How to Keep Up With AI Without Losing Your Focus (2026)

In the second week of September, three things happened in four days. On Thursday the 11th the top of Hacker News was an Ask HN titled "Can we please limit the AI news flood?" — 868 points, four hundred comments, and by the afternoon two separate "Hacker News without AI" front pages had been built, posted, and upvoted onto the same front page they were filtering. On Saturday the 12th Dario Amodei published "We Must Pace the Frontier," asking the entire industry to slow the rate at which it improves its own models. On Monday the 14th AI stocks fell on the news — Nvidia down 3%, SoftBank down 11% — and OpenAI said it would not go public this year.

That same Monday a founder I coach told me, without irony, that he spends about ninety minutes a day reading about AI and could not name one thing he had changed because of it in the last month. He is not lazy and he is not stupid. He is doing what the environment asks of him: keeping up. Keeping up has quietly become a job with no output, and the people who feel it most are exactly the ones who cannot afford it — founders and operators whose real work only happens in long, uninterrupted blocks.

This post is the system I actually run. I build AI systems for a living, so my "need to know" is higher than most people's: if I can stay current on forty-five minutes a week, you probably can on less.

The number that reframes the problem

Gloria Mark's group at UC Irvine has been timing how long people stay on one screen before switching since 2003. In the 2004 fieldwork the average was roughly two and a half minutes. In the logging studies from 2016 onwards it settled near 47 seconds, with a median of 40 — half of all screen-attention periods ended within 40 seconds. Mark is careful to say this is not a biological ceiling; it is a description of how information workers now behave in front of an information supply. In Attention Span she also notes that once you get past that first 47 seconds you can actually go deep — the threshold is the whole game.

AI news is engineered, not on purpose but by selection, for the 47-second brain. A launch post, a benchmark chosen by the lab that shipped the model, a founder announcing the model replaced his engineers before lunch, a reply pointing out it cannot count letters. Every item is short, every item feels consequential, and every item is a switch. The cost is not the reading time; it is the resumption tax Mark's research has documented for two decades, paid a hundred times a day.

Cal Newport's Slow Productivity gives the frame I use on the other side of the ledger. Its three principles — do fewer things, work at a natural pace, obsess over quality — are the exact opposite of what a daily AI feed does to you. The feed adds things, it accelerates your pace to the industry's, and it rewards being first rather than being right. So the question is not "how do I read faster." It is "how do I convert an industry running at a frontier pace into inputs I can take at a natural one."

Why filtering AI out is the wrong fix for a founder

The obvious answer is the one Hacker News reached for in a fit of pique: hide it. hcker.news added an ?ai=exclude switch; unslop.news runs an AI classifier to remove stories about AI, which its own author called ironic. One commenter reported the count on the day it launched: "92 of 179 articles survived the filter." The creator replied that on some days it is closer to 75% removed. So a little over half of the front page of the most important technical forum in the world is now AI, and the community's first instinct was a mute button.

I understand the instinct and I think a founder should resist it, for the reason a rebuttal posted the same day put well: the category has broken. "A vulnerability found by an agent is security news. A browser redesigned around an agent is browser news. A chip built for inference is hardware news." Filtering it produces "a tidier front page and a worse account of the world." That is exactly right for anyone whose business is built on software, which in 2026 is most businesses. I covered the fatigue side in what the Hacker News anti-AI backlash means for high performers; the fatigue is real, the signal underneath it is real too, and muting one to fix the other is a bad trade.

The founder's problem is not too much AI news. It is that he consumes it at the wrong unit. He reads at the unit of the headline, and there are a hundred a day. He should be reading at the unit of the decision, and there are perhaps three a month.

The Four-Question Ledger

Here is the filter I apply to every AI item that reaches me, and it takes less time than reading the item. Does this change:

1. What I can build? A capability shift — a model that does something the previous one could not, at a price I can pay, for a task I actually have. 2. What I must protect? An incident, an attack class, a permissions change in a tool that touches my data or acts in my name. 3. What I pay or agree to? Pricing, terms, policy, ads, an incentive change in a product I depend on. 4. Nothing, yet. Everything else: benchmarks without a task attached, launch prose, opinion, and the growing genre of news about the volume of news.

Only the first three earn a line in a ledger I keep — a plain text file, one line per item, dated, with the action it implies. Category four gets nothing, no matter how many points it has. Most weeks the ledger gains two or three lines. Here is what September looked like through it, because the abstract version of any framework is worthless.

Look at the last column. The most-upvoted item of the month earned zero minutes of working attention. The item that earned the most — three frontier models in three days — earned it as a scheduled block with a defined output, not as a week of reading reviews. That asymmetry is the whole method. Social proof measures how many people are switching; the ledger measures whether you should.

The weekly intake, as I actually run it

Daily: nothing. No AI newsletters in the inbox during work hours, no HN before the first deep-work block, no model-release notifications. This is the "work at a natural pace" principle applied literally. The industry's pace is not my pace; the industry is racing itself, and Amodei's essay is a lab CEO saying out loud that even the labs would like to stop. If the frontier can consider pacing itself, a founder in Tallinn can wait until Friday.

Friday, 45 minutes, three inputs. The first is a current model with a search tool and a fixed prompt: what changed since last Friday in the models I use, the tools I pay for, and agent-security incidents — dated, with sources, nothing without a source. I rotate which model runs it, which doubles as the cheapest evaluation I know. The second is one human-curated source — currently the HN front page, once, with the AI filter off, because AI stories are the ones I am there for. The third is a look at the shelf: the release date and training cutoff of the models my agents run, from a page that links every date to the lab document it came from. Output: ledger lines, five at most, and one experiment for the following week, put into the calendar as a real block. Which model gets the digest job, and why the answer is not "whichever is newest," is in Perplexity vs Claude vs ChatGPT for executive research; the calendar half is in Motion or Reclaim for protecting deep work.

Monthly, two hours. Re-read the ledger. Kill any experiment that did not ship. Promote the strategic signals — the Amodei essay, the ads inside ChatGPT, whatever the incident of the month was — from "noted" to "decided," or explicitly to "not deciding yet." Anything that has sat in category four for a month and still feels important gets read properly, in NotebookLM with the primary documents, the way I described in how founders use NotebookLM to learn faster. Slowly, once, is cheaper than skimming it eleven times.

The kill rule. If four consecutive weekly intakes have produced no change in what I build, protect, or pay for, I am not keeping up; I am watching a sport. That is allowed. But it moves to evenings and it stops being called work. Newport's "obsess over quality" cuts both ways: it applies to the reading too.

Where the AI helps here, and where it does not

The digest prompt is the part people copy and it is the part that fails most quietly. A model is a poor source on models, and the dead-king test shows why: over two thousand calls to sixteen models that each had a web search tool available, the frontier models searched when they should and the weaker ones stated settled facts that had changed — one of them with a system prompt that explicitly told it its training ended in February and to search for anything later. The decision to search runs on the same stale weights that hold the stale fact. I went through the practical consequences in does a model's training cutoff still matter if it can search the web. For a weekly digest the fix is simple: demand dates and links, and discard anything without them. An undated claim about AI is category four by definition.

The second failure is subtler. A model can tell you what changed. It cannot tell you what changed for you unless you have told it what you are building, what you are exposed to, and what you pay for — and a model that has not been told will cheerfully rank everything as important, because importance is what the launch posts assert and sycophancy fills the rest. So the ledger questions go into the prompt, in that order, every time. The model's job is retrieval; the judgement stays with me. That division is the same one I keep for decisions in how to get AI to challenge your decisions, and for the same reason.

The third bites me personally. A tool that summarises the week is also a tool for not reading anything, and after a few months of digests I was retaining almost nothing about the tools I had supposedly evaluated. The countermeasure is unglamorous — the one experiment per week is done by hand — and it is the same problem as using AI to actually remember what you read. The digest protects the calendar. It does not build the understanding.

What it costs

I miss things. I found out about Sponsored Agents five days after the announcement, from the Friday digest rather than from the day's discourse. The delay cost me nothing, and the five days of not knowing bought me five uninterrupted mornings. I have been late to two model releases this year that turned out to matter, and early to nothing that did. When a capability shift is real, it is still real on Friday.

The bigger cost is social. Founders keep up with AI partly to be able to talk about it — with investors, with peers, on the panel — and a weekly cadence makes you slightly worse at that on a Tuesday. I have decided that being a day late in conversation is the price of being able to think; the pattern in why AI saves hours but the week never gets shorter is largely made of people who chose the other way and did not notice.

Where this could be wrong

My "need to know" is unusually high — I architect AI systems and coach founders who build on them — so the ledger fills faster for me than for a founder running a logistics company, and the honest recommendation for that founder is a monthly intake, not a weekly one. My evidence is biased: the people who bring this problem to a coach are the ones already drowning, and I do not see the founders who read AI news for an hour a day and ship fine. Newport's frame assumes you control your calendar, which many operators only partly do. And September was an unusually loud month — three frontier releases, a major agent launch, and a lab CEO asking for a slowdown in the same fortnight — so the signal-to-noise ratio in the table flatters the method.

What I am confident about is the failure mode. Gloria Mark's 47 seconds is not a fact about brains; it is a fact about what people do when the supply of small, urgent-feeling items is unlimited, and the AI industry has become the most productive supplier of those items on earth. The people asking Hacker News to turn the flood off are not wrong that it is a flood. They are wrong about the remedy. You do not need less of it. You need to stop drinking it by the headline, and start taking it by the decision — three a month, on Friday, with the rest of the week left alone for the work the news was supposed to help you do.

Sources

Ask HN: Can we please limit the AI news flood? (Hacker News, 11 September 2026): 868 points; "92 of 179 articles survived the filter"; creator: "closer to 75% on some days"
hcker.news — A Better Hacker News Reader (?ai=exclude filter) and unslop.news (classifier-filtered front page), both posted 11 September 2026
Limiting AI news would hide what is happening to software (lord.technology, 11 September 2026): "a tidier front page and a worse account of the world"
Dario Amodei — We Must Pace the Frontier (September 2026): recursive self-improvement; the OpenAI–Hugging Face incident; "we must slow the pace at which we improve the capabilities of AI models"
Los Angeles Times — Tech leaders say they want to slow down the development of AI (14 September 2026): Nvidia −3%, SoftBank −11%, OpenAI IPO delayed
Silicon Canals — from two and a half minutes to 47 seconds: Gloria Mark's screen-attention research (30 August 2026): 2004 CHI fieldwork (14 workers, 477 hours); 2016 logging study (40 workers, median 40 seconds)
Gloria Mark, Attention Span (2023) — "averaging just 47 seconds on any screen"
How stale is your AI? Release age and training cutoff for 20 models (data checked 16 September 2026): GPT-6 Astra 3 Sept, Claude Fable 5.1 1 Sept, Gemini 3.8 Flash 2 Sept; 10 of 20 models publish a cutoff
Does AI training cutoff matter with web search? 16 models, one dead king (September 2026): five of sixteen search-enabled models named a dead king; ~2,000 calls
OpenAI — Reimagining advertising with AI (16 September 2026): Sponsored Agents tested with select US advertisers; HubSpot and Shopify integrations
Meta Newsroom — Introducing Muse (8 September 2026)
Cal Newport, Slow Productivity (2024) — do fewer things, work at a natural pace, obsess over quality