TWO NEW TERMS AND A NEW DISTINCTION • AEO (Answer Engine Optimization) — shaping content so a search or answer engine picks it as the answer itself, not just one link among many. You'll also hear it called GEO. • E-E-A-T (Experience, Expertise, Authoritativeness, and Trustworthiness) — Google's framework for judging whether a page deserves trust. Trust is the one the other three can't make up for. • CPU vs. GPU vs. TPU — the planner that handles the one-off pieces, the crew that does repetitive work in parallel, and the machine built for one kind of AI math. AEO and CPU vs. GPU vs. TPU each come with an illustration of their own. That takes the starter dictionary to 69 acronyms and terms, plus 4 distinctions. Entries you've rewritten are left exactly as you wrote them; the new ones are simply added. JUMP STRAIGHT TO A LETTER On iPhone and iPad, the term list is now grouped by first letter, with an index down the edge — as in Contacts — that takes you straight to one. It steps aside while you search. PICTURES THE SIZE OF THE WINDOW On a Mac, clicking an illustration now opens it as large as the window, ready to zoom in, and you can still resize it from there.

THREE NEW DISTINCTIONS Some terms are easiest to understand next to the one they're mistaken for. Three new entries set them side by side, each with an illustration of its own: • Agents vs. Assistants — an agent is proactive and goal-driven, and acts for you; an assistant is reactive, and helps when you ask. • ASI vs. Singularity — ASI is a level of intelligence; the Singularity is a transformative event it might bring about. • MoE vs. Model Router — a Mixture of Experts routes inside one model; a model router chooses between whole models. That takes the starter dictionary to 67 acronyms and terms, plus 3 distinctions. Entries you've rewritten are left exactly as you wrote them; the new ones are simply added. TWO NEW ILLUSTRATIONS Coding Agent and Singularity now come with artwork of their own, shown in the list and on the term's page. BIGGER PICTURES ON BIGGER SCREENS On a Mac, and on an iPad with room to spare, the illustrations in the term list are larger again — about a third bigger than before.

A NEW TERM Scaling laws — the finding that a language model's error falls as a predictable power law as you add parameters, training data, and compute — arrives with an illustration of its own. That takes the starter dictionary to 67 acronyms and terms. Entries you've rewritten are left exactly as you wrote them; the new one is simply added. THREE NEW ILLUSTRATIONS ASR, CV, GPU, and SLM now come with artwork of their own, shown in the list and on the term's page. BIGGER PICTURES ON BIGGER SCREENS On a Mac, and on an iPad with room to spare, the illustrations in the term list are now larger. A ROOMIER MAC WINDOW On a Mac, the app opens in a larger window, and can't be shrunk so far that the term list and the Recent column crowd each other.

A NEW TERM HYSET — hyperedge-based set-level tool retrieval — helps an AI agent choose tools from a large library by scoring whole sets of tools that are used together, rather than each tool in isolation. That takes the starter dictionary to 66 acronyms and terms. Entries you've rewritten are left exactly as you wrote them; the new one is simply added alongside. EIGHT NEW ILLUSTRATIONS API, BERT, Coding Harness, Embedding, FFN, MLX, Multimodal Model, and NLP now come with artwork of their own, shown in the list and on the term's page. TWELVE REDRAWN ILLUSTRATIONS ACP, AGI, AI, CNN, Guardrails, LoRA, MCP, Model Router, NN, Quantization, RL, and RSI have new artwork. PICK UP WHERE YOU LEFT OFF The Ask screen reopens the definition you were reading when you last left the app. COPY PART OF A DEFINITION Select any text in an entry — the definition, the analogy, the uses, or the source — to copy just the part you need. MORE ANSWERS IN SETTINGS The Frequently Asked Questions page now explains how to get back to Browse from the Ask screen, and confirms that your edits and deletions carry over when you set up the app on another device.

SIX NEW TERMS GRPO — group relative policy optimization — is a PPO variant for language models that drops the separate critic and grades each sampled answer against the others for the same prompt. Hallucinations — model output that is fluent but false: invented facts, citations, or details the input never supported. Misalignment — the gap between what an AI system is optimized to do and what its designers or users actually want, including reward hacking and goal drift. PPO — proximal policy optimization — is a reinforcement-learning algorithm that improves a policy in small, clipped steps so it never strays far from the last version. It's the RL stage of RLHF behind InstructGPT and many later chat models. Swarms — a group of agents that split a job and coordinate, often without a single controller, instead of one agent doing every step. Tokens — the pieces of text a language model reads and predicts, usually subwords or bytes rather than whole words. Context length, price, and generation are all counted in them. All come with an illustration of their own. That takes the starter dictionary to 65 acronyms and terms. Entries you've rewritten are left exactly as you wrote them; the new ones are simply added alongside. FOUR NEW ILLUSTRATIONS CoT, DiT, GGUF, and LLM now come with artwork of their own, shown in the list and on the term's page.

FIVE NEW TERMS Hook — user-defined code that a coding agent runs at a lifecycle event. PreToolUse fires before a tool call and can allow, deny, or change it; PostToolUse fires after, and can log or format but not undo. Its source is Anthropic's Claude Code hooks release of July 2025, and it arrives with its own illustration. Skill — a folder of instructions, scripts, and resources, usually a SKILL.md, that a coding agent loads only when a task matches, instead of keeping the whole procedure in its system prompt. Its source is Anthropic's "Introducing Agent Skills" of October 2025, and it has its own illustration. SFT — supervised fine-tuning — is further training of a pretrained model on labeled examples, so a language model follows instructions instead of only continuing text. Its source is the 2022 InstructGPT paper, and it has its own illustration. STT — speech-to-text — turns spoken audio into written text, the first stage of a classic voice assistant. Its source is Bell Labs' 1952 paper on recognising spoken digits. S2S — speech-to-speech — is a model that takes speech in and gives speech back without passing through a separate transcript, keeping pauses and tone. Its source is the 2024 Moshi paper. That takes the starter dictionary to 59 acronyms and terms. Entries you've rewritten are left exactly as you wrote them; the new ones are simply added alongside. NEW ILLUSTRATIONS RLHF, RNN, and Tool Calling now come with artwork of their own, shown in the list and on the term's page. REDRAWN ILLUSTRATION ML has new artwork. SWIPE BACK TO BROWSE Swipe right across a definition on the Ask screen to return to the Browse list. DELETED TERMS STAY DELETED A starter term you delete no longer comes back after an update, or when you set up the app on another device. YOUR EDITS ARE SAFE ON NEW DEVICES Setting up the app on another device no longer risks replacing a term you rewrote, or a picture you added, with the original. ILLUSTRATIONS YOU CAN REACH FROM THE KEYBOARD Open an illustration full screen with Tab and Space. On Apple Vision Pro and iPad it now highlights when you look or point at it. CORRECTIONS Tool Calling has its full term back, so Restore Original and Siri work for it; three sources are now cited in full; and a leftover duplicate of Embedding, listed as EMB, is gone.

TWO NEW TERMS FFN — feed-forward network — is a neural network whose connections run in one direction only, from inputs through its layers to outputs, with no loops back. It's the shape of the classic multilayer perceptron and of the position-wise block inside every transformer layer. Its source is Frank Rosenblatt's 1958 paper on the perceptron. SDD — spec-driven development — is a way of building software in which a structured, versioned specification is the source of truth, and AI coding agents derive their plans, tasks, tests, and code from it instead of improvising from one-off chat prompts. It's the blueprint to vibe coding's text message. Its source is AWS's July 2025 preview of Kiro, a spec-driven agentic IDE, and it arrives with its own illustration. That takes the starter dictionary to 54 acronyms and terms. Entries you've rewritten are left exactly as you wrote them; the two new ones are simply added alongside. NEW ILLUSTRATIONS ASI, Distillation, DL, GPT, MLP, MoE, ReLU, SGD, and VAE now come with artwork of their own, shown in the list and on the term's page. REDRAWN ILLUSTRATIONS ACP, CNN, Context Window, GAN, and LoRA have new artwork. ARTWORK FOR YOUR OWN ENTRIES TOO A term you've rewritten still gets its new illustration, as long as it has no picture yet. A photo you chose yourself is never covered up.

FOUR NEW TERMS ASI — artificial superintelligence — is the stage past AGI: a hypothetical intellect that greatly exceeds the best human performance in virtually every domain, from scientific creativity to social skill. It arrives with its own analogy, the places the idea is actually argued over rather than deployed, and Nick Bostrom's 2014 definition as its source. RSI — recursive self-improvement — is the loop said to get there: a system that improves the methods it uses to improve itself, so that each gain speeds up the next. Its source is I. J. Good's 1965 paper on the "ultraintelligent machine." The Singularity is where that loop is said to lead: the point past which machine intelligence outruns ours and change comes too fast for our old ways of reasoning about human affairs to hold. Its source is Vernor Vinge's 1993 NASA paper, which gave the word its AI meaning. DeepSWE is the benchmark that stopped reusing the answers. Where earlier sets mine public GitHub fixes — so a high score can be recall of a patch the model had already read — DeepSWE writes 113 original, long-horizon tasks from scratch across 91 active open-source repositories and five languages, keeps the reference solutions off the public record, and passes any implementation that meets the spec. That takes the starter dictionary to 52 acronyms and terms. Entries you've rewritten are left exactly as you wrote them; the four new ones are simply added alongside. PICTURES IN THE LIST A term with an illustration — the app's or your own — now shows a small copy of it at the end of its row, so you can spot it before you open it. ACP, Guardrails, MCP, RAG, and RSI get new or updated artwork. WHAT'S NEW, IN THE APP Updating now opens on a page describing what changed, shown once and then out of your way. If you want it back, it's at the top of Settings, with the welcome screen close by. WIDGET The widget shows up to seven lines of a term's definition instead of four, so a longer entry no longer stops mid-thought. STEADIER ICLOUD SYNC On a new device, the starter terms no longer show up twice while iCloud catches up: the copies are merged as soon as they arrive. And a starter entry that comes back from another device with older wording is brought up to date. Entries you've edited are never touched.

TWO NEW TERMS ASI — artificial superintelligence — is the stage past AGI: a hypothetical intellect that greatly exceeds the best human performance in virtually every domain, from scientific creativity to social skill. It arrives with its own analogy, the places the idea is actually argued over rather than deployed, and Nick Bostrom's 2014 definition as its source. DeepSWE is the benchmark that stopped reusing the answers. Where earlier sets mine public GitHub fixes — so a high score can be recall of a patch the model had already read — DeepSWE writes 113 original, long-horizon tasks from scratch across 91 active open-source repositories and five languages, keeps the reference solutions off the public record, and passes any implementation that meets the spec. That takes the starter dictionary to 50 acronyms and terms. Entries you've rewritten are left exactly as you wrote them; the two new ones are simply added alongside. WHAT'S NEW, IN THE APP Updating now opens on a page describing what changed, shown once and then out of your way. If you want it back — or the welcome screen — both are in Settings.

USES Every one of the 48 built-in acronyms and terms now says where you'd actually meet it — the concrete products, systems, and tasks it turns up in — shown just after the analogy. It's a field like any other: yours to rewrite or clear, filled in by Apple Intelligence for terms you look up, and carried along in shared text and in JSON export and import. FREQUENTLY ASKED QUESTIONS A new page in Settings answers the questions the app gets asked. How do I edit a field? What does editing cost me? How do I put the original back? What happens to a bad date in an imported file, and why are some entries dated January 1st? SPANISH Every screen, label, and prompt is now translated into Spanish. The dictionary entries themselves stay in English. FIXED Importing a JSON file that says nothing about illustrations — an export from an earlier version, or a hand-written list — no longer strips the artwork from your entries. Confirming an overwrite on the Create screen no longer does either.

Analogies for acronyms and terms to aid understanding of the concepts. There is also the option to restore their original definitions in case you edited them.