How to read an AI paper without being a researcher

A few days ago I was researching what it really costs to train a model from scratch, and every summary I found about OPT-175B — the 175-billion-parameter model Meta trained in 2021 — said the same thing in nearly the same words: 992 A100 GPUs, about two months. Full stop. It sounded like the whole story. It wasn't. When I opened the technical paper instead of settling for the summary, section 2.5 — "Training Processes," subsection "Hardware Failures" — told a story no summary carried: at least 35 manual restarts, engineers changing the learning rate on the fly, someone deciding in real time to cut gradient clipping from 1.0 to 0.3 to keep the run from blowing up. That finding wasn't a curiosity — it was the entire argument of my previous piece, and it sat no more than three clicks past where everyone else had stopped reading. Here's the route I followed, told as a method, so you can follow it too.

A paper isn't what people say it says

Between the original finding and what you read, there are usually three layers, and almost no AI coverage gets past the second one. Layer one: the paper, the document the authors actually wrote and published. Layer two: the press release the lab or university writes about the paper, summarizing what it wants told. Layer three: the news story a media outlet writes about that press release, often without ever opening layer one. Every jump loses nuance and gains simplification, and by the third layer the description bears roughly the same relationship to the original finding that a movie trailer bears to the movie.

There's a second, separate distinction: whether the paper has been peer-reviewed or not. Platforms like arXiv, where most recent AI research lives, are preprint repositories: arXiv says so plainly itself — "material is not peer-reviewed by arXiv" — and its moderation only checks that a submission is topical and has "scholarly value," not that its conclusions are correct. That doesn't make a preprint suspect on its own: a large share of serious, careful research is published this way, before or alongside formal review. It does mean "it's on arXiv" isn't the same thing as "other experts have checked it and agreed," and mixing up the two is a common mistake even in coverage that otherwise takes itself seriously.

What to read, and in what order, if you're not reading the whole thing

Almost nobody reads a paper start to finish, and you don't need to. What you need is to know where to look.

The abstract is the sales pitch: the authors wrote it themselves, and it's in their interest that you walk away with the most flattering conclusion. It's rarely a lie, but it's the best-lit version.

After the abstract, it's worth jumping straight to the results — the actual numbers, in their tables, before the prose frames them — rather than reading through the introduction, which usually repeats the abstract in more words.

And then there's the section almost nobody opens, the one that in most papers is called something like "Limitations," "Considerations," or, as with OPT-175B, "Training Processes." That's where the interesting part lives, and not by accident: many academic venues now require an explicit limitations section, so authors are obligated to name, somewhere, what didn't go perfectly. Those sections aren't the fine print of a contract someone hopes you'll skip — they're often the real news, written in plain prose, not in the equations that scare people off.

Two more places are worth knowing about, even though they don't always exist: the appendices, where evidence that didn't fit the main body sometimes ends up, and the "model card" or "datasheet" when a paper includes one — a short form specifically meant to summarize what is and isn't known about the system described. None of these sections require understanding the whole paper to get something out of them: you read them standalone, the way you'd look up one entry in an index, not the way you'd study a chapter.

The case: the 35 restarts no summary mentioned

This is the part you can reproduce yourself in ten minutes, with the same document. Open the OPT-175B paper. Don't reread the abstract a second time: use the table of contents or search for "2.5" to land directly on "Training Processes," and inside it, the "Hardware Failures" subsection. Right there, in two paragraphs of ordinary prose, is everything no echo carried: "hardware failures contributed to at least 35 manual restarts and the cycling of over 100 hosts over the course of 2 months," with roughly 70 additional automatic restarts; when the loss diverged, the team fixed it by "lowering the learning rate and restarting from an earlier checkpoint," with "a number of mid-flight changes" to that rate and a reduction of gradient clipping from 1.0 to 0.3 to stabilize training.

You don't need to understand every technical term in those paragraphs, or know exactly what gradient clipping does, to grasp what they're saying: training that model wasn't turning machines on and waiting, it was watching them in real time and making calls on the fly. The section has a name that says exactly what it contains — nobody hid it — and yet none of the summaries I'd read before mentioned it, not the lab's own write-up and not the coverage that echoed it. It wasn't buried. It was simply past the point where everyone else stopped reading.

When you don't need to do this

Opening the original document shouldn't become a ritual you owe every article you read, and presenting it that way guarantees nobody ever will. If a serious outlet already links its source and the figure it gives doesn't sound suspiciously round or too flattering to whoever announced it, you probably don't need to redo the work. Save it for three moments: when a claim carries no link backing it at all, when the number looks designed to impress rather than inform, or when you're about to base a decision that actually matters to you on that figure. It doesn't need to become a full audit every time, either: it's enough to open the table of contents, find the section with the most procedural-sounding name — "Method," "Limitations," "Considerations" — and read that one piece on its own. It isn't rereading the whole paper more carefully; it's knowing which single paragraph to go to first. And if what you need isn't to read the whole document but to check whether one specific number is real as quoted, that's a different, faster tool — opening the full paper is for when you want to understand the terrain that number lives in, not just confirm it exists.

Next time a headline says "a paper shows that…," you already know what to do with the next ten minutes: not read the whole thing, not trust the summary either, but go straight to the section with the most boring name, the one that promises procedures instead of headlines. That's where the real news has been hiding this whole time, in plain sight, waiting for anyone willing to keep scrolling past the bottom of the abstract.