Most people choose AI sources the way tired consumers choose snacks: by visibility, speed, and whatever is easiest to grab at the moment. That is a terrible way to build judgment.
The question is not "who posts a lot?" It is "whose work would still help me if frontier progress paused for six months?" That is the filter. If someone's value disappears the moment launches slow down, they were probably selling recency, not understanding.
This matters because sources do not merely inform you. They train your instincts. Read learning in the step-change with that in mind. The people you read repeatedly become the voice in your head when you decide what is important, what is risky, and what deserves more study. Bad sources do not just waste time. They distort the shape of your curiosity.
I like three source archetypes.
The first is the operator who ships. This person is close to real product or workflow decisions. They talk about tradeoffs, failure modes, costs, user behavior, rollout mistakes, and what changed after contact with a team or customer. They tend to be less theatrical because reality has already punished them enough. When they describe a new model or capability, they anchor it in jobs, constraints, and outcomes.
The second is the researcher or explainer who reduces confusion rather than increasing status anxiety. This person takes technical shifts and makes them legible to non-specialists without flattening the nuance into nonsense. They distinguish what is proven, what is plausible, and what is still speculative. They make you calmer and sharper at the same time. That combination is rare and valuable.
The third is the case writer. This person studies how products and companies actually adopted a capability, where the narrative got ahead of the evidence, and what aged well or badly. Case-based learning matters because it breaks the spell of abstract certainty. GitHub Copilot's adoption curve, Klarna's AI deflection, and Harvey's legal AI are valuable not because they settle every debate, but because they show how claims fare in the wild.
Now for the source types I would downgrade hard.
First, the frontier-poster. This person builds identity around being early. Their content is usually framed as access: the latest thing, the hidden trick, the future before everyone else sees it. Sometimes they surface useful material. More often they train you to overvalue novelty.
Second, the obituary account. This is the "X is dead" operator. Search is dead. Prompting is dead. SaaS is dead. Junior engineers are dead. These accounts monetize absolutism because absolutism travels. You can learn one thing from them: if someone keeps declaring entire categories dead, they are not doing analysis. They are manufacturing attention.
Third, the summary aggregator. This person reads the real sources and turns them into a smoother feed product. That may be useful occasionally. It is not where you should build your base. If you only consume summaries, you inherit their blind spots and never notice.
The recommendation here is not to create a massive reading roster. Quite the opposite. Build a small barbell: two or three durable sources you revisit regularly, plus a handful of tactical sources you scan lightly for change detection.
Lever: a small, high-trust source portfolio improves your calibration because the same few voices keep teaching you how to reason, not just what to notice.
Risk: a narrow source list can become an echo chamber if all your chosen sources share the same assumptions.
Rollback: deliberately include sources with different vantage points. For example, combine one operator, one explainer, and one case-oriented source. Diversity of method matters more than diversity of posting style.
You should also run periodic source audits. Ask four questions.
- Does this source make me better at decisions, or just better at small talk?
- Does the source show work, evidence, or concrete tradeoffs?
- Would I still value this person's writing if launches slowed down?
- After consuming this source, do I feel clearer or merely more stimulated?
That last question is underrated. A lot of modern AI media is stimulation masquerading as education. It gives you the sensation of relevance without the substance of orientation.
This is where naming people can become a trap. Specific names date quickly, and the skill is not celebrity worship anyway. The skill is source archetype recognition. You want the ability to see patterns before you see profile photos.
A useful exercise is to keep two lists: a trust list and an ignore list. The trust list is small and earned. The ignore list is not a moral blacklist; it is a boundary. If a source repeatedly produces urgency without evidence, or wide claims without rollback logic, move on. You do not need to stay open to everyone forever.
Notice that strong sources usually think in lever, risk, and rollback terms even when they do not use those exact words. They tell you what adopting a view buys you, what could go wrong, and how to back out if the bet proves too broad. Weak sources skip the rollback because certainty performs better.
This filter also helps inside companies. When a teammate forwards a hot take and says, "Should we react to this?" you can ask, "What kind of source is this? Operator, explainer, case writer, or frontier theatre?" That question often clears the room faster than arguing about the claim itself.
Trust is not about perfect foresight. No source gets everything right in a step-change. Trust is about error quality. When they are wrong, are they wrong in ways that still help you think? Do they update cleanly? Do they separate evidence from ego? Those are durable signals.
Build your source system with the same seriousness you would use for hiring or vendor selection. It is a long-duration decision. The people you keep reading will shape what you notice, and what you notice will shape what you learn next.
Rules from this lesson
- Prefer sources that would still be useful if AI progress paused for six months.
- Build around three archetypes: the operator who ships, the researcher who explains, and the case writer.
- Downgrade frontier-posters, obituary accounts, and summary aggregators as core teachers.
- Audit sources by their effect on your decisions and calibration, not by their reach or posting speed.