Following AI Research Without a PhD

The volume of machine learning publishing exceeds what anyone reads, including full-time researchers, who cope by ignoring nearly all of it outside a narrow area. If your plan is to keep up, the plan is already wrong. What works is a filter you trust enough to stop worrying about what it discards.

The premise: you will miss things

Two facts make the anxiety unnecessary. First, results that matter are repeated: anything genuinely important will reach you through several independent channels over the following months, because everyone will be building on it. Missing a paper on the day it appears costs you nothing except the feeling of being early.

Second, most papers do not matter to you. This is not a judgement about quality — it is a consequence of specialisation. A paper improving a training technique for a model class you will never train is excellent work and irrelevant to your week. Reading it is a hobby, not professional development, and it is worth being honest about which you are doing.

The goal is therefore not coverage. It is to notice, within a month or two, the small number of results that change what you would build.

Two hours a week

Set the budget first and design inside it. A shape that works, and that can be scaled up or down without changing its structure:

The written record does a second job that only becomes apparent after a few months: it shows you what you keep reading about and never acting on, which is usually a sign that the topic is interesting to you rather than useful to you. That is worth knowing, and it is fine either way.

The filter

Applied at the skim stage, in order. Each question takes seconds.

• Could this change a decision I have made or will make? Not “is it interesting”. Interesting is unbounded; decision-relevant is not.
• Is it a method I could adopt, or a fact I should know? Both qualify. A paper that is neither is entertainment, which is allowed, on a different budget.
• Has it been checked by anyone? An artefact, a reproduction, or several independent people describing having used it. For anything you might act on, this outweighs novelty.
• Is the excitement about the result or the framing? If the discussion is entirely about what the result implies rather than what it measured, wait a month. The reaction will settle and the careful summaries will exist by then.
• Would waiting cost anything? Almost always no. Waiting is an underrated filter: it lets other people do your pass one for free.

Source types, and what each is good for

Specific sources change and lists of them go stale, so the durable thing to internalise is what each kind of source can do for you and where it fails.

• Category feeds from the preprint server. Complete and unfiltered. Good as a raw input, hopeless as a reading list — the volume is the problem you are trying to solve, not the solution.
• Curated newsletters. The best value per minute available, because someone competent has done your skim. Failure mode: you inherit their interests and their blind spots, so read the ones that link the paper rather than only summarising it, and click through often enough to keep the summariser honest.
• Open review pages. Underused and excellent. The reviews tell you the objections of people who know the area, which is information the paper cannot contain.
• Reading groups. The highest-retention format by a distance, because explaining a paper out loud is where you discover you did not understand it. Two colleagues and a recurring slot is enough.
• Social feeds. Fast, and optimised for reaction rather than accuracy. Useful for noticing that something exists; unsuitable for deciding what it showed, because the thread is written against v1 and never updates. See why the version history matters.
• Survey papers. The most efficient way into an unfamiliar area, and the correct move when you realise you are three years behind on a subject you now need. One good survey replaces twenty papers.
• Release notes and model cards. Not research, and often more decision-relevant than research if you are building on hosted models.

Where the depth goes

The habit that separates people who understand the field from people who can recite this week’s papers is counter-intuitive: spend most of your depth budget on old work.

Foundational papers have three advantages over new ones. They have been checked, so you are not spending effort on results that will not survive. They are prerequisites for reading anything current, so the investment compounds rather than depreciating. And there are far fewer of them, so the set is finishable — which is the opposite of the relationship you have with the arXiv feed. A structured way through them is the annotated reading list.

And one permission, since the anxiety is the real problem this page is addressing. Nobody is keeping up. The people who appear to be are specialists in one narrow area plus consumers of the same summaries you have access to. A defensible two hours a week, sustained for a year, leaves you better informed than an unsustainable ten hours a week abandoned in March.