Posts

Guides, essays, and deep dives.

43 posts
  1. Light 1-bit sim screen: a conveyor of tiny platformer rooms stamped accepted or rejected, weight sliders, and an inspector with red rubric bars.Developer toolsInteractiveHow can a game keep only the good generated levels?A generator makes platformer rooms nonstop, a simulated System One model scores each on four rubrics, and weights in code decide which ones you play.
  2. Light 1-bit sim screen: a dark dungeon room with the hound's red two-button question, a calibration bar chart on a dashed diagonal, an inspector.Developer toolsInteractiveWhat changes when a game companion acts only when confident?A dungeon hound acts alone above a confidence threshold you set and asks you below it, with a live chart that checks whether its confidence is calibrated.
  3. Light 1-bit sim screen: a guard at a torchlit gate saying a caught-you line, a red suspicion bar, and an inspector of intent bars and gauges.Developer toolsInteractiveHow can an NPC judge what you type without an LLM?Talk your way past a gate guard with 40 scripted lines. A simulated System One model judges every line you type, and code decides what the guard says.
  4. Light 1-bit sim screen: grazers, scouts and a black predator in an ink meadow at dusk, beside a personality editor and red probability bars.Developer toolsInteractiveCan a game creature's AI be a paragraph of English?Eight meadow creatures whose brains are a paragraph of text. Edit a sentence and a simulated System One model changes what each one does next.
  5. Light 1-bit sim screen: an arena with a caster, three dummies, a wall and a red fire-trap ring, beside an inspector of red and black probability bars.Developer toolsInteractiveHow can a game build spells from typed text without an LLM?Type a spell and a simulated System One model answers seven typed questions. Game code builds it from designer-made parts. No text is ever generated.
  6. Light 1-bit sim screen: a red player among black enemies in a grid arena, a beat timeline over a red tension line, and an inspector of red dials.Developer toolsInteractiveHow does an AI game director choose the next encounter?Play an arena run while a simulated System One model scores your tension twice a second and a director samples the next of 14 designer-written beats.
  7. Light 1-bit desktop: a pull request card with four yes/no checks and margin bars, beside a five-line policy with the block line highlighted in red.Developer toolsInteractiveTypeSafe after Claude Code: verify, then escalateFour typed checks on each change a coding agent makes, then code decides auto-merge, confirm, review, or block. Interactive demo with adjustable floors.
  8. Light 1-bit desktop: a request card and four probability bars on the left, a red-highlighted policy line and route count bars on the right.Developer toolsInteractiveTypeSafe before Claude Code: a typed preflight gateA preflight gate asks TypeSafe's Jev four typed questions about each dev request, then local code decides whether Claude Code runs at all. Interactive demo.
  9. Light 1-bit desktop with two racing lanes of worker rows, a red-titled System One window marked done, and a scoreboard of black and red bars.Developer toolsInteractiveTypeSafe instead of Claude Code: judgment, not generationThe same 80 messages triaged by an LLM agent loop and by a System One model returning typed answers. Compare wall time, cost, and malformed outputs live.
  10. Dark flow diagram showing a push to main feeding Workers Builds, which produces a generated wrangler.json and deploys a Worker beside a static assets store, with the deploy command printed in monospace.Developer toolsDeploying Astro to Cloudflare Workers, not PagesHow to deploy an Astro site to Cloudflare Workers with the v14 adapter: the generated wrangler config, asset-first routing, redirects, and the traps we hit.
  11. Dark illustration of a browser window showing a wireframed web page with a red comment pin and a note card reading move this above the fold, beside a crossed-out email icon.Developer toolsClient feedback on staging sites without screenshot emailsReplace screenshot emails with on-page comments on your staging site. How co-mment and Annotate.js compare, and a review round that actually closes.
  12. Four numbered layers stacked top to bottom: content as JSON, Astro 7, Tailwind 4, and Cloudflare Workers, each with a one-line note and a code sample.Developer toolsHow this site is built: Astro 7, Tailwind 4, and WorkersThe stack behind protrailblazer.com: Astro 7 on Cloudflare Workers, Tailwind 4 tokens, content as JSON blocks, Polar checkout, and Claude Code doing the typing.
  13. A red model block feeding glowing token chunks down two dark wires, one painting text word by word, one held behind a buffer gate.Machine learningInteractiveWhy ChatGPT types one word at a time: streaming visualizedWatch a model write tokens one at a time while streaming rushes each one to your screen and buffered mode holds the whole answer behind a gate.
  14. Six red tile fields of increasing area on a near-black canvas, each labeled 1B through 400B, with a log-scale memory ruler below.Machine learningInteractiveWhat does 7B mean? LLM parameter counts explainedThe B in 7B counts the learned weights stored in a model. See how count, precision, and architecture determine file size in this interactive scale tunnel.
  15. A grid of word tokens in red on near-black, glowing arcs fanning from one highlighted query token out to every other token, arc thickness varying by weight.Machine learningInteractiveSelf-attention: how a token decides which tokens to readSee how self-attention works. Tap any word to make it the query and watch scaled dot-product attention weight every other word in the sentence.
  16. Red contour map of a policy landscape on a dark canvas, a glowing dot trailing from a white reference ring toward a dashed target circle.Machine learningInteractiveRLHF and DPO: how models learn what people preferLabel response pairs and watch two preference-optimization pipelines, RLHF with a reward model and DPO's direct update, reshape a conceptual policy landscape.
  17. Two vertical ranked lists side by side on a dark canvas, red rank-change arrows bridging them as documents shuffle positions during a reranking pass.Machine learningInteractiveReranking: why top-K retrieval isn't good enoughA fast retriever pulls top-N candidates, then a cross-encoder style reranker reads query and document together and reorders them. Traps fall, answers rise.
  18. Two red pipeline lanes on a dark canvas, one fed by glowing document chips and one by a frozen timeline, with paired bars between them.Machine learningInteractiveRAG vs fine-tuning: which one your project needsRun the same question through RAG and a fine-tuned model side by side. Push the source date past the cutoff and watch freshness, citations, and style trade off.
  19. Dark pipeline diagram with a red shard of injected text flowing through trust-boundary checkpoints toward an agent terminal on a near-black canvas.Machine learningInteractivePrompt injection: when data becomes instructionsWatch untrusted content compete with the real task inside an agent pipeline, then test which layered defenses actually hold.
  20. A dark horizontal ribbon of token cells with one glowing red needle cell at its middle and a red U-shaped recall curve dipping below it.Machine learningInteractiveLost in the middle: where context windows go blindDrag a needle fact through a long context and run seeded trials to watch recall dip in the middle, plus the strategies that flatten the U.
  21. A large frozen gray weight grid with a translucent red low-rank sheet floating above it, wired to two thin red adapter bars on a dark canvas.Machine learningInteractiveLoRA fine-tuning: adapting a model without rewriting itSee how LoRA trains a task update in two thin matrices while the base model stays frozen, and why rank sets a hard floor on what the adapter can learn.
  22. A dark horizontal GPU memory bar, red weight and KV cache segments spilling past a bright capacity line into a tray below.Machine learningInteractiveCan my computer run an LLM? VRAM requirements visualizedA live VRAM calculator: watch model weights, a growing KV cache, and runtime overhead fill a GPU memory bar, spill into system RAM, and compare GPUs.
  23. Red weight histogram snapping onto a sparse comb of quantization levels, with a partly filled red bit tower on a dark canvas.Machine learningInteractiveLLM quantization explained: fewer bits, less memorySee how quantization shrinks a model's weight memory by storing FP16 values on a coarser grid of levels, and where the error actually lands.
  24. Rows of red token bars and a scatter of dots split by a dashed threshold line beneath a red gauge needle just past its cutoff.Machine learningInteractiveHow AI detectors work, and why they are wrong so oftenScore text with the signals AI detectors use: predictability, burstiness, repetition, and style. Formulaic humans get flagged, edited model text slips by.
  25. Isometric stack of four dark graph layers, sparse on top and dense at the base, a red search path descending toward a glowing query point.Machine learningInteractiveHNSW vector search: how a layered graph finds neighbors fastA layered graph walks coarse to fine toward your query, checking a fraction of the vectors a full scan would. Tune efSearch and M and watch recall respond.
  26. Two stacked panes of horizontal GPU lanes filled with glowing red token cells, the top pane broken by pale striped idle gaps, the bottom packed solid.Machine learningInteractiveContinuous batching: why your API latency variesA GPU scheduler timeline runs static and continuous batching on the same requests. Watch idle lanes refill, queues drain, and latency percentiles move live.
  27. Red probability bars on a dark console, a bright scanline collapsing masked bars to dashed outlines beside a glowing JSON fragment.Machine learningInteractiveConstrained decoding: forcing an LLM to emit valid JSONWatch a schema grammar mask illegal tokens, renormalize the survivors, and force a simulated LLM to emit valid JSON while a prompt-only run fails beside it.
  28. A dark triptych in red on near-black: a directed node graph, a ranked stack of glowing feed bars, and a bright attention heatmap grid.Machine learningInteractiveHow attention works: search, feeds, and LLMsSee how attention works in search, social feeds, and LLM self-attention. One interactive runs the same score, normalize, allocate loop through all three.
  29. Two stacked horizontal token streams on a dark canvas, a short 2k lane shedding red tokens off its left edge while a longer 32k lane below keeps them.Machine learningInteractiveContext window: what an LLM sees, and what it forgetsWatch two LLM context windows, 2k and 32k tokens, slide through the same text and drop exactly the tokens that scroll off the back of the smaller one.
  30. Top-down 3D view of six AI assist nodes labeled with web project tasks, three rendered as red wireframe octahedrons and three as glowing white orbital atoms, set on a dark red grid plane.Machine learningInteractive36 ways AI works alongside a freelance web developerA 3D tour of six freelance web project phases with six AI assists each, plus a toggle showing which 16 of 36 handoffs can drop the human approver.
  31. A 3D scatter of faint red document points in dark space with a brighter central query node and short red connecting lines to its nearest neighbors, against deep black negative space.Machine learningInteractiveRetrieval-augmented generation: how LLMs look things upA 3D demo of a RAG pipeline: a query vector searches a corpus, the top-K nearest chunks light up, and the LLM grounds its answer in the retrieved text.
  32. Dark glass-panel prompt engineering interface, red technique tabs on left, streaming model output on right, red probability histogram strip at the bottom.Machine learningInteractivePrompt engineering: how technique shapes what models sayZero-shot, few-shot, chain-of-thought, and role prompting explained with an interactive demo. See how technique shifts token probability and output quality.
  33. Dark isometric pixel grid with a central red MCP host cube surrounded by six smaller red server cubes, glowing message cubes arcing between them.Developer toolsInteractiveModel Context Protocol: how AI apps plug into tools and dataWatch JSON-RPC messages flow between an MCP host and its servers as the client discovers tools, invokes them, and streams results back to the model.
  34. A 3D cloud of red particles radiating outward from a glowing white point in dark space, with three faint nested wireframe spheres marking truth, plausible, and hallucinated zones.Machine learningInteractiveLLM hallucination: why confident answers drift from truthA 3D demo of language-model sampling: tokens drift outward as text moves from grounded to fabricated, while the model's confidence barely moves.
  35. A dark radial diagram with a central red orchestrator node surrounded by a ring of glowing specialist agent panels, with faint curved trails showing tasks routed between them.Machine learningInteractiveAgentic workflows: how AI agents chain business tasksA radial simulation of an agentic AI workflow, plus a plain-language look at what agents do, why the orchestrator matters, and what breaks first for SMBs.
  36. A dark data-viz scene showing a red histogram of token probabilities along the bottom with the tallest bars labeled in monospace, a bright red vertical cutoff line slicing the visible bars, and a haze of glowing red dots accumulating above the tallest columns.Machine learningInteractiveTop-k and top-p sampling: how an LLM picks its next tokenA plain-English look at the two cutoff strategies that decide which tokens a language model is allowed to sample from, with a live interactive demo.
  37. A dark 3D scene showing six stacked layers of red cylindrical towers arranged in rings, with a glowing red token particle rising through the center and bright beams connecting it to two illuminated towers per layer.Machine learningInteractiveMixture of Experts: sparse routing for huge models at fast-model costA plain-English explainer on MoE models with a live 3D demo showing how tokens route through a small subset of experts at each layer.
  38. A dark side-by-side data visualization with two red bar charts: the left labeled 'logits unchanged' with bars rising above and below a zero line, the right labeled 'prob at T=1.00' with bars rising from a baseline, monospace token labels below each chart.Machine learningInteractiveLLM temperature: one number that reshapes the next-token distributionA plain-English look at temperature in language model inference, with a live demo that shows exactly what the parameter does to the probability distribution.
  39. A dark flow diagram with two red candidate panels feeding into a central judge node, with a single verdict arrow exiting to the right against near-black space and faint red guide lines.Machine learningInteractiveLLM as judge: one model grading anotherOne LLM grades another: a live demo of pairwise, rubric, and best-of-N judging, plus the position, length, and self-preference biases you have to probe for.
  40. A dark viewport filled with a red 8 by 8 grid of independent reasoning-chain cells, most cells lit bright red in agreement on a majority answer, a handful at lower opacity as outliers, with a tally bar above the grid.Machine learningInteractiveInference-time compute: more samples, smarter answersA plain-English look at inference-time compute, with a live grid of reasoning chains that vote on the same question and surface the real accuracy-vs-N curve.
  41. A dark 3D scene showing a horizontal sequence of red token blocks with monospace labels above each, vertical KV cache stacks below each token, and a glowing red write head at the end with bright attention beams fanning back to every prior token.Machine learningInteractiveAutoregressive generation: how an LLM writes one token at a timeA plain-English look at how LLMs actually produce text, with a 3D demo of the KV cache that makes inference feasible and what would happen without it.
  42. 3D loss landscape shown as a deep red saddle-shaped surface, with a red ball near the bottom of the valley and a dashed path showing the optimizer's descent from the top.Machine learningInteractiveGradient descent: how a model rolls downhillWatch a simulated model find a low spot on a 3D loss surface. A plain-English look at the optimizer that drives modern machine learning.
  43. Embedding Space Explorer showing the sentence 'This is Pro Trailblazer' tokenized into red spheres in 3D space, with an embedding vector panel and nearest-neighbors list.Machine learningInteractiveEmbeddings in 3D: how models turn words into coordinatesType a sentence, watch each token land as a point in 3D space. Click any token to inspect its vector and see which other tokens it's nearest to.

Open the full glossary