{"course":{"slug":"ai-careers","title":"Working in AI","url":"https://addaly.com/learn/ai-careers","level":"beginner","track":"work"},"lesson":{"slug":"keeping-up-without-drowning","title":"Keeping up without drowning","url":"https://addaly.com/learn/ai-careers/keeping-up-without-drowning","position":81,"of":85,"minutes":7,"answer":"Follow problems and primary sources rather than launches, because almost nothing announced this month changes what you should do this month.","keyIdea":"Follow problems and primary sources rather than launches, because almost nothing announced this month changes what you should do this month."},"body":{"format":"markdown","text":"## Why it feels urgent\n\nThe volume is enormous and it is not slowing. But notice the mechanism behind the feeling: launch announcements are written to feel urgent, and the person posting is rewarded for your attention rather than for your career. Your sense of falling behind is the product being sold, and it is sold by people with model releases, courses and newsletters to promote.\n\nOnce you can see that, the noise becomes much easier to ignore without guilt.\n\n## The three-source rule\n\nPick three sources you actually read, and let everything else reach you by accident.\n\n- **One primary.** The documentation and release notes of the tools you actually use. Model provider docs. The paper, when there is one, rather than the thread about the paper.\n- **One practitioner.** Somebody who ships things and writes about what broke. Their failures are worth ten launch posts.\n- **One local.** Your country's or language's community, because hiring, pay and regulation are local and the global feed will never tell you about them.\n\nThree. Everything else you sample quarterly, which is often enough for anything that survives a quarter.\n\n## The test for whether something matters\n\n**Would this change a decision I am making this quarter?**\n\nIf no, note the name and move on. You can learn any tool in a week when it becomes relevant to you. You cannot recover the six months spent skimming announcements about tools you never used.\n\nThis rule feels reckless the first time you apply it and is almost always right. The important developments do not disappear; they keep appearing, in your primary sources, attached to problems.\n\n## What is worth tracking\n\nFour things, because they change decisions rather than opinions:\n\n- **Prices.** They fall unevenly and they change build-versus-buy arithmetic. A feature that was uneconomic last year may be trivial now.\n- **Context limits and capability changes** in the models you actually build on.\n- **Deprecations.** This one bites hardest and gets the least attention: a provider retiring a model version breaks an application that has run untouched for a year. Providers publish deprecation schedules; check them for anything you have in production, and diarise the dates.\n- **Licence changes on open models and libraries.** An open weight model whose licence changes can make a shipped product non-compliant overnight.\n\n## The half-life argument\n\nSome knowledge decays fast and some barely decays.\n\n**Fast:** the API surface of a specific framework, the current best model, the fashionable orchestration library, prompt tricks tied to one model version.\n\n**Slow:** SQL. Statistics. Debugging. Writing clearly. How data actually gets messy. Your domain: how insurance claims work, how a hospital admits patients, how freight is priced.\n\nNearly everyone spends their learning time on the fast-decaying half because it is what gets announced. Deliberately spending most of yours on the slow half is the entire strategy of this lesson, and after five years the difference between two people who studied equally hard is almost entirely this allocation.\n\n## A monthly routine that fits a working life\n\nTwo hours, once a month:\n\n1. **Thirty minutes:** release notes and deprecation pages for the things you actually run.\n2. **Thirty minutes:** one long-form piece or paper, read properly rather than skimmed.\n3. **One hour:** build something tiny with one new thing. Not a tutorial: the smallest possible real use.\n\nThat is it. It sounds inadequate against the volume of what exists, and it produces more durable capability than daily skimming, because the hour of building is the only part that becomes knowledge.\n\n## The free path\n\nEverything you need is free and primary: provider documentation, arXiv, model cards on Hugging Face, project changelogs, and the community for your language or region. The paid layer (newsletters, courses, video subscriptions) mostly repackages those sources with a delay and a personality attached. Some of it is good. None of it is where the information originates, and paying for it does not reduce the volume problem.\n\n## The limitation, stated honestly\n\nThis approach trades novelty for depth, and it will occasionally make you late to something real. You will hear about a important shift a few months after the people who read everything.\n\nThat is the correct trade for a career, and the cost is smaller than it sounds: the real shifts persist for more than a quarter, so they reach you anyway, at the point where they have stopped being speculation and started being something you can use."},"check":{"question":"A significant new model is released with capabilities beyond anything you have used. You are three weeks from delivering a project on the current one. What is the sound response?","options":[{"text":"Evaluate switching immediately, since delivering on an outdated model wastes the work.","correct":false,"tempting":"Treats novelty as obsolescence, and reopens a working design three weeks from delivery on the basis of an announcement rather than a measurement."},{"text":"Deliver as planned and avoid the new model until it has been in the market for a year.","correct":false,"tempting":"Turns a sensible focus rule into a blanket avoidance, which is how teams end up two generations behind and paying for it."},{"text":"Add it as a fallback provider now, so the project benefits without changing the main path.","correct":false,"tempting":"Adds an untested dependency and a second code path to a project three weeks from delivery, in exchange for no measured benefit."},{"text":"Finish the project, note the release, and after delivery run your existing evaluation set against both models: a measured comparison on your own cases is the only thing that should move a production decision.","correct":true}]},"prev":{"slug":"what-gets-automated","title":"Which parts of this work are most exposed","url":"https://addaly.com/learn/ai-careers/what-gets-automated"},"next":{"slug":"the-refusals","title":"The things you will be asked to build","url":"https://addaly.com/learn/ai-careers/the-refusals"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Keeping up without drowning', in Working in AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-careers/keeping-up-without-drowning (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Keeping up without drowning","courseTitle":"Working in AI","module":{"index":9,"slug":"staying","title":"Staying: the first ninety days and the ten years after"},"answer":null,"keyIdea":"Follow problems and primary sources rather than launches, because almost nothing announced this month changes what you should do this month.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"why-it-feels-urgent","runs":[{"text":"Why it feels urgent"}]},{"t":"p","runs":[{"text":"The volume is enormous and it is not slowing. But notice the mechanism behind the feeling: launch announcements are written to feel urgent, and the person posting is rewarded for your attention rather than for your career. Your sense of falling behind is the product being sold, and it is sold by people with model releases, courses and newsletters to promote."}]},{"t":"p","runs":[{"text":"Once you can see that, the noise becomes much easier to ignore without guilt."}]},{"t":"h2","id":"the-three-source-rule","runs":[{"text":"The three-source rule"}]},{"t":"p","runs":[{"text":"Pick three sources you actually read, and let everything else reach you by accident."}]},{"t":"ul","items":[[{"text":"One primary.","b":true},{"text":" The documentation and release notes of the tools you actually use. Model provider docs. The paper, when there is one, rather than the thread about the paper."}],[{"text":"One practitioner.","b":true},{"text":" Somebody who ships things and writes about what broke. Their failures are worth ten launch posts."}],[{"text":"One local.","b":true},{"text":" Your country’s or language’s community, because hiring, pay and regulation are local and the global feed will never tell you about them."}]]},{"t":"p","runs":[{"text":"Three. Everything else you sample quarterly, which is often enough for anything that survives a quarter."}]},{"t":"h2","id":"the-test-for-whether-something-matters","runs":[{"text":"The test for whether something matters"}]},{"t":"p","runs":[{"text":"Would this change a decision I am making this quarter?","b":true}]},{"t":"p","runs":[{"text":"If no, note the name and move on. You can learn any tool in a week when it becomes relevant to you. You cannot recover the six months spent skimming announcements about tools you never used."}]},{"t":"p","runs":[{"text":"This rule feels reckless the first time you apply it and is almost always right. The important developments do not disappear; they keep appearing, in your primary sources, attached to problems."}]},{"t":"h2","id":"what-is-worth-tracking","runs":[{"text":"What is worth tracking"}]},{"t":"p","runs":[{"text":"Four things, because they change decisions rather than opinions:"}]},{"t":"ul","items":[[{"text":"Prices.","b":true},{"text":" They fall unevenly and they change build-versus-buy arithmetic. A feature that was uneconomic last year may be trivial now."}],[{"text":"Context limits and capability changes","b":true},{"text":" in the models you actually build on."}],[{"text":"Deprecations.","b":true},{"text":" This one bites hardest and gets the least attention: a provider retiring a model version breaks an application that has run untouched for a year. Providers publish deprecation schedules; check them for anything you have in production, and diarise the dates."}],[{"text":"Licence changes on open models and libraries.","b":true},{"text":" An open weight model whose licence changes can make a shipped product non-compliant overnight."}]]},{"t":"h2","id":"the-half-life-argument","runs":[{"text":"The half-life argument"}]},{"t":"p","runs":[{"text":"Some knowledge decays fast and some barely decays."}]},{"t":"p","runs":[{"text":"Fast:","b":true},{"text":" the API surface of a specific framework, the current best model, the fashionable orchestration library, prompt tricks tied to one model version."}]},{"t":"p","runs":[{"text":"Slow:","b":true},{"text":" SQL. Statistics. Debugging. Writing clearly. How data actually gets messy. Your domain: how insurance claims work, how a hospital admits patients, how freight is priced."}]},{"t":"p","runs":[{"text":"Nearly everyone spends their learning time on the fast-decaying half because it is what gets announced. Deliberately spending most of yours on the slow half is the entire strategy of this lesson, and after five years the difference between two people who studied equally hard is almost entirely this allocation."}]},{"t":"h2","id":"a-monthly-routine-that-fits-a-working-life","runs":[{"text":"A monthly routine that fits a working life"}]},{"t":"p","runs":[{"text":"Two hours, once a month:"}]},{"t":"ol","start":1,"items":[[{"text":"Thirty minutes:","b":true},{"text":" release notes and deprecation pages for the things you actually run."}],[{"text":"Thirty minutes:","b":true},{"text":" one long-form piece or paper, read properly rather than skimmed."}],[{"text":"One hour:","b":true},{"text":" build something tiny with one new thing. Not a tutorial: the smallest possible real use."}]]},{"t":"p","runs":[{"text":"That is it. It sounds inadequate against the volume of what exists, and it produces more durable capability than daily skimming, because the hour of building is the only part that becomes knowledge."}]},{"t":"h2","id":"the-free-path","runs":[{"text":"The free path"}],"role":"free"},{"t":"p","runs":[{"text":"Everything you need is free and primary: provider documentation, arXiv, model cards on Hugging Face, project changelogs, and the community for your language or region. The paid layer (newsletters, courses, video subscriptions) mostly repackages those sources with a delay and a personality attached. Some of it is good. None of it is where the information originates, and paying for it does not reduce the volume problem."}]},{"t":"h2","id":"the-limitation-stated-honestly","runs":[{"text":"The limitation, stated honestly"}],"role":"limit"},{"t":"p","runs":[{"text":"This approach trades novelty for depth, and it will occasionally make you late to something real. You will hear about a important shift a few months after the people who read everything."}]},{"t":"p","runs":[{"text":"That is the correct trade for a career, and the cost is smaller than it sounds: the real shifts persist for more than a quarter, so they reach you anyway, at the point where they have stopped being speculation and started being something you can use."}]}],"check":{"question":[{"text":"A significant new model is released with capabilities beyond anything you have used. You are three weeks from delivering a project on the current one. What is the sound response?"}],"options":[{"text":[{"text":"Evaluate switching immediately, since delivering on an outdated model wastes the work."}],"correct":false,"tempting":[{"text":"Treats novelty as obsolescence, and reopens a working design three weeks from delivery on the basis of an announcement rather than a measurement."}]},{"text":[{"text":"Deliver as planned and avoid the new model until it has been in the market for a year."}],"correct":false,"tempting":[{"text":"Turns a sensible focus rule into a blanket avoidance, which is how teams end up two generations behind and paying for it."}]},{"text":[{"text":"Add it as a fallback provider now, so the project benefits without changing the main path."}],"correct":false,"tempting":[{"text":"Adds an untested dependency and a second code path to a project three weeks from delivery, in exchange for no measured benefit."}]},{"text":[{"text":"Finish the project, note the release, and after delivery run your existing evaluation set against both models: a measured comparison on your own cases is the only thing that should move a production decision."}],"correct":true,"tempting":null}],"order":[2,1,3,0],"why":null},"prev":{"slug":"what-gets-automated","title":"Which parts of this work are most exposed","minutes":7},"next":{"slug":"the-refusals","title":"The things you will be asked to build","minutes":8}}}