Engineering Notes

The LK Forge Blog

Writeups from building this site: how the game AIs work, what an AI-assisted pipeline gets right and wrong, and real numbers instead of vibes.

Data Studies

Do the Top 1,000 Sites Actually Use robots.txt, Open Graph and llms.txt?

Every SEO tool assumes these files matter — so we scanned the Tranco top 1,000 sites to see who actually ships them. Of the 1,000, 554 served a browsable homepage (the rest are CDN/DNS infrastructure); among those, the basics are near-universal — meta description 82%, robots.txt 82%, Open Graph 73% — but it falls off fast: structured data (JSON-LD) is on 41% and FAQPage schema on just 2%. The 2026 shift is on the AI side: 27% now name at least one AI crawler in robots.txt to keep it out (GPTBot 20%, ClaudeBot 19%, Google-Extended 17%), and 16% already publish an llms.txt — the latter concentrated in developer-tool and infrastructure companies (Cloudflare, GitHub, Stripe, Shopify). Caveat: the top 1,000 skews tech-heavy, so adoption runs ahead of the general web. Reproducible from a public list and a short script.

 ·  Read → Data Studies

What the Average Big Five Result Actually Looks Like — 19,718 People, Measured

Is a 70% in Openness high? Does scoring high on one trait predict the others? We scored 19,718 real IPIP-50 Big Five responses — the same instrument our free test uses — into the five OCEAN traits. The headline: the traits barely correlate (the strongest link between any two is r = 0.33, Extraversion–Agreeableness; the average pair is 0.17), so your score on one says almost nothing about the rest — and not one of the 19,718 people scored in the top band on all five. The distributions differ by trait (Openness and Agreeableness skew high, Extraversion and Neuroticism spread widest), and 46% land in the middle band on three or more traits. One honest caveat: it is a self-selected online sample, so the distribution shapes and near-zero correlations are the robust findings, not the absolute means. Reproducible from a public dataset and a short script.

 ·  Read → Data Studies

Which Video Format Is Actually Smallest? H.264 vs VP9 vs AV1, Measured

"Convert to MP4 to save space" is mostly wrong — MP4 is a container, not a compression method; the codec inside sets the size. We encoded two 720p clips (Big Buck Bunny + Jellyfish) across a full quality sweep in H.264, VP9 and AV1 and scored each with VMAF, Netflix's perceptual metric. At equal quality (VMAF 90) VP9 is ~40% smaller than H.264 and AV1 ~60% smaller — the same 10-second 720p clip runs 3.3 MB as H.264, 2.0 MB as VP9, 1.3 MB as AV1, while the .mp4/.webm/.mkv extension does nothing on its own. AV1 and VP9 just cost far more CPU to encode. Every number is the mean of the two clips, reproducible from a short script.

 ·  Read → Data Studies

How Much Quality Do You Lose Compressing an Image? We Measured It on 24 Photos

"Compress without losing quality" is a promise lossy formats can't fully keep, so we measured the trade. Across the 24-image Kodak reference suite we saved every photo as JPEG and WebP at qualities 20–95 and scored each on SSIM (structural similarity to the original, 0–1). The curve bends hard: the first kilobytes buy most of the quality, then it flattens — pushing JPEG from quality 85 to 95 nearly doubles the file (14% → 26% of the original) for an SSIM gain of just 0.025 (0.95 → 0.98). WebP sits about 30% smaller than JPEG at the same SSIM (29.6% at 0.95, 31.9% at 0.97), and staying lossless costs roughly 7× a quality-85 JPEG. The practical takeaway: the sweet spot for photos is quality 80–85, and the live preview beats any single number. Every figure is the mean over the 24 images, reproducible from a short script.

 ·  Read → Engineering

We Asked ChatGPT and Grok to Benchmark Our Game AI. Then We Ran the Code.

We handed ChatGPT and Grok the same brief — build a tool proving classical game AI beats an LLM at Tic-Tac-Toe and 2048 — then ran what each produced. ChatGPT shipped an honest, runnable tool that refused to invent LLM numbers and wired a real proxied model; Grok shipped a self-contained script whose "LLM" was a simulated straw man (12% random moves plus noise), with "illustrative" results and a DETERMINISTIC ADVANTAGE DEMONSTRATED banner. Running Grok's own code, expectimax reached 2048 in 6 of 8 games (median tile 2048, avg score 27,976) versus 0 of 8 for the noisy heuristic — a ~22x quality gap — yet a full 100-round run would take about 4.5 hours, which is exactly why those numbers were never run. And on Tic-Tac-Toe minimax was ~300x slower than the straw man, so the honest advantage is consistency, not speed.

 ·  Read → Engineering

Three Hard Problems, One Method: Chaining Math Calculators

Real problems are never one answer — they are a chain of them. We take three genuinely multi-step problems and solve each end to end by wiring the LK Forge math calculators together: a ball thrown off a ledge (Polynomial → Quadratic Formula → Derivative → Integral, from launch to a 55 ft total path and a 43.8 ft/s impact), a set of ten quiz scores (Mean/Median/Mode → Variance → Standard Deviation → Z-Score, deciding the top 100 is high but not an outlier at z = 1.86), and a 3×3 linear system (Determinant → Inverse → Multiply-to-verify → Solve → Rank, proving a unique solution of x=2, y=3, z=−1). The method — decompose, solve each piece with the right tool, recombine — across fourteen calculators, plus a dependency-free script that reproduces every number.

 ·  Read → Live Data

204 Asteroids Passed Closer Than the Moon This Year

Using NASA/JPL's close-approach data, we counted a year of near misses: 204 known asteroids passed within one lunar distance of Earth, 99 within half, and 12 crossed inside the geostationary satellite belt. The closest, 2025 UC11, skimmed roughly 228 km above the surface on October 30 — below the orbit of the ISS. All were small, all missed, and JPL was tracking them. A distance histogram, the five closest passes, why 204 is a floor not a ceiling, and a keyless snippet to reproduce the counts.

 ·  Read → Live Data

Climate Vital Signs: The 2026 Snapshot of CO₂, Temperature, Methane and Arctic Ice

Four primary-source climate numbers in one dated card: atmospheric CO₂ at 429 ppm (up 35.9% since 1958), global temperature +1.39 °C above the 1850s, methane 1,938 ppb (up 19% since 1983), and the Arctic summer sea-ice minimum down 32.7% since 1979 to 4.75 million km². Each with the full series behind it, its source agency (NOAA, the Met Office, NSIDC) and baseline, plus a one-liner-per-metric snippet to reproduce every figure from public feeds.

 ·  Read → Live Data

How Far South Did the Aurora Reach This Year? A Year of Kp in One Chart

A full year of official NOAA/GFZ planetary K-index data — 2,921 three-hourly readings — reduced to one chart. Geomagnetic activity peaked at Kp 8.67 (a G4-severe storm) on two separate occasions, and 62 of 366 nights reached storm level. Under NOAA's own scale, a G4 storm can bring the aurora as far south as Alabama and northern California. What Kp is, why it decides whether you see the lights, and a NOAA G-scale table of how far south each storm level reaches. Every number reproducible from the public GFZ feed.

 ·  Read → Engineering

We Built Three Difficulty Levels for a Crazy Eights AI. Then We Measured Them.

A seeded self-play study of our Crazy Eights AI. Heads-up, the Hard opponent beats the Beginner just 50.5% of the time — a coin flip — because only about 30% of turns even offer a choice and the shuffle decides the rest. Skill shows up a little at a four-player table (a 30% win share against a 25% fair share), where going first matters more than playing well. Why a shedding card game barely rewards skill, and why we skipped minimax and MCTS. Every number reproducible from a seed.

 ·  Read → Engineering

Watch a Game AI Think: Minimax and Alpha-Beta, Live

An interactive game-tree explainer: step through minimax on a real Tic-Tac-Toe position, toggle alpha-beta pruning, and watch the node count drop — same move, fewer nodes. Plus runnable pseudocode, a depth-vs-tree-size widget across our five games, a link to a real unminified engine, and how Tic-Tac-Toe, Connect 4, Checkers, Othello and Chess each apply the algorithm.

 ·  Read → Engineering

What Is One Ply of Search Worth? Four Board-Game Engines, Benchmarked

Connect 4, Checkers, Othello and Chess all use depth-limited negamax. We ran 2,400 headless self-play games to measure what each extra ply of search actually buys. The first ply is worth a fortune — up to +953 Elo, because a depth-1 engine is barely better than random — but after that the gains fall off a cliff and stop behaving: small, game-specific, non-monotonic, with a real odd/even parity wobble in Connect 4. The mirror image of our Go MCTS result, where more search scaled up smoothly. Every number reproducible from a seed.

 ·  Read → Engineering

Does Thinking Twice as Long Make a Go AI Twice as Good?

We ran our 9×9 Go engine against weaker copies of itself, 600 headless games, to measure what a bigger MCTS search actually buys. The surprise: no diminishing returns. Each doubling of playouts adds more Elo than the last — +171 at the bottom of the ladder, +357 at the top — with no flattening out to 1,600 playouts a move. Plus a side finding: at this strength, 7.5 komi slightly over-compensates on 9×9 (Black wins 41%). Every number reproducible from a seed.

 ·  Read → Guides

Free Puzzle Games Online: The Ones Worth Bookmarking

A guided tour of the eleven puzzle games we build at LK Forge — Sudoku, 2048, Minesweeper, Connect 4, unbeatable Tic-Tac-Toe, Color Lines, the 15-puzzle and block puzzles, Memory Match, plus a sliding puzzle, maze generator and word-search maker — grouped by the kind of thinking each one trains. All free, in your browser, no sign-up or download. Includes an honest look at the 2019 PROTECT study on puzzles and cognition (association, not proof).

 ·  Read → Guides

JSON Formatter Online: Stop Editing JSON by Hand

A single missing comma in minified JSON can cost you an hour. This guide covers what a JSON formatter online actually does, why formatting and validating are different jobs, when to minify, how to keep sensitive payloads private (check the Network panel yourself), and the four mistakes that quietly waste the most time. Everything runs in your browser — no sign-up, nothing uploaded.

 ·  Read → Data Studies

Friday the 13th, and the Gregorian Calendar's Hidden Patterns

Counted across a full 400-year cycle, the 13th of the month lands on a Friday 688 times — more than any other weekday, but only just (the rarest gets 684). Plus why every year has one to three Friday the 13ths and never zero, how the calendar repeats exactly every 20,871 weeks, and which years get 53. Every figure reproducible with one script.

 ·  Read → Data Studies

The World Clock Time-Zone Landscape: What 162 Places Say About Time Zones

We read the standard UTC offset of all 162 places on our World Clock straight from the IANA database. Eleven sit off the whole hour — Kathmandu at UTC+5:45 is the only :45 offset on the board — 46% never change their clocks for daylight saving, and the whole set spans 22 hours from Honolulu to Fiji. Every figure is reproducible with one script.

 ·  Read → Live Data

What a Weather Forecast Tool Shows That Your Default App Hides

Your phone's default weather app compresses the forecast into one icon and one number. A real forecast tool exposes the values behind it — rain as a probability, the drivers of feels-like, an honest signal that day 7 is a weaker claim than tomorrow, and UV and sun timing as first-class data. Free, in your browser, on credited Open-Meteo data, no account.

 ·  Read → Live Data

How to Track Active Wildfires in Your Browser (Without Fooling Yourself)

A satellite hotspot is not a confirmed wildfire — it can be a gas flare, an industrial source or a volcano. Here is the honest three-step browser workflow: NASA FIRMS detections on our wildfire tracker, official incident status from InciWeb and NIFC, and smoke from AirNow. Free, no account, and never a substitute for emergency services.

 ·  Read → Live Data

Track Earthquakes, Volcanoes and Wildfires in Your Browser

Three live hazard trackers, one workflow: earthquakes from USGS, volcanoes from the Smithsonian GVP, wildfires from NASA FIRMS. Each free, in-browser and account-free — and each honest about the gap between a coloured map marker and what is really happening on the ground.

 ·  Read → Data Studies

London Is 5 Hours Ahead of New York — Except for 28 Days a Year

We measured all 17 city pairs we publish against the IANA time zone database, every day of 2026: 14 of them change their time gap during the year. London is 4 hours ahead of New York, not 5, for 28 days — the three weeks each spring and one week each autumn when the US and UK clocks are out of step. London–Sydney takes three different values.

 ·  Read → Live Data

Which Volcanoes Are Erupting Right Now? (And Why 75% Is the Wrong Number)

The Smithsonian's current-eruptions list carried 40 volcanoes on 16 August 2026, of which roughly 20 erupt on any given day — and "active" does not mean any of them are erupting today. Plus the Ring of Fire statistic almost every article gets wrong: 687 of 1,214 Holocene volcanoes, which is 57%, not the 75% usually quoted.

 ·  Read → Data Studies

We Ran Our Own Blog Through Our Word Counter: 22.1 Words a Sentence

A word counter's least useful output is the word count. We put all 22 posts on this blog — 24,003 words, 1,084 sentences — through our own counter: 22.1 words per sentence against a 15–20 guideline, 39.7 in the worst post, and a top crutch word ("every", 93 uses) that no generic list would have flagged, while the words those lists always name barely appear.

 ·  Read → Live Data

Live Space Data in Your Browser: What to Actually Open

Where the ISS is right now, which asteroids pass this week, and what a patch of sky really contains — from a browser tab, with no account or API key. Plus two counts we pulled live rather than quoted: 6,336 confirmed exoplanets and 36,233 open datasets on 16 August 2026, each reproducible with a single query, and both well ahead of the figures still in circulation.

 ·  Read → Data Studies

Tipping on the Post-Tax Total Costs About 1%: We Ran It for All 50 States

Card terminals compute their suggested tip from the post-tax total; etiquette says use the pre-tax subtotal. Computed from our own 51 state base rates: at a 20% tip the choice is worth $1.02 per $100 at the average rate, $1.45 in California, and nothing at all in the five states with no sales tax — about $64-$90 a year for a twice-weekly diner, not the hundreds it is often claimed to be.

 ·  Read → Guides

The Best Free Online Games You Can Play Without Downloading Anything

Free browser games with no download, no account and nothing uploaded: Sudoku, 2048, Minesweeper, Connect 4, Tic-Tac-Toe, Color Lines, Memory Match and Wordle. Which to open for the time you actually have, why the deduction games are the ones worth your attention, and how the AI opponents work — 0 losses in 1,200 benchmarked tic-tac-toe games, the 2048 tile reached in 69.6% of runs, ~0.3 ms per move on your own device.

 ·  Read → Guides

Free Printable Sudoku Puzzles: How to Print Them and Actually Solve Them

Where to get free, print-ready sudoku with answer keys and no sign-up, which of the four difficulty levels to start on, how to print the grid so it comes out clean (100% scale, test page, save a PDF), and the solving techniques — naked singles, pencil marks, naked pairs, X-Wing — that finish a grid. Plus the honest version of what the research does and does not show.

 ·  Read → Data Studies

US Sales Tax by State: The Full Ranking of All 50 States

California's 7.25% is the highest statewide base sales-tax rate; five states (Alaska, Delaware, Montana, New Hampshire, Oregon) charge 0%. Across all 50 states plus DC the average base rate is 5.11% and the median is 6%. The complete ranked table, plus the base-vs-combined-rate distinction most rankings gloss over.

 ·  Read → Data Studies

A Cup Is Not a Cup: How Much 9 Baking Staples Actually Weigh

A cup is a unit of volume, not weight. Across nine common ingredients the same one-cup volume ranges from 89 g of oats to 237 g of water — a 2.66× spread. A cup of sugar is 65% heavier than a cup of flour, and granulated sugar outweighs powdered sugar by 75%. Every weight is the King Arthur chart the cooking converter ships.

 ·  Read → Live Data

Planet & Sky Live: How Many Earthquakes, Storms & Asteroids Right Now

Three live counts of the natural world, fetched from public feeds: M4.5+ earthquakes in the last 30 days (USGS), active US weather alerts right now (NWS), and near-Earth asteroid close approaches in the next seven days (JPL). Each number updates live and links to the raw feed so you can reproduce it.

 ·  Read → Data Studies

What Is the Average Typing Speed?

A live experiment, currently gathering data: take a standardized 60-second typing test and your anonymized net WPM and accuracy join a public dataset. Add a result and watch the speed distribution grow.

 ·  Read → Engineering

Real Game AI, Not a Chatbot: Why Our Games Use Game-Tree Search, Not LLMs

Every "AI" now means a language model — ours doesn't. Why the games run on minimax, expectimax, and BFS instead of an LLM, point by point, plus our measured numbers: a provably-optimal tic-tac-toe move in ~0.3 ms, on-device, with zero network calls and 0 losses in 1,200 games.

 ·  Read → Data Studies

How Much Does It Cost to Charge an Electric Car? We Ran the Numbers for 10 EVs

Using one transparent model we computed the cost to charge 10 popular EVs at home, on public Level 2, and on DC fast chargers — per full charge and per mile. Home charging runs about 5-6 cents a mile; DC fast can meet or top a 30-mpg gas car at roughly 2.8× the home rate.

 ·  Read → Data Studies

How Accurately Can You Stop a Stopwatch?

A live experiment, currently gathering data: stop a stopwatch at exactly 10.00 seconds and your anonymized attempt joins a public dataset. Add a data point and watch the error distribution grow.

 ·  Read → Guides

Are Online PDF Tools Safe? How No-Upload PDF Tools Work

Many online PDF tools upload your file to a server. How client-side, no-upload PDF tools work, why it matters for contracts and scans, and how to tell whether a site uploads your files.

 ·  Read → Engineering

How We Built a 12-Tool PDF Suite That Runs Entirely in Your Browser

Twelve PDF tools running fully client-side with pdf-lib, pdf.js, jspdf and SheetJS — no WebAssembly, nothing uploaded — plus the limits we chose to ship honestly.

 ·  Read → Guides

PDF Compression, Honestly: What "Compress PDF" Really Does

Downsampling embedded images vs. rasterizing pages, why rasterizing removes selectable text, and why a small or text-only PDF may not shrink — or can grow.

 ·  Read → Guides

PDF to Word: What Actually Transfers (and What Doesn't)

Converting a PDF to Word transfers the text, not the layout. What carries over, what is lost, why scanned PDFs need OCR, and when to use Word's own import.

 ·  Read → Data Studies

WWF vs Scrabble Scoring: Where the Two Games Diverge

We scored all 172,835 ENABLE words in both games. A word can gain up to 11 points playing Words With Friends over Scrabble, but the Scrabble edge never exceeds 3 — and only 8.02% of words score identically in both.

 ·  Read → Data Studies

How Random Is a Random Word Generator?

We ran 20,000 seeded draws against our word generator's default 365-word pool and tested the sampling with a chi-square goodness-of-fit test: chi-square 320.572, 364 degrees of freedom, p = 0.9509 — no word is favored over any other.

 ·  Read → Data Studies

How Long Is Golden Hour? We Measured It for 27 Cities

Golden hour is almost never an hour — we measured its length for 27 cities with the NOAA solar-position algorithm; it runs from ~24 minutes at the equator to about 58 minutes in Berlin in December, and never reaches an hour.

 ·  Read → Data Studies

The ENABLE Lexicon Study: What 172,835 Words Say About Scrabble's Tile Values

We counted every letter across the 172,835-word ENABLE list. Scrabble's tile values don't match how letters actually occur: C outranks three cheaper tiles, H beats B, and 11 letters are mispriced against frequency — with the longest isograms and richest anagram set thrown in.

 ·  Read → Engineering

Six Games, Three Classic Algorithms: Shipping Real Game AI in Vanilla JS

The six games here run on three textbook algorithms — minimax with alpha-beta pruning, expectimax, and breadth-first search. What it takes to ship them correct and frame-fast in vanilla JS, with the measured numbers: 0 losses across 1,200 tic-tac-toe games, 69.6% of 2048 runs reaching the tile.

 ·  Read → Engineering

Proving It's Actually Unbeatable: How We Benchmark a Game AI Before Publishing a Number

Unbeatable is a testable claim, not a slogan. The headless self-play harness that runs before any strength number ships: 1,200 tic-tac-toe games with zero losses — and the honest finding that a depth-2 search still lost 2 of 200.

 ·  Read → Engineering

Fast Enough to Feel Instant: Search Algorithms Inside One Browser Tab

No backend, no worker thread — adversarial search runs between two frames on the player's own device. How alpha-beta pruning (a 93% node cut, 549,945 → 36,528) and adaptive depth keep it under the frame budget at ~0.3 ms per move.

 ·  Read → Engineering

26 Repos in 29 Days With an AI Pipeline: What Actually Broke

Real numbers from a one-month AI-assisted build sprint: 26 repositories, 1,549 commits, 335 pages — the commit data, the structural failures that mattered, the 93% token-cost finding, and the rules that survived.

 ·  Read →

Deep dives that live with their tools

Algorithm writeups sit next to the tools they explain: