changelog · the project, dated
How this site grows.
Pacific Design is an ongoing project; this page is its commit log at human resolution — what shipped, when, newest first.
- 2026-08-29
The scroll stop, corrected
A review of yesterday's scroll stop found three things worth fixing. Keyboard focus now outranks snapping — tabbing to a link just below an instrument used to scroll it barely into view and then get yanked back, leaving the focused link off-screen. The stop is now desktop-only, because an open WebKit bug disables momentum scrolling on touch devices whenever snapping is active, which made every flick through the writing stop dead; phones and tablets keep the full-size visualization and scroll normally. And the canvas was only re-syncing its drawing buffer when its width changed, which was fine when heights were fixed numbers and wrong the moment they started following the viewport.
- 2026-08-29
Reachable by keyboard
Two fixes for people who do not use a mouse. Every page now opens with a skip link, because tabbing past seventeen navigation items to reach the article is a tax this site had been charging on every visit. And in-page links — the jump list on the reading paths, every cross-reference to a section — now land below the sticky bar instead of underneath it, which had been quietly hiding the first line of whatever you clicked toward.
- 2026-08-29
The instrument gets the screen
Pages carrying a live visualization now hold a scroll stop: scroll toward the instrument and the page settles on it, sized to the screen instead of to a fixed panel — the whole screen in the Lab, where the instrument is the page, and four-fifths of it on a section hub, so the first entries still show underneath. It is proximity snapping, not mandatory — it catches you when you land nearby and never fights a deliberate scroll into the writing below, which stays free the way long reading should be. Readers who ask for reduced motion keep the plain layout. The favicon was redrawn at the same time: the old one used two-pixel strokes that rendered at half a pixel in a browser tab and vanished.
- 2026-08-29
Two questions the site was silent on
What it costs the grid — the energy, water and carbon question, with the arithmetic actually done and an honest account of why the numbers in circulation disagree by orders of magnitude. And one model, many senses, on the shift from bolting an image encoder onto a text model to training on everything at once: what the shared representation buys, and what it costs in tokens, data and attack surface.
- 2026-08-29
Leak a model on purpose
The Lab's sixteenth instrument, and the first for the new applied section: a real logistic regression on a thousand loan applications, scored twice — once the way your validation set scores it, once against a future it has never seen. Turn on a column that only exists after the outcome and watch validation leap into the nineties while the model gets no better at its actual job. Then turn on a genuinely good feature and watch both numbers rise together, because the reflex this page could easily teach — big jump means leakage — is wrong. Every combination you try gets plotted against the line where the number you saw was the number you got.
- 2026-08-29
How to stop reading and start building
The site's whole purpose is that someone learns something, and until now nothing here said how to get from reading to doing. Learning by building is six small projects in the order that teaches most, the habit that makes them stick, and an honest account of why a year of reading can leave you unable to recognize a failure you can define perfectly. There is also a sixth reading path now, for people whose problems already arrive as tables.
- 2026-08-29
The machine learning nobody blogs about
A fifteenth section: Applied & Classical ML, sitting second in the reading order because it should have been there from the start. Behind the headlines, most deployed machine learning reads rows and columns — and on that data a gradient-boosted tree still routinely beats a neural network. Six entries on why that is, on the feature work and the leakage bug that decide most projects, on recommenders and forecasting and rare-event detection, and on the scientific results that changed what any of this is for. The hub's visualization is real boosting: watch each tree fit what the last one missed.
- 2026-08-29
Break an agent, on purpose
The Lab's fifteenth instrument is an agent loop you can compromise. Plant a sentence in a page the agent reads and watch the taint spread forward, step by step, until it reaches the send button — then turn the two architectural defenses on and watch them work without filtering anything. Beside it, the arithmetic nobody quite believes until they see it: two hundred and forty runs of the same task, and the exponential that decides how many finish. This site keeps saying the defense is architectural; here it is, operable.
- 2026-08-29
Four new entries, and a pass over everything else
New: measuring retrieval separately, learning a policy from decisions already logged, judging generated media, and the open-weights turn. Then a site-wide consistency pass: every entry is now reachable from another entry, the self-checks grew from forty-two questions to seventy, and a dozen places where two articles disagreed with each other were reconciled. A free site earns trust by being checked, including against itself.
- 2026-08-29
Lighter, and usable without a mouse
Article pages stopped loading the visualization engine they never used — about 25KB less on most of the site. Form controls got borders that actually meet the contrast standard, and every instrument that previously needed a click now has a button that does the same thing, so the Lab works from a keyboard. A free site should load fast and work for everyone; those are the same job.
- 2026-08-29
Three more instruments, and the questions behind them
The Lab grew to thirteen: a reliability diagram you can break and repair, the fairness impossibility as a slider you move, and a retrieval pipeline that fails in front of you — the answer is in the document every time, and whether it reaches the model is your chunking decision. Each one pairs with a new written entry: calibration, bias & fairness, and the retrieval ceiling.
- 2026-08-29
Answer before you open
Every section now ends with three self-check questions — the answer hidden until you ask for it. Trying to recall something teaches it better than re-reading does, and this site would rather you learned the material than admired it. Forty-two questions, no score, no account, nothing recorded.
- 2026-08-29
Built for learning, not just reading
Two additions aimed squarely at the reason this site exists. Reading paths: finishable routes through the map — new to the field, building something, working with data, deciding about it, understanding the risks, or learning by turning knobs — because "read in any order" is unhelpful advice on day one. And a glossary of every term the site uses, defined in a sentence and linked to the entry that explains it, so hitting an unfamiliar word costs you a glance instead of a detour. Free, no account, nothing tracked — as always.
- 2026-08-29
Two entries the map was missing
Reasoning models — the second scaling dial, bought with serving compute instead of parameters — and interpretability, the attempt to read the weights rather than trust the outputs. Both were living as passing mentions in other entries; both now have their own.
- 2026-08-29
Wide frame, deeper instruments
The site now fills a 1600px frame, and every visualization was rebuilt to earn the extra room rather than stretch into it: gradient descent races momentum against plain descent, attention grows its causal matrix with the real weights in every cell, the context window shows a second request hitting the prefix cache, the agent loop tracks its own transcript cost, retrieval ranks and reranks with scores, the Q-learner plots its learning curve, the robot arm reports its control signal, and the neural network gained a stats panel and a live confidence histogram.
- 2026-08-29
Ten deeper cuts
The map thickened where the territory had: computer use, where training data comes from, fine-tuning in practice, video understanding, 3D & world models, privacy & memorization, agentic retrieval, local & edge inference, context engineering, and self-driving — ten entries slotted into their home sections, each sent through the same adversarial review as everything else. The Lab grew a tenth instrument: an embedding space that trains on whatever you type.
- 2026-08-29
Reviewed, adversarially
Eight independent review passes went over every entry on the site — the new sections, the originals, the late additions, and the Lab's own write-ups, two reviewers re-running demos' math against their source — and 65 fixes landed: a mislabeled equivariance, a causal mask overstated away, policy datelines that had aged in a month, an invented autonomy tier, and demo copy that promised things its code didn't do. The publication grades itself the way it tells you to grade models.
- 2026-08-29
The site index
Every entry on one filterable page at /suchen/ — matching runs locally in your browser, nothing leaves the page. Plus structured data on every article and a build-time audit that refuses to publish a broken internal link.
- 2026-08-29
The Lab opens
Ten interactive instruments at /lab/: train a real neural network, walk a loss landscape, train a tokenizer, drive an attention head, sample a tiny language model, steer a diffusion sampler's guidance knob, grow an embedding space from your own text, race exact against approximate retrieval, teach a Q-learning agent in a maze you draw, and see your own voice as a spectrogram — all in the browser, nothing installed, nothing recorded.
- 2026-08-29
The roadmap, completed
Eight sections written and shipped in one push — Computer Vision, Speech & Audio, AI Safety & Alignment, Reinforcement Learning, Robotics, AI Infrastructure & MLOps, Policy & Regulation, and History of AI: 40 entries, each section with its own live visualization. Every major section of AI this project committed to is now live; the map moves with the territory from here.
- 2026-08-27
v2: the AI publication
The one-day-old software-design field guide was scrubbed and the site relaunched as an AI publication: six sections, 28 entries, the latent-space design system, and live canvas visualizations on every hub.
- 2026-08-27
v1: first light
DNS repaired, TLS issued, and pacific-design.com served its first real page.