AI Watch: Our Method for Staying Informed Without Drowning

The StartingUp Summer keeps taking the long view. Mapping out a year this dense calls for an AI watch that goes the distance: how do you stay informed without drowning in it? After the questions AI made obsolete, here is the one it asks every Monday morning.

A model that outclasses the previous one within months, tools that are born, pivot or disappear in a single quarter, use cases that reinvent themselves from one demo to the next: the pace of AI news probably has no equivalent in the history of our trades. Even spending entire days on it, nobody keeps up with everything. We are an agency whose practice AI has transformed, and we accept letting dozens of announcements go by unread every week.

That's not negligence, it's a method. Because the real risk isn't missing an announcement: it's FOMO, that fear of missing out which scatters teams, stacks up subscriptions and pushes toward gadget decisions, the kind where you adopt a tool because it's making noise rather than because it solves a problem. Here is how we organize our watch, and above all what we've deliberately chosen not to follow.

Doing effective AI watch comes down to four decisions: telling apart three levels of news (research, products, your own use case) and seriously following only the last one; sticking to a small number of durable sources rather than a feed; keeping only what can be tested on a real problem within a week; and only sharing with the team what has been verified. The rest of this article breaks down each of these choices.

The first mistake is treating "AI news" as a single block. In reality it plays out on three tiers that don't deserve anywhere near the same attention.

Fundamental research is fascinating, and that's exactly its trap. If a breakthrough is real, it will land in a product, with a button and documentation, within the following months: that's the moment it will concern you. An executive or a marketing team reading research papers spends attention it can't put to any use.

The products and models tier does deserve some follow-up, but by sampling. Genuinely important information has one reassuring property: it will reach you more than once. If an announcement is still being commented on by serious sources a week after publication, it deserves a read. If it has vanished, the filtering already happened on its own. The last tier, your own use case, reverses the logic: there, waiting no longer makes sense, and it's up to you to go and fetch the information.

How do you decide that something new falls into your own use case? We apply a simple rule: an announcement only deserves our attention if we can test it on a real problem within the week. Not a demo run on a toy example: a real problem, drawn from a live project, with an observable result.

This filter has two virtues. It turns watch into first-hand knowledge: what we know about a tool comes from confronting it with our own projects, not from a launch video. And it lets us ignore everything else without guilt. The evolutions of coding agents, for instance, are testable the same day on our own repositories: we follow them closely, and our real workflow with these agents is the direct product of that loop. Conversely, a spectacular announcement that meets no concrete problem at ours or at our clients' simply joins a waiting list, one we'll reopen the day the problem exists.

The natural reflex when facing a fast-moving field is to subscribe to the fastest feed: social media. That's the worst possible choice. AI influencer threads obey an attention economy that structurally rewards overpromising: every tool there "changes everything," every model there "buries" the previous one. We prefer a small number of durable sources, chosen by type:

• The labs' official blogs (Anthropic, OpenAI, Google DeepMind): the announcement at the source, without the layer of interpretation. You read what the tool does, not what you hope it will do.
• One curation newsletter, and only one: dense daily digests like TLDR AI or The Rundown AI do the sweeping work for you, in a one-line-headline-and-two-sentences format. Picking one is enough: they largely cover the same launches.
• One or two practitioners who actually test things: blogs like Simon Willison's, who has been documenting his experiments for years, are worth more than ten commentators. The selection criterion stays constant: does this person show what they tried, or comment on what others have announced?

The names may come and go, the types will stay: a primary source, a curation service, a practitioner's eye. Three well-chosen subscriptions cover the essentials; the tenth only adds noise.

There's a certain justice in turning the tool against the problem it creates. AI assistants have become part of our watch setup: summarizing a twenty-page announcement into ten lines, comparing two tools on the criteria that matter to us, and above all questioning the news from within our own context ("what does this announcement actually change for a headless WordPress site?"). That last query is the most valuable one: it turns generic information into situated information.

With one limit you need to know about: freshness. A language model doesn't know its own current events; its knowledge stops at a training cutoff date, and it can describe a months-old landscape with perfect confidence. For watch purposes, systematically use the modes with web search, demand sources, and date whatever you're told. An undated summary, in AI, is stale information that doesn't know it.

What's left is keeping watch from becoming an individual matter again, with everyone piling up tabs in their own corner. Our answer comes down to a ritual: a short check-in, at a fixed cadence, fifteen minutes is enough, with a single rule of admission: we only share what we've tested. No "I saw something amazing go by," but "I tried this tool on this project, here's what it gives, here's what it costs."

The effect is twofold. Watch becomes verified collective knowledge, instead of internal rumor. And FOMO dissolves on its own: if nobody on the team found it worth testing something new within a week, its urgency was an illusion of relevance. The ritual serves as much to filter as to inform.

We're regularly asked "how do you manage to follow everything?" We don't follow everything, and no one does: those who claim to are confusing exposure with understanding. Missing an announcement rarely costs anything: if it matters, it will come back, and you'll find it again the day it meets one of your problems. Scattering your attention, on the other hand, always costs something.

That's why watch, properly understood, isn't a consumption of news: it's an attention policy. Deciding what to ignore, choosing your sources the way you choose your tools, keeping only what survives contact with a real problem, and sharing it once verified. In a field where everything changes from one week to the next, this discipline is paradoxically what changes the least: it's this, not the list of subscriptions, that's worth passing on.