the daily briefing, while older evidence remains available in the wiki as context.
How The Wiki Works
The Karpathy style second brain wiki is the memory layer behind the briefing. It cross-links concepts, signals, sources, and supporting evidence so you can move from a short daily summary into the larger map.
The wiki lets you:
- open a concept and see the supporting signals
- click a keyword from the newsletter into the relevant concept
- filter evidence by notable voices
- filter the map by recent pull date or by the source's own publication date
- open the original source material
- follow related concepts through the map
- download a portable Markdown version for personal notes, including scoped
vaults such as the last 5, 7, 30, or 90 days
Think of the newsletter as "what should I read today?" and the wiki as "how does this fit into the bigger picture?"
The default wiki date filter uses pull date because it reflects what 3signals learned during recent collection runs. Source-date filtering is also available when you want to study material by when the original author published it. Both views preserve the same cross-links between concepts, signals, authors, sources, and supporting evidence.
Why The Seed List Matters
The seed list is the editorial starting point. It defines the people, organizations, topics, and source types that the system should pay attention to.
This matters because AI news is not evenly distributed. Some sources create original work. Some explain difficult ideas well. Some mostly repeat other people's work. The seed list helps the system start with better judgment before the ranking algorithm ever sees an item.
The list can evolve. New voices can be added, stale sources can be reduced, and topic coverage can expand as the AI landscape changes.
What To Expect Each Day
Each briefing is designed to answer three questions:
- What are the strongest AI signals right now?
- Why did those signals rise above the noise?
- Where can I go deeper if I care about one of them?
The result should feel less like a pile of links and more like a living research map that gets updated every day.
If you miss an issue, the archive keeps previous briefings available so the daily email and the long-running wiki stay connected.