Skip to the content.

2026 (rolling) — How to find hot topics systematically

Why this matters

The other entries in this repo each cover what — a topic, a launch, a pattern. This one covers how — the systematic method for finding which topics are worth writing entries about in the first place. It’s the meta-discipline that makes the rest of the repo’s currency possible.

Without a system, you end up with two failure modes: (1) chasing every shiny launch and producing low-signal coverage, or (2) missing major paradigm shifts because you’re not looking in the right places. The 2026 ecosystem moves fast enough that both failure modes are common.

following.md is the output of this methodology — the source list. This entry is the process.

The five-tier intake stack

A reliable 2026 information diet for agentic engineering has five tiers, ordered by signal-to-noise (highest first) and update cadence:

Tier 1 — first-party release channels (highest signal)

What ships from the vendors themselves. Cadence: per-release, typically daily for active products.

Discipline: subscribe via RSS where available; don’t rely on algorithmic surfacing.

Tier 2 — curated newsletters + podcasts (high signal, daily-weekly)

Filter-and-digest layer. Senior practitioners doing the synthesis you’d otherwise do yourself.

Discipline: don’t subscribe to more than 6-8. The headline 2026 stat from the Readless aggregator review: AI publications hit 242,000 per year and growing. Quality filtering matters more than source quantity. Start with 8 trusted feeds; revisit the list monthly to drop low-signal sources.

Tier 3 — named individuals (highest signal-per-post)

The people-not-platforms layer. Specific named operators whose output is worth following directly rather than via aggregation. The cited 2026 set for agentic engineering:

Person Signal type
Boris Cherny (Anthropic, Head of Claude Code) Direction-of-travel; podcast circuit; @bcherny
Cat Wu (Anthropic, founding eng on Claude Code) Engineering-side framing; UX choices
Andrej Karpathy Broader LLM commentary; the 80% pivot post; the LLM Wiki spec
Simon Willison Live-event coverage; deep dives on tool behaviour
Swyx AI Engineer infrastructure; Scaling without Slop framing essays
Cole Medin Long-form YouTube agent deep-dives; PIV methodology
Paul Gauthier Aider perspective on precision tools
Tom Blomfield YC partner; economics framing
Hamel Husain Practitioner-engineer reflection on production ML

Discipline: follow on Twitter/X via lists (not the algorithm feed); subscribe to their personal substacks. Don’t follow more than ~15 names — beyond that the signal-to-noise drops sharply.

Tier 4 — conference circuit (timed-event signal)

When the field gathers, the conversation crystallises. The 2026 calendar:

Discipline: read the post-event debrief posts (Simon Willison live blogs, Latent Space debrief episodes) rather than trying to consume every talk individually. Conferences produce both signal and industry-PR noise; the digest layer separates them.

Tier 5 — aggregators (broad, lowest signal-to-noise, useful for surface coverage)

The 2026-current aggregator picks:

Discipline: don’t read aggregators daily. Schedule a fixed weekly skim (e.g., Sunday morning, 30 minutes). The rest of the week, work — don’t let aggregator FOMO drive the schedule.

The hot-topic detection rules

A topic is hot enough to write an entry when at least three of these signals fire within ~30 days:

  1. Multiple independent writeups — not the same essay reposted across three sites; different practitioners covering the topic from different angles.
  2. A senior named individual (Tier 3) talks about it — Cherny / Karpathy / Swyx tweet, post, or appear on a podcast about it.
  3. An arxiv paper exists or is in submission. The 2026 ecosystem moves fast enough that papers lag practice; but when both happen, the topic is durable.
  4. A conference talk is dedicated to it. Conferences are tier-4 timed events; their topic selection is curation.
  5. A standards body discusses it — Linux Foundation Agentic AI Foundation working groups, IETF, W3C. Standardisation is a late-stage signal but a durable one.
  6. Multiple open-source implementations spring up — the Karpathy LLM Wiki spec spawned 6+ implementations within months; that’s the signal, not Karpathy publishing the spec alone. Implementations are votes.
  7. Anthropic / OpenAI / Google announces an integration covering the topic. Vendor-side adoption is structural, not aspirational.

Discipline: when fewer than three of these fire, the topic might be interesting personally but isn’t yet load-bearing across the field. Wait. Or write a brief mention in following.md without committing to a full entry.

The 3-source verification (anti-hallucination discipline)

Before writing an entry, verify the topic against three independent sources:

  1. A primary source — vendor docs, official announcement, the actual code repo
  2. A community source — a writeup by an independent practitioner with citations
  3. A secondary source — a third writeup that doesn’t just recycle the first two

This is the discipline that prevents zeuikli-style errors (see the hooks entry’s call-out of plausible-but-fake event names). Specifically:

When all three sources agree on the load-bearing facts, the topic is trustworthy enough to write an entry about.

The cadence model

A reliable rhythm:

Cadence What
Daily Check first-party release channels (5 min)
Weekly Skim curated newsletters; review one or two named-individual posts (30 min)
Monthly Audit follow-list, drop low-signal sources, add emerging ones (60 min)
Per-conference Read the post-event debrief; decide if any topics warrant new entries (variable)
Quarterly Re-read the existing repo entries for supersede-notes opportunities (45 min)

The quarterly re-read is the discipline that keeps the rest of the repo current. Every quarter, walk through the entries and ask “what claim in this entry is now wrong?” Add supersede notes; don’t silently edit.

The “★ earns its place through use” rule

Per following.md’s discipline: a source only goes on the follow-list if a real entry cites it.

The reverse failure mode of underconsumption is aspirational consumption — subscribing to feeds you “should” read but don’t. Aspirational consumption produces guilt + a cluttered feed + nothing else.

The rule: when an entry cites a source, the source goes in following.md with a ★. When following.md gets a source that doesn’t have a ★ for 6 months, drop it (move to a “former sources” section if it was once cited but stopped producing).

The follow-list is evidence of use, not aspiration.

Try this

If you don’t currently have a systematic information diet for this space:

  1. Pick one channel from each tier — one changelog (Tier 1), one newsletter (Tier 2), one named individual (Tier 3), one conference’s post-event coverage (Tier 4), one aggregator (Tier 5). That’s five sources total. Less is more.
  2. Block 30 minutes weekly on the calendar for the skim. Same time every week. Treat it as a meeting with the field.
  3. Maintain a “things to dig into” backlog — a plain text file, an Obsidian note, whatever. When a topic looks interesting but you can’t tell if it’s hot yet, write it down with the date. Revisit weekly. Topics that survive 30 days of mention are entry-worthy.

The point isn’t to consume more. It’s to consume less, more deliberately. A reliable five-source diet outperforms a chaotic 50-source one.

Sources