Staying Current Without Drowning

You cannot keep up. Not as a personal failing — arithmetically. More is published each week than anybody can read, and the fraction of it that will change what you do is small and not correlated with how loudly it is discussed.

Why the feed cannot be kept up with

The material that reaches you has been selected for engagement, not for relevance to your work. That selection is not malicious; it is just what a feed optimises. The consequence is a systematic bias toward the novel, the dramatic and the speculative, and away from the incremental and the boring — which is where almost all of the material that would actually change your decisions lives.

There is a second, sharper problem. Reading about a capability is not the same as being able to use it, and the feeling of being current that comes from reading is almost entirely disconnected from the ability to build. A person who read every release note for a year and a person who shipped three systems have very different knowledge, and only one of them has the kind that transfers.

So the goal is not coverage. The goal is to not be surprised by something that affects you, while spending an amount of time you can sustain — which for most working engineers is somewhere under two hours a week.

The filter: three questions

Apply these to anything competing for your attention. It takes about fifteen seconds and discards the large majority of what arrives.

• 1. Would this change a decision I have already made, or one I am about to make? Not “is this interesting”. A new model is relevant if you are choosing a model; a new retrieval technique is relevant if your retrieval is the bottleneck. If you have no decision it touches, it is entertainment, and you may read it as entertainment as long as you know that is what you are doing.
• 2. Can I check it? Is there a method, a repository, a number with a sample size, something reproducible? A claim you cannot check is not knowledge you can act on. This question alone removes most of what circulates.
• 3. Will it still be true in six months? A structural fact — how attention scales, why output costs more than input, why retrieval fails the way it does — pays back forever. A fact about which model currently leads a benchmark expires almost immediately and can be looked up when needed rather than remembered.

The three questions have a useful property: they are increasingly strict. Most things fail the first. Anything that passes all three is worth an hour, and there are perhaps a handful of those a month.

Primary sources, and why they are shorter

This page will not hand you a list of newsletters and accounts, because such a list decays — the people change, the quality changes, and a recommendation made now cannot be vouched for in a year. What does not decay is the category of source, and the categories worth your time share one property: they are written by whoever actually did the thing.

• Provider documentation and changelogs for the models and services you actually use. This is the highest-value category by a wide margin, because it is the only one that tells you about a change to something you depend on. Subscribe to the changelog; ignore the marketing announcements.
• Model cards and system cards for models you might adopt, read with an eye for what they omit as much as what they claim.
• The paper itself, when a result matters to you, rather than the summary of the summary. Papers are less work than they look if you read them in passes and stop at the point you have the answer you came for.
• Release notes and issue trackers of the open-source tools in your stack. Dull, specific, and the place where the problems you are about to have are already being discussed.
• Post-incident and post-project write-ups from teams describing something they ran. Rare, disproportionately valuable, and identifiable by the presence of numbers and admitted mistakes.
• One or two aggregators of your own choosing, used as a discovery layer only. Their job is to point at primary sources; if you find yourself reading the aggregator instead of what it points at, it has stopped working.

Judge any individual source you are considering by a single test: how often does it say “I was wrong” or publish a negative result? Sources that never do are optimising for something other than accuracy.

A weekly and a quarterly pass

Two loops, deliberately different in character. Most people attempt only the first and then wonder why their mental model is a year old.

Weekly, thirty minutes. Skim changelogs for the things you depend on. Note anything that would change a decision. Resist reading anything that fails question one. The output of this pass is usually nothing, and nothing is the correct output — its purpose is to catch the two or three changes a year that matter to you, not to produce material.

Quarterly, half a day. This is the pass that actually updates you. Pick one thing you believe about the field and check whether it is still true — by building something small, not by reading. Re-run an evaluation you already have against a current model. Rebuild a piece of your system with an approach you dismissed a year ago. The output is a changed belief with evidence behind it, which is worth more than a quarter of feeds.

A specific and cheap version of the quarterly pass: keep a small frozen evaluation set from real work, and run it against whatever is new. It converts “is this model better” from a reading exercise into a measurement, and it is the same set your regression tests already use.

What to deliberately not read

• Benchmark leaderboard movements. Interesting as sport, and they do not transfer to your task. Look one up when you are choosing; do not track it.
• Capability demonstrations without a method. A screenshot of an impressive output is one sample from a distribution, selected for being impressive.
• Predictions about the next two years. Nobody has been reliably right, they are unfalsifiable at the moment of publication, and acting on one has no expected value.
• The same news, retold. A release gets covered dozens of times; the marginal value of the third retelling is negative because it feels like new information.
• Anything about a model you have no ability to use. It will still be there when you can.

Permission, since it is rarely given explicitly: falling behind on the discourse is not falling behind on the work. The structural knowledge that makes someone good at this — how retrieval fails, what a failure costs, how to tell whether a change helped — has been stable for years, and the exhaustion of trying to track everything has a real cost that nobody puts on the other side of the ledger.