What's new

The rule about advertisers who are also tools we teach has changed from a ban to disclosure. Yesterday an advertisement could not be for an AI product or any vendor a lesson names. From today anyone may advertise, and what is never allowed is hiding the conflict from the learner it could affect. An advertiser that a lesson or tool card also describes is named as an advertiser beside that description, on the card and in every lesson that mentions it, while the advertisement runs and for 90 days after, and its advertisement never appears on a page about it. Every advertiser and promoter is listed on a public register on the Support page. Learners still never pay and nothing is behind money. No advertisement is shown or approved.

Sky is now livelier, and every part of it is yours to turn down. It can glow, flash and play one short quiet chime when you do something, cheer when you open it, say kindly when a streak is at risk, show a pace timer for a quick quiz, and show an opt-in leaderboard and a daily bonus star. You can drag it anywhere on the page, or move it by keyboard, and reset it. There are three levels of reaction (calm, gentle, lively), calm is chosen automatically when your device asks for reduced motion, flashes are capped well below the safe limit, and nothing ever cuts a timer short. Points and streaks still come only from learning, never from opening Sky. The leaderboard shows a generated handle, never a name; the practice pals that fill it are made up, change each day, are each labelled "Practice pal" and are never counted as learners. The welcome clip on the home page now plays in Safari and on iPhones as well. Badges you earn float and celebrate when you open them.

When read-aloud is on, you can now choose the voice, grouped as Indian and Global, with a play-a-sample button and a speed control. These are the voices your own device has, ranked so the most natural one comes first; nothing is sent anywhere. Recorded narration by people is supported but none has been recorded yet, and the page says so rather than implying otherwise.

Spread the word and Feedback are now in the top menu, and Share has a button for X (Twitter) alongside WhatsApp, Telegram, LinkedIn, Reddit and email. A Rate us on Google button appears on the Feedback page once Lrnon has a Google review link. The Privacy page now describes the one anonymous page count the site keeps (no cookie, identifier, IP address or browser string; not sent if your browser sends Do Not Track), and the sign-in record kept for people with accounts.

Each community section now shows its real state: planned, closed for now, or open. Administrators can open or close sections with one click. Anything where learners write, such as Ask Doubts and rooms, cannot be opened until the child-safety conditions hold, and the database refuses otherwise.

Sky now opens properly on an iPhone and on other phones. On a phone the panel is a sheet from the bottom of the screen, and a styling mistake had placed it a whole screen above the top edge, so Sky seemed to vanish when tapped. Dragging Sky now works by finger as well as by mouse. Sky’s panel no longer lets one long word or wide row push its content off to the side. "Ask a Homework Question" is now "Ask a Question".

Sky’s sounds were silent because browsers only start sound after a touch, and Sky asked too late. They now work, and are on by default, quiet, with a speaker button in Sky’s header to turn them off in one tap. Sky also has a voice of its own: it borrows the brightest natural voice on your device, tunes it, and opens with its own chirp, or uses a recording once one exists. You can choose the voice and hear it in Sky settings. Sky never talks over a lesson being read aloud.

Translation has been built, under one rule: nothing machine-translated is shown to a learner until a person who reads the language has checked it and put their name to it. A translation made from English that has since changed comes off the site until it is checked again. A language picker appears in the header once any language has something checked. The Languages page says where each language stands. No language is live yet; a Hindi draft is waiting for a reader of Hindi.

A volunteer given a role could not open anything: the console let in only owners and admins. Every role now has a desk that says what it covers, and opens the pages it is for. Admins can give and end the four volunteer roles (assessment reviewer, moderator, editor, language reviewer) from the user portal, with a stated purpose and an end date within a year, recorded with their name. Pages open faster: the page is no longer held up waiting for a stylesheet, Sky’s picture is a fraction of the size, and the visit counter no longer waits on the database.

Lrnon is now open to funding, and the earlier promise of no advertising and no sponsorship of any kind has been replaced. Until today Lrnon was paid for by its founder alone and promised never to take advertising or sponsorship of any kind. It now accepts donations from anyone, learners included, grants, and advertising under published rules. What did not change is the part that matters to a learner: nobody pays to learn, nothing is behind money, and giving unlocks nothing. What changed in the promise is stated on the Support page. Nobody can pay to be in a lesson, a tool card or a comparison. An advertisement may not be for an AI product or any vendor the lessons name, may not track anyone, and may appear only on a few browsing pages, never in a lesson, quiz, review, assessment, the accessibility panel or Sky. Those rules are checked by tests that fail the build. Nothing has been raised, no advertisement is shown or approved, and Lrnon is not a registered charity. The Privacy page now also says what happens to a question typed to Sky's question-answering feature, and the Safeguarding page now describes advertising and Sky as they are.

Sky, the green owl-shaped robot, is now a study buddy on every page. It greets you, suggests one next step with how long it takes, and shows your own progress. It follows your light or dark theme, the high-contrast setting, the dyslexia-friendly font, text size and line spacing, and it can read its message aloud with your device's own voice. It never flashes, glows or plays sound, its one slow movement stops when your device or Lrnon is set to reduce motion, and the streak and level numbers can be switched off. Sky awards nothing, so opening it earns nothing, and it has no leaderboard, timer or talk of losing a streak. The part of Sky that answers written questions is a separate thing and is still open only to the Lrnon team.

A freshness review of every Explorer and Practitioner lesson, the first since the tool descriptions were last checked in August. Four statements were wrong or had aged and were corrected. E2·L11 said most open-weight models are released under a custom licence with conditions; several labs now use permissive licences such as Apache 2.0 or MIT, so the lesson now says the terms vary by model and version and must be read for the version in use. E2·L9 said free tiers run last year’s flagship model, which is no longer reliably true, and now says a smaller or earlier model than the vendor’s best. E2·L2 said ChatGPT reached a hundred million users in late 2022; it launched then and was reported to pass that mark within about two months. E8·L2 now says the EU AI Act’s transparency duties are being phased in on different dates. Card summaries for Meta AI, Llama, Grok, Microsoft Copilot, GitHub Copilot and Notion AI were corrected: Meta AI is no longer described as built on Llama, xAI is now part of SpaceX, Microsoft is merging its personal and work Copilot apps, and GitHub Copilot is no longer called the most widely used. Lessons that named tools without citing them now declare them, and P8 now cites two official sources it previously could only describe as unread: the Government of India’s press release on the DPDP Rules, and Google’s own Gemini API terms, whose paid-use wording was previously seen only in a third-party copy. Seven lessons that name a law with no cited source are listed in the weekly report as still outstanding. Lessons can now carry dated outside sources, and a weekly check watches tool pages and those sources, ages lessons as well as tool cards, and reports to a single issue instead of staying silent behind an old one.

When Not to Automate published, which completes the Practitioner track at eleven modules. Eight lessons on the decision that comes before everything else in the track. Five kinds of work worth leaving with a person, namely where the judgement, the relationship, the irreversibility, the rarity or the protective friction is the point; fixing a process before building around it, since automation preserves flaws and makes them harder to change; the skill that fades when a reliable automation stops being practised by hand; why accountability does not automate, and what must stand behind an automated decision; the people on the receiving end, and the comfort test of whether they would be unbothered knowing exactly what is happening; switching an automation off, with a stop control tested while running and a manual path that still works; things built and deliberately not switched on, and where such conditions should be enforced; and a one-page decision record written before building. Lrnon is the worked example throughout, including a candid contrast between two launch conditions enforced in the database and one, widening the assistant, enforced only in the console. The Builder track has one module planned and none published.

Vendor Risk and Portability published — eight lessons on what happens over a long enough time when the companies a workflow depends on stop behaving as assumed. Drawing the dependency map in layers, including the people, and asking what happens if each is gone for a day, a week and for good; six ways a vendor relationship changes against you without the vendor leaving, and what a free tier does and does not promise; lock-in as a gradient of six kinds, each estimated in person-days; an exit plan and why only a rehearsal on a small slice makes it more than a hope; how thin an abstraction layer should be and what it cannot hide; open-weight models as an exit, with what they remove and what they add; the clauses in a vendor agreement that decide whether you can leave or recover, with a note that this is general information and not legal advice; and concentration of risk, exit triggers and an annual review. Lrnon is the worked example, including a plain list of what the project does not yet have, among them a rehearsed exit. Practitioner is now ten of its eleven modules.

Rolling It Out to a Team published — eight lessons on handing a tool to other people, which is a different problem from building it and often a harder one. Why a working tool is not an adopted one, and finding the real problem by watching the work before proposing anything; piloting with a few willing people on one frequent, low-stakes, checkable task, with success and failure written down in advance; teaching judgement rather than buttons, since interface walk-throughs expire in weeks; a one-page acceptable-use note, and why a flat ban drives use out of sight; treating fears and objections on their merits, promising only what you control, and never monitoring individuals in secret; measuring against a baseline and counting the checking and rework, with a warning about self-reported savings; support, feedback and keeping a rollout current; and reading non-adoption as information rather than blaming the people who did not adopt. Lrnon is the worked example, including that its assistant’s review mechanism existed before anyone used it. Practitioner is now nine of its eleven modules.

Data Protection for Workflows published — eight lessons on what India’s Digital Personal Data Protection Act and the EU’s GDPR ask of whoever builds a workflow that touches personal data. Drawing the path data takes, including the model provider and the logs; what the two laws share and where they differ, with the DPDP Act’s main duties not yet in force on this date; sending less, and what a pattern-based redaction step can and cannot catch; why the provider is a recipient whose terms differ by plan; every second copy of the data and how long each is kept; children, who under the DPDP Act means anyone under eighteen; answering a request to see or erase what you hold; and what to do in the first hours of a breach. The module states plainly that it is general information and not legal advice, and that its sources were reproductions of the statutes and law-firm summaries rather than the official texts, which should be read before any figure is relied on. It also records a correction to Lrnon itself: the assistant’s footer promised that nothing typed was used to train anything, which had not been checked against the provider’s terms for the plan in use, and now says only what holds on any plan. Practitioner is now eight of its eleven modules.

Corrected how the entries for P2, P3, P4, P5 and P6 describe the size of the Practitioner track. Each said Practitioner was a number of modules of twelve. Practitioner has eleven modules, P1 to P11; the twelfth in that figure is the first Builder module, B1, which belongs to a different track. The correct reading of the earlier entries is a number of eleven Practitioner modules. Those entries are left as they were published, wrong count included, because a correction belongs in a new dated entry and not in a quiet edit to an old one.

Cost, Monitoring and Reliability published — eight lessons on running a model-backed feature in production without surprises. Estimating cost from tokens per call, calls per day and price per token, for a typical day and a worst one, and measuring to find which of several limits actually binds; the three levers, which are fewer calls, less sent per call and a smaller model; why a cap that checks, acts and then records is a race, and how reserving the worst case first and settling afterwards makes one hold; recording the shape of every call with the reason beside every failure, over a window longer than the gap between looks; rate limits that are set on you and how to find them by measuring; four ways a provider changes under you with no change on your side; deciding between failing open, failing closed and degrading, and writing a message that is true; and a short list of signals, a weekly look and a runbook. Lrnon is the worked example throughout, including an honest account of which of its own signals are in good shape and which are empty. Practitioner is now seven of its eleven modules.

Multi-Step Workflows and Agents published — eight lessons on systems where the model chooses the next step. The one question that separates a workflow from an agent, and why most things sold as agents are workflows with a single real judgement in them; the loop mechanically, why every turn re-reads the whole history so cost climbs roughly with the square of the turns, and reserving a budget for the whole run before it starts; the tool list as the whole risk, judging each tool by its worst single call; why reliability multiplies down a chain, so ten steps at ninety-five percent each give a run that is right about six times in ten; what an agent actually remembers, and why hostile text read early in a run stays in play; defining done as something code can check, and recognising a stuck run; why an agent’s report of its own work is not evidence of it; and four questions for deciding where an agent is the right answer at all. Lrnon’s assistant is the worked example, as a deliberate non-agent: one retrieval, one model call, no tools the model can reach. Practitioner is now six modules of twelve.

Connecting Tools published — eight lessons on automations that act without anyone talking to them. Every automation as a trigger, steps and an action, with the model as one step among them and the trigger deciding whose text enters; building the chain with fixed values first so the plumbing and the model cannot be blamed for each other; what connecting an account hands over, including scopes, whose account it is and where the keys are kept; treating what a model returns as input to the next step, and never letting it choose the target; the three ways an automation fails quietly, running a step twice, and rate limits; where a human approval belongs and why it decays into a reflex; a ladder from sample data through shadow mode to a staged rollout; and the inventory, one-page description and clean retirement that stop an automation being forgotten. Lrnon is the worked example throughout, including a place where the rollout console describes a safeguard that the database does not enforce, which the module states plainly. Practitioner is now five modules of twelve.

Custom Assistants published — eight lessons on building a saved configuration other people will use. What the feature actually is, which is instructions, optional knowledge and permissions rather than a trained model, and why that misunderstanding leads people to expect improvement with use and to pay for fine-tuning they did not need; writing instructions for someone busy and unfamiliar who gets no second turn, including what to do when a request is unclear; attaching knowledge, and the characteristic failure, which is age rather than error; why a shared assistant holds no secrets, since its instructions can be recited and its files quoted; ownership and a change record, because most platforms keep no history of what the instructions said before; a dozen test cases written before handover, with the author named as the weakest tester; what transfers when you move platform and what does not; and when a document, template or script is the better answer. Sky is the worked example throughout, including an honest account of why its own usefulness is not yet established and a geographic refusal that went unnoticed for five weeks. Practitioner is now four modules of twelve.

Untrusted Input and Prompt Injection published — eight lessons on building workflows that read text somebody else wrote. Mapping every surface where text enters and who controls the bytes; what prompt injection is and why it is a structural property rather than an unpatched bug, since escaping needs a grammar natural language does not have; indirect injection and the carriers that deliver it, including text that survives extraction while being invisible to a reviewer; ranking impact by capability rather than by how clever an attack looks, and the three ingredients whose combination turns a wrong answer into a breach; what fencing and system prompts genuinely buy, which is a lower rate and not a closed class; least privilege as the control that actually holds because it is enforced outside the model; building an attack corpus organised by shape and reading a passing suite without overclaiming; and designing so that a successful injection stays recoverable. Lrnon’s own assistant is the worked example throughout, including an honest account of its exposure once outside contributions reach the retrieval index, and the occasion when its kill switch wrote a log row and changed nothing while reporting success. Practitioner is now three modules of twelve.

Your Own Documents and Sources published — eight lessons on grounding an answer in material you control: why supplying the passage beats relying on training, cutting documents so each piece still means something alone, keyword against meaning-based retrieval and why rare words carry the weight, building a twenty-question set that catches retrieval failures before users do, assembling a request that keeps your instructions and the retrieved text in separate channels, keeping an index current when documents are withdrawn, refusing when the passages fall short, and deciding what belongs in an index at all. Several lessons use Lrnon's own assistant as the worked example, including the measured case where "What is the capital of France?" scored higher than any genuine question because a lesson contains that phrase as a word-prediction example. Practitioner is now two modules of twelve.

Corrected how the site describes certification, which it had been promising in more places than it could deliver. There are two things, and they were being said as one: a Module Record, earned by finishing every lesson, which is issued today; and a Lrnon Certificate, which additionally requires passing that module's assessment, and which nothing on the site can issue yet. The landing page said 'free certificates', E7 said it was 'compulsory for certification', and the achievements page listed the two together. All of that now names the tier and says what it takes. Two words are retired: 'compulsory', because nothing on Lrnon is, and 'certified' as a verb, because it implies an authority we are not. The entry below of 26 August is left as it was published, wrong claim included — a correction belongs in a new dated entry, not in a quiet edit to an old one. Anyone who finished E7 under the old wording will be told directly when the assessment opens, and will not be charged to sit it.

Every lesson that names a tool now shows where the description came from and when it was last checked, with a warning when that is over 120 days old. A new About page states the position on brands and sources: no sponsorship ever, neutral rather than silent, and the vendor is always the authority over us.

Market review of the whole Explorer track. Added E2·L11 on open-weight and regional assistants — DeepSeek, Llama, Mistral and Qwen now have registry cards, closing a gap where E10 and E11 both referenced a tier the landscape module never covered. Added E13·L9 on whether AI use counts as cheating, which research showed is the question learners actually ask.

Health, Money and Legal published, completing the Explorer track at 14 modules and 112 lessons. Eight lessons on preparing for appointments, why symptom questions spiral, financial concepts versus financial advice, understanding legal documents, the four hard stops where AI is the wrong tool entirely, getting to a real person, and six rules for high-stakes questions.

Using AI to Actually Learn published — eight lessons on the illusion of understanding, getting explanations pitched at you, retrieval practice, spacing and forgetting, teaching it back, where AI is a bad teacher, what explanation cannot teach at all, and a study loop that survives a bad week.

AI You Didn't Choose published — eight lessons on AI embedded in tools you already had: spotting it, turning it off, the AI in your phone, features that read your messages, decisions made about you and your rights, workplace tooling, and a one-hour settings audit.

What It Actually Costs published — eight lessons on tokens as the unit you pay in, why bills move, right-sizing the model, the time budget nobody counts, getting real work done for nothing, paying from outside the dollar economy, spending caps before automation, and judging honestly whether it was worth it.

Working in Your Own Language published — eight lessons on why quality varies by language, testing your own, back-translation and what it cannot catch, code-switching, names and scripts, honorifics and register, learning a language, and running a two-language working life.

AI as an Assistive Tool published — eight lessons on reading, writing, seeing, hearing, focus and communication when those are barriers, plus how to make things other people can actually get at. Written after an audit found the curriculum had no coverage of assistive use at all.

Advanced Prompting published — the first Practitioner module. Eight lessons on few-shot examples, structured output, splitting a prompt into single-purpose steps, chaining without compounding errors, persistent instructions, long inputs, testing a prompt against a saved evaluation set, and building prompts that survive a model update.

AI Ethics & Responsible Use published — eight lessons on authorship, disclosing AI use, where bias comes from and lands, who answers when something goes wrong, the environmental and labour costs, automating work that belongs to colleagues, how to refuse, and writing your own code of conduct. This completes the Explorer track: 64 lessons across eight modules.

Truth, Verification & Critical Thinking published — eight lessons on why fluency defeats scepticism, triaging which claims to check, a two-minute verification method, judging sources, handling disagreement, how manipulation finds you, your own bias, and a habit small enough to survive. This module is compulsory for certification.

Creative AI: Image, Video & Voice published — eight lessons on how diffusion models actually work, describing an image, iterating with seeds, the artist-style question, the current limits of video, voice cloning and consent, labelling what you made, and the categories that are not grey areas.

Documents, Spreadsheets & Presentations published — eight lessons on outlining before prose, drafting in your own voice, using an assistant as a critic, spreadsheet formulas you can test, recomputing every number, assertion-headline slides, rewriting for a different reader, and the final consistency pass.

Everyday Productivity published — eight lessons on where AI actually fits in your day, email, reading long documents, research that cites its sources, meeting notes, planning, a twenty-minute routine, and what not to hand over.

Reading preferences added: text size, line spacing, line length, contrast, a reading typeface and data saver. Dark mode also now follows your device — it had been built but never switched on.

Answer options are now shuffled. 94% of correct answers had been sitting in the middle slot, so always picking the middle passed every module quiz without reading the question.

Eight lessons on briefing, iteration, correcting a wrong answer, and the requests no wording can rescue. Adds a 22-question quiz bank and 24 review cards.

NotebookLM was renamed Gemini Notebook by Google in July. The lesson title and the tool card now say so, and the documentation link it pointed at had been dead for six weeks.

Claude's tool card claimed no voice support. Voice mode has shipped on every plan, so the capability grid was showing learners a wrong fact.