Changelog

New models, new benchmarks, fixed prediction bugs, named failure modes, library version bumps. Reverse chronological. Subscribe via RSS.

  1. v0.9.0 Phase 9 — sharing and retention

    site

    The site now produces shareable links that convert.

    • OG cards. Every surface (home, calculate, compare, validate, learn) generates a 1200×630 PNG OG card at build time using satori + resvg-js. Slack, Discord, Twitter, and Notion unfurls render with the site’s visual language — warm paper, slate accent, JetBrains Mono numerals.
    • Page titles. Browser tab titles encode state. /calculate becomes “Llama-3-70B FP8 · 2× H100 · $1.08/M — Tokenomy”; tabs are now navigable.
    • Beat anchors. Every /learn beat heading copies its full URL to clipboard on click. Bloggers writing about decode bandwidth can link to the exact paragraph.
    • First-time tips. One inline tip per surface, dismissible, persisted to localStorage. Never modal.
    • Changelog + RSS. This page. Plus /changelog.xml for readers who follow new model coverage and named failure modes.
  2. v0.8.0 Phase 8 — five-stage learn spine

    site

    The Learn spine is complete. Stages 2 through 5 land alongside Stage 1.

    • Stage 2 — Decode bottleneck. 7 beats, 6 widgets including the roofline plot and four-knob explorer. Derives the decode equation from a spec sheet.
    • Stage 3 — KV cache and capacity. 7 beats, 5 widgets. Names the superlinear capacity cliff at long context and the GQA / MLA / paged-attention answer stack.
    • Stage 4 — Optimisation toolbox. Resolved §4.8: hub + 3 sub-pages (memory levers, compute levers, system levers) over the originally proposed 5 sub-pages.
    • Stage 5 — Designing a deployment. 10 beats walking a real product spec to a defensible scorecard. Closes the spine.
    • Glossary. 21 entries seeded across all stages.
    • Falsify-a-claim callouts. 11 across the spine, each with a /calculate deep-link for hands-on verification.
  3. v0.7.0 Phases 5–7 — Compare, Stage 1, and Validate

    site

    Three big surfaces land in one push.

    • /compare. Three view modes (Table · Cards · Pareto) plus a sweep generator with 7 curated dimensions. URL forms accept either arbitrary scenarios (?a=&b=&c=) or sweep descriptors (?base=&sweep=&values=). Differences-summary pills above the content. Min 2 / max 6 (Table+Cards) / max 12 (Pareto).
    • /learn/token-economy. Stage 1 of the spine. 6 beats, 5 widgets, ~2,500 words. Derives the input-output price asymmetry from first principles.
    • /validate. The trust play. Coverage grid, sortable benchmark table with 7 named failure modes, methodology / reproduce / tools / feedback sections. Confidence badges across the site link here.
  4. v0.4.0 Phases 1–4 — monorepo, design system, calculator

    site

    The foundation goes live.

    • Monorepo scaffold. lib/ (Python math), api/ (FastAPI), ui/ (Astro + React). Strict layered dependencies; each workspace independently buildable and tested.
    • Design system. Three-tier React component library (primitives / molecules / organisms) with strict downward dependencies. Inter + JetBrains Mono via @fontsource. Light + dark themes.
    • /. Hero, live demo strip, three-path cards, featured scenario. Less than 100 ms calc-to-paint on the demo strip.
    • /calculate. Cockpit layout. EssentialsBar + TuningSliders + 2×2 output cards. URL state with stable param names.

    Library version: 1.0.0. 111 library tests + 5 API smoke tests passing; 4/4 high-confidence benchmarks within ±30%.