The capability-scaling-strategy supplement now states more clearly an existing recommendation in its Stage 1 strategy: the Consortium should make capability progress via training compute scaling rather than software progress to the greatest extent possible, because software improvements can be used by covert projects while compute cannot.

An added paragraph after the four areas of impact suggests categorizing them by whether they concern an especially important group's decisions (like AGI project and government decision-making) or society more broadly (like public influence via voting), and classifying interventions accordingly as "targeted" or "broad".

The section listing levers for AIs improving epistemics in the AI-for-epistemics supplement is now titled "Example interventions" instead of "Interventions", and the summary's "more" link points to the new heading.

The discussion of epistemic tools possibly being obsoleted by future AIs now links a Forethought appendix on when meaningful acceleration is still achievable; "Personalized leaning systems" is corrected to "learning systems", and "how conducive the culture … are" to "the cultures … are".

All saved drafts of the experimental playbook essay were taken down along with the feature that displayed them, including the main draft's policy recommendations ("Track where all the compute is", "Invest in verification R&D", "AI Training Transparency") and an imported copy of the "Explaining Plan A" essay.

The work-in-progress interactive playbook was removed from the site while it's reworked: readers can no longer browse saved essay versions on their own pages, edit the essay in the browser, or use the floating outline of expandable choice boxes and section tables of contents that came with it. A few related pieces elsewhere went with it.

• in-browser essay editor removed

• version pages and switching removed

• standalone METR graph page removed

Supplement essays no longer show the section navigation rail along the side that was added with the playbook work; they return to their earlier single-column reading layout.

Section IV previously held a placeholder expandable ("[TODO, includes citizens dividend]") sketching job loss as AI reaches human level. It now defines "Expert-Dominating AI" — AI whose advice even top experts would do better to follow — warns that we must prepare for mass unemployment, and adds an "AI Dividend" box whose preview proposes a small tax on datacenter computing paid out to all citizens, like Alaska's oil dividend.

The "pretty heavy political lift" preview previously gave one reason (superintelligence sounds like science fiction); it now lists two — the science-fiction problem plus the need for robust global coordination. Two boxes under "What would sharing a world with wildly superhuman AI be like?" also gained first-draft previews.

• "pretty heavy political lift"

• "Minimal Safety and Property Rights"

• "Universal Basic Galaxies"

The "Why should I expect superhuman AI?" expandable now opens "Because AI capabilities seem to be rising at an exponential rate" instead of "Broadly, because AI capabilities seem to be going up at an exponential rate", and the Wildly Superhuman AI section's "there will be AI not just smarter than any human" became "there'll be".

Before, hovering a box could stack its peeked section under an already-pinned one in the same row. Now a row shows exactly one section: the hovered box's peek temporarily displaces the pinned box's section and gives it back when the pointer leaves. Peeks also open after a short delay and wait a moment before closing, so crossing the gap between boxes doesn't snap them shut.

Editing plan versions previously only worked when running the site locally; it now works on the live site after entering a shared password, which is remembered for a year. Once entered, the version list also shows its Edit, Duplicate, Star, and rename controls.

Rows previously packed in as many boxes as would fit at a minimum width, with special rules forcing five boxes onto one line and splitting six into 3+3; now each box takes half the row and extras wrap, unless a row is explicitly set to fit on one line — a new style alongside "compact" and "multi-open". Sections also open with a quicker, subtler animation, and headings inside opened sections are sized down so they no longer match the essay's own section titles.

With the plus-style outline marks, hovering a box used to flip its + to a −. Now the + grows and darkens on hover and only becomes a − once the box is actually picked into the portfolio, and hovering the mark shows a tooltip reading "Add to policy portfolio" or "Remove from policy portfolio".

The sidebar's headings switched from small uppercase, tinted labels to the essay's serif face in black at larger sizes, box outlines were lightened from mid-gray to a faint border, row spacing was tightened, and its buttons gained the same add/remove tooltips as the in-text marks.

A new page presents a plain-language explainer of the plans: the four dangers of superhuman AI (loss of control, concentration of power, destabilization, mass job loss), Phase 0 incremental policies, Phase 1 where the US and China build a verifiable "break pedal" by declaring compute and pausing training, Phase 2's choice of a long-term plan, and Plan A's four principles. Two new flowcharts map the choices: a Phase 1 path diagram landing on Plan D (Race to ASI), Plan C (Burn the lead), Plan B (Uncoordinated Slowdown), and Coordinated Slowdown, and a Phase 2 choicepoint splitting "Shut it all down" into Plans S+/S and "Verified Slowdown" into Plans A/A*.

The six incremental-policy boxes in the essay's Phase 0 section — Transparency, Enforce (or repeal) export controls, Invest in verification R&D, AI Compute Tracking, Improve government AI capacity, and Limit R&D budgets — were bare labels with nothing behind them; each now reveals the full policy description on hover, taken verbatim from the scenario's Incremental AI Policy Wishlist.

The essay's rows of labelled boxes (Phase 0 policies, Phase 1 paths) got a new interaction: hovering previews a box's content in a card, and clicking now opens that content as an outlined section spanning the row below, with a disclosure chevron that points right while closed and turns down when open. Previously the content only floated over the page while the pointer hovered. Rows size themselves so every label fits in at most two lines, and long rows split into balanced rows (3+3 rather than 5+1).

The playbook interface on the main scenario page swaps its reddish "race" accent for the Plan S accent color throughout: HUD panels and hover swaps, policy appearance, the extrapolated graph, datacenter labels on the globe, and indicators like alignment-researcher count, employment rate, and median income.

The essay page previously stacked two titles — the site header's "Plan A" plus the essay's own "Explaining Plan A (and S and...)" — along with the Plan A author list, which isn't this post's byline. It now shows only the compact site header and nav above the essay's own title.

The cards that pop up when hovering links grew from 420px to 460px wide (and their maximum from 440px to 480px), with slightly adjusted text sizes, so multi-paragraph previews read comfortably instead of wrapping into a tall narrow column.

Selecting a large date range used to fire a burst of queries at once, and the resulting overload often replaced the whole dashboard with a "PostHog not queryable yet" error. Data now loads one panel group at a time — the active site first, then the other sites, then the history graph — and momentarily rejected requests are waited out and retried, so every panel eventually fills in.

The custom date range was previously capped at 366 days; the cap is now high enough that a site's entire launch-to-today lifetime is selectable, with the daily charts covering the whole span.

A query error used to replace every panel with the error notice, even when most data had already arrived. The notice now only replaces the panels when the site has no data at all; otherwise it appears above whatever loaded successfully.

The site measures which link hover-previews (image previews, curated preview cards, and note tooltips) readers linger on, to guide which previews get improved. Previously nothing was counted; now a hover of half a second on either the link or its open preview card registers once, with repeat-visited variants of the same destination counted together. Nothing about how previews look or open has changed.

A new interactive widget at /companies-widget lets readers explore Plan A company finances: a params card of adjustable assumptions beside the figures, with a strip of headline summary values above the charts. The model assumes compute permits are auctioned at 80% efficiency and applies Cournot competition margins from 2030 on; an earlier AI-industry cash-flow panel was dropped before publishing.

The changelog on the About page previously listed each day as a short headline plus a single expandable of minor items; now each day is split under bold "Content changes" and "Design changes" headers, with every individual change as its own expandable whose body ends in a red swoopy-arrow link jumping to the exact spot on the site. Text sizes across the changelog also grew, and on narrow screens the date now sits above its entry instead of squeezing beside it.

• bold Content/Design section headers

• red arrow deep link on each change

• bigger title and 19px entry text

On the analytics dashboard, the site tabs and column sort headers previously only updated the view in place, so cmd/ctrl-clicking or middle-clicking them did nothing. They now open the resulting view in a new browser tab on modified or middle clicks, while a plain click still switches the view in place as before.

Plan A's footnote on colonizing the reachable universe previously said it's physically possible to "send out colonizing probes to the entire reachable universe in a short amount of time (e.g., 6 hours)"; it now says this takes "a surprisingly small amount of time and energy (e.g., only ~6 hours of the sun's energy output)", matching the cited "Eternity in six hours" paper.

In Plan B, the reflection beginning "We have complicated feelings about Plan B" was preceded by a literal italicized ":::box" line instead of appearing inside a styled box; it now displays as the intended boxed aside.

In the "Automated scenario planning and exploration" paragraph, the link to Forethought's writeup was titled "Scenaio planning on tap"; it now correctly reads "Scenario planning on tap".

On the extrapolated capabilities graph, the red loss-of-control region used to misbehave when a plan's danger boundary sat far above the visible chart: the red faded to almost nothing mid-chart, and where the boundary climbed steeply near the top it compressed into a sharp red sliver with a hard diagonal edge. The red now deepens to full intensity at the top of the chart and fades out smoothly as the danger band exits the frame, even along steep boundaries.

In the capability block's risk key, the entry formerly titled "Delayed Actor Risk" ("The black dots: how far a project racing outside the deal has likely gotten by then") is now "Trailing Actor Risk," described as how likely a non-deal actor has reached the frontier at this point, e.g. because the deal broke down.

The "Are Mythos cyber capabilities overhyped?" hover preview now shows a chart built from Epoch data — replacing a static screenshot — illustrating the spike in reported software vulnerabilities (CVEs) its text discusses; that text now calls the spike "large" rather than "significant".

The Congress-hearings passage in "2027: The Writing on the Wall" previously said lawmakers read "the 2016 OpenAI emails" about preventing an AGI dictatorship; the emails are from 2017, and the text now says so.

Several hover previews shown on external links were corrected: the OpenAI-emails preview now credits the email to both Ilya Sutskever and Greg Brockman rather than Sutskever alone, the DeepSeek synthetic-data example drops from 800,000 to 600,000 worked solutions, the intelligence-explosion preview now notes the explosion is "driven by AIs improving AIs", the dataset-growth claim now cites Epoch's dataset-size trend, and a few typos ("feedbackloop", "capabilities climbs") were fixed.

• Email credited to Sutskever + Brockman

• DeepSeek solutions: 800,000 → 600,000

• "feedbackloop" → "feedback loop"

On a fresh page load, the choose-your-path flowchart region used to sit empty for several seconds — long enough that the plan buttons floated up over the "2029: Choose a Path" heading. The diagram now appears immediately with the rest of the page.

In the phone HUD strip's plan fan, each unselected plan line used to carry its own dash pattern (long, medium, dotted, dash-dot); all unselected lines now share one even dash, and the plan you point at or pick is identified by its letter (S/A/B/C/D) floating just above its line, with the label sliding to keep a constant gap from the line and steer clear of the year pill. The fan also animates in faster.

The About page gains a Changelog section listing a short AI-generated summary of each day's changes to the site, and /changelog now redirects there.

Several hover previews of external links were rewritten with concrete figures instead of placeholders: the AI Futures Model preview now cites ~1% of US GDP going to AI datacenters and human text running out around 2028 (replacing "[GDP fraction]" and "[x?]" placeholders), the recursive-self-improvement preview drops its long task-length quote for a short two-line definition, and new previews were added for takeoff speeds, the Superintelligence FAQ, and several Epoch and Astral Codex Ten sources.

• ~1% of GDP now cited for datacenters

• new smooth vs. lumpy takeoff preview

• NSA breach preview adds summary line

Some link previews now include miniature charts illustrating the concept behind the link, rather than being text-only.

• smooth vs. lumpy takeoff curves

• AI investment growth sketch

• sigmoid branches & paradigm timeline

The data tables in the Deal Decline supplement now carry descriptive titles above them, where before they appeared with no label at all.

Links in the Making Our Own Luck chapters and the Plan A public-perspective epilogue previously opened through a google.com redirect wrapper; they now point straight at their destinations.

The marker that flips the timeline to 2028 used to sit at the very top of the "2028: AI on the Ballot" section; it now appears several paragraphs later, so the year changes just before the election-campaign discussion rather than at the section's opening lines.

Hovering anywhere on the capabilities graph now swaps the HUD's lower half to a stacked two-section explainer: a capabilities block with a subtitle, live R&D speedup readout, and the agent-level ladder, above two key lines explaining the graph's shaded fields — Takeover Risk and Delayed Actor Risk. The red risk field keeps its full strength while hovered instead of dimming.

The compact HUD strip that replaces the full heads-up display on phones was substantially expanded in two rounds of changes, and the graph's zoom animation gained an extra step on small screens.

The hover-preview cards for external links now truncate long destination URLs cleanly instead of letting them overflow, and the card layout adapts better so preview text stays visible.

The scenario and its supplements picked up the newest edits: footnote URLs were corrected across the Plan A footnotes and several supplements (covert AI projects, deal decline, Plan A assumptions, space governance), a citation to a Newsweek report on an AI breach of US classified systems was added, and the plan B/C/D/S branch openings regained the timeline date markers that pace their side panels.

• new Newsweek AI-breach citation

• corrected footnote URLs

• branch timeline dates restored

The AI-vs-non-AI × cluster-vs-single-device figure previously carved 2% of the AI fleet off as single-device and 0.5% of non-AI compute as cluster, contradicting the cluster-split table and its own legend. It now uses the table's categories — the whole AI fleet as AI × cluster, the specialty-consumer forecast as AI × single-device, servers as non-AI × cluster, consumer devices as non-AI × single-device — with the legend relabeled to match. Stock bars are also relabeled 2025–2029, the Jan 1 snapshots they actually show, instead of reading one year early; flow bars keep calendar years 2024–2028.

The distribution-of-power block's shares — world and region totals, the global top-20 actors, and each region's remaining tail — are now taken directly from the model spreadsheet's concentration model, and during Plan A's 2029–2030 deal window the displayed date sweeps in the same January/May/September steps as that data. The plan B/C/D metric series also shifted slightly after being re-read from the spreadsheet.

Following the source document's rename, the supplement now lives at /supplements/takeoff-forecast with its title, cover, and supplements-list entry updated; links to the old takeoff-supplement address, including explorer deep links, redirect to the new one.

Five external links now show preview cards on hover, including Wikipedia's singularity and recursive-self-improvement pages and two news reports cited in the scenario.

• Wikipedia: Technological singularity

• Futurism: OpenAI–Musk AGI emails

• Newsweek: AI-breach report

The reference to the Forethought writeup in the automated-scenario-planning paragraph read "Scenaio planning on tap"; it now reads "Scenario planning on tap".

The black trailing-actor risk field is now the same base shape for every plan, slid in time per plan: 1 year left for Plan A, 7 for B, 9 for C and D, and 2 years right for Plan S, where it fades in faintly instead of drawing solid. The field is also lighter overall, drawing only on higher-risk tiles and at roughly half its previous peak density.

While the pointer is over one of the plan choices, the graph now sketches that plan's projected line and fades the other plans' lines, then fades the preview back out when the pointer leaves.

Choosing a plan at a branch point now brings the choices to a consistent position in the viewport, and deselecting a plan waits for scrolling to settle before returning to the main view instead of jumping immediately.

Footnote markers inside expandables previously showed a sequential position rather than the footnote's real number, so numbers could repeat or skip once expandables were opened; each marker now displays its actual footnote number, and its margin note appears whenever the expandable is open.

A link inside an open hover preview can now open its own preview instead of breaking the first one, and preview cards position themselves more accurately beside the link they describe.

Passages using single tildes for "approximately" — e.g. "~$0.7–1.1M per wafer (~$25–40K…)" in the compute material — previously lost the tildes and rendered the span between them as struck-through text; single tildes now display literally everywhere on the site.

The opening no longer just says the supplement lists all main recommendations — it now explains the scenario's stance: roughly the author's median view on timelines and takeoff, but "realistic about empirical facts and optimistic about US government actions." The appendix table rating how plausible and desirable each key policy is was also tweaked.

The transparency-spectrum figure's label was renamed, and the same rename flows through the assumptions, space-governance, and transparency-plan supplements; long labels near the figure's edges now stay fully on the chart instead of running off.

Hovering some external links in the text now pops up a short preview card; previews were added for links to ai-2027.com and Anthropic's recursive self-improvement post.

The box at the start of the Plan B chapter recapping Plan A wasn't rendering correctly — text meant to sit inside it now displays within the box.

In "2029: Choose a Path," the MSS link led to a malformed Wikipedia URL; it now opens the Ministry of State Security article. Footnote 186 previously said colonizing probes could spread to the entire reachable universe "in a short amount of time (e.g., 6 hours)"; it now says this takes "a surprisingly small amount of time and energy (e.g., only ~6 hours of the sun's energy output)".

The choose-your-path branch diagram at the scenario's decision point now lays out five plan branches side by side instead of four, with the plan graphs and arrows updated to match.

The capability graphs' y-axis jump from ~144x to 1000x now happens later: Plan D at 2031 (was 2030) and Plan C at May 2031. Plan B previously never zoomed out at all — its axis stayed pinned at 143x forever — and now expands to 1000x at 2032, a year before its 2033 takeoff. Mobile is unchanged.

The black dithered region marking the too-slow trailing actor wasn't drawing at all on the branch-plan graphs (e.g. Plan D) despite the underlying data being correct; it now fills its region of the visible plot.

Each chart embedded in the economics supplement had been loading the entire interactive explorer — controls and all charts — instead of the one chart it illustrates. Embeds now isolate to their single chart, and the simple-model embed shows just the calculator.

Blockquotes across the site now show a left border only instead of a full box, and a new bordered callout-box style is used for the plan summary boxes in the Plan B–D and S chapters, with matching print styling.

The scenario HUD's forecast numbers were shown with spurious precision — a rough estimate displayed as "3,771" safety researchers or "×1,937" AI speed implied accuracy the model doesn't have. Safety-researcher counts now round to one significant figure, and median income, AI speed multipliers, and welfare-panel researcher counts round to two, while small values like "1.5K" keep their meaningful fractional digit.

• 3,771 researchers → "4K"

• "$10.3M" income → "$10M"

• "×1,937" AI speed → "×1,900"

The "DEAL-RACE BREAKDOWN" chart now shows a new Monte-Carlo snapshot with 14,466 draws instead of 5,767 (built 2026-07-11 rather than 2026-07-09). The percentages shift slightly throughout — e.g. in the 1–2y scaling band without detection, the covert→milestone share moves from 36.3% to 37.5%.

Straight apostrophes and quotation marks across the scenario chapters, About page, insider and public perspectives, supplements, and chart labels now render as proper typographic curly quotes.

• AI 2027's → AI 2027’s

• "explosion" → “explosion”

• don't → don’t

Three changes to the scenario HUD: the below-HUD power breakdown now splits the world into country columns by frontier-AI access, with icons and an underlined key; the capability ladder is marked by a gliding red level-meter segment alone, replacing the black rail line (the details-panel ladder matches); and the graph's capability line, point markers, and hover dot now cut off at the top of the red dither field instead of floating up over the page navigation.

On desktop, the Plan C and D graphs jumped their y-axis maximum to 1000 at the start of 2030 while capability was only around 3, cramming the line into a flat sliver at the bottom of the frame. The axis now ramps gently (6 → 12 → 60 for Plan D, 6 → 12 → 80 for Plan C), so the line sits around 25–40% of the frame through 2029–2031 and only shoots off the top when the plan's intelligence explosion actually hits — 2031 for D, 2032 for C — matching how mobile already behaved.

Pressing Cmd-F on the scenario now expands the selected plan's full storyline plus every collapsed expandable, so the browser's find-in-page can search all the text; it no longer also pops open the plan's epilogue as part of doing so.

While scrolling through Plans B, C, or D, the date pill on the graph used to glide through every intermediate month, even though the timeline itself snaps in thirds of a year. The pill's month now snaps to January, May, and September to match. Plans A and S, which show year-only pills, are unaffected.

The earth's crown previously poked slightly above the top edge of the red takeover-risk dither field; the globe is now seated 5 pixels lower so it stays fully inside the field.

Three clarifications: the per-wafer cost build-up is spelled out more explicitly, controlled hardware R&D is restated as roughly 10%/yr progress (the post-2010 cost-per-transistor pace), and the description of inference-specialized chip architectures is refined.

In the Distribution of Power treemap, Plans C and D now count an actor as frontier only if it is within 2x of the leader (from 2031 on), and under Plan B the merged US National Project appears as the sole frontier cell.

The world-compute history chart's dense-INT8 throughput figure for the RTX 4060 was wrong; it now reads 968 TOPS, with the 4060 Ti at 1,412.

The acknowledgments line and the Model Explanation call-out box both linked to an out-of-date writeup PDF; both links are now gone.

The italic "Plan A — capability-shaped uplifts" label that sat above the chart's left milestone ladder has been dropped; the milestone labels themselves are unchanged.

The verification get-involved section now cites Cankaya 2026 on bit-exact inference.

The dashboard's hover details got a quieter treatment: the separator above each detail is now a thin dotted black line hung in the gap (it was a thick solid red rule), details start where the content they cover starts, and the AI Policies definition was trimmed to one line ("…or have not begun") so the policy map sits higher.

• Separator: 1px dotted black, was 2px red

• Titles and descriptions 13px → 11px

• "Click for details" hint removed

As you scroll a branch plan's story, the dashboard date previously leapt from one dated event straight to the next; for Plans B, C, and D it now advances between them in four-month steps, and the year pill shows the month ("Nov 2030"). Plans A and S, which move on annual phases, keep year-only dates. Also fixes the date occasionally stalling below the story's real markers (Plan D could stick at late 2030).

Plan A's ending reframes the capability graph around the active plan's peak, and this wrongly fired for Plan S too: past 2041 its flat ~1.8 capability squashed the whole y-axis to a 0–1.8 range. Plan S now keeps the standard frame through its 2045 endpoint.

On narrow phones the expanded dashboard's four-across stat row overflowed its column and ran underneath the Distribution of Power diagram; the stats now sit flush with tighter gaps and the power diagram narrows on the smallest screens so both fit side by side.

Tapping the collapsed dashboard strip could misfire: depending on where the finger landed, the just-opened dashboard immediately showed the AI Capabilities details view on top, or the sheet opened and instantly closed. A tap now only opens the dashboard; tapping the graph inside the opened dashboard still works as before.

On mobile, embedded economics-explorer charts could each render the full explorer chrome — title, Full/Simple tabs, controls, and Parameters tab — repeated down the page; embeds now show only the single chart they illustrate, even on slow devices.

Several rounds of wording edits landed across the Plan A scenario — from the 2029 Choose a Path chapter through the epilogue and postscript — plus the footnotes, both perspective pages, and several supplements; the covert AI projects supplement was cut down substantially, losing long stretches of its earlier discussion. Representative examples:

• covert AI projects heavily trimmed

• footnotes revised throughout

• edits in every 2030–2040 chapter

The voiceover was replaced with Nicole's re-edited recording — a cut inside the 2029 chapter shortens the total running time from 3862.5s to 3854.5s — and every chapter timestamp from 2030 onward was realigned so chapter skipping stays in sync.

The AI-2040.pdf download was a 70-page A4 file from before the print redesign; it's now a 90-page US Letter render of Plan A with the other plans as appendices, including the new Foreword section, white figure backgrounds, ink-black headers and page numbers, and smaller figure captions.

The RAND CAST card now carries an expression-of-interest form for working with the team plus a link to the RAND CAST center page, and the Community Night card links its Partiful RSVP page instead of the stale "Link will be posted here."

The chart's horizontal axis previously read "less plausible" / "more plausible" at its two ends and the vertical axis "less desirable" / "more desirable"; each axis now carries a single centered label — "plausibility at time of adoption" and "desirability" — with the arrowheads still marking the "more" ends.

The dashboard's hover details gained a consistent click-for-more hint, now cap their rows on short screens so they fit the viewport, and the power treemap's hover no longer misses in the gaps between cells or fails to light a country group when hovering its "Rest of X" block. The agents hover shows its speed-split line as an equation that visibly sums to the displayed fleet total, and the audio bar now drops out cleanly when the datacenter list or capability ladder takes over, and hides with the rest of the HUD during the epilogue.

• treemap hover works in cell gaps

• "Rest of X" hover lights its region

• agents line sums to fleet total

The print pages grew a full bound-book treatment: pages moved from A4 to US Letter, sidenotes alternate to the outer edge on left and right pages, and a new cover-wrap maker renders a back-cover/spine/front-cover sheet sized to the book's page count. Many smaller print fixes landed alongside.

• redesigned part-divider pages

• designed back cover added

• covers print on white, new AIF logo

Tapping into a metric detail panel on mobile showed the current story year in its header but data frozen at the timeline start — the Distribution of Power panel could read "2029" while showing ~2026 compute numbers (America 15.8M instead of 819M). All panels now show the data for the year the dashboard was displaying.

Before the reader reaches the choice point, the black projected-future segment of the mobile chart strip didn't render at all on iOS; it now draws on iPhones the same way it does everywhere else.

Frontier labs without a brand mark (Meta AI, ByteDance, G42, the anonymous "US co. N" holders) all showed the same generic AI-company logo, which wallpapered the treemap at late scenario dates when a dozen-plus of them turn frontier; now only the two biggest brand-less labs by compute show it and the rest render plain.

In Safari at a non-default page zoom, the labels along the edges of the choose-your-path decision widget drifted out of position; they now sit on their branches at every zoom level.

The narration player now offers links to listen on Spotify and Apple Podcasts alongside the on-site playback controls.

Points on the chart were hard to hit on small screens, where the chart scales down; a tap now selects the nearest point within a comfortable ~48px touch radius at every screen size, and tapping empty space clears the selection.

Links shared to social platforms now preview with the 2029 choose-a-path decision-tree card at 1200×675 (X's native card ratio) instead of the old Plan A curve card.

Pressing cmd+left/right on macOS to move through browser history was being intercepted on the scenario page; it now navigates back and forward as expected.

Two content imports updated the scenario throughout: every Plan A chapter from 2029 through the epilogue, the footnotes and postscript, the public and insider perspectives, and several supplements — the Plan A assumptions supplement gains a key-policies appendix.

• all Plan A chapters + footnotes

• public & insider perspectives

• new key-policies appendix

The "Where the covert project acquired compute" figure now shows China's 2028 inflow of new compute as 13M H100e (was 15M), of which 3M smuggled (was 5M), matching the compute supplement's §1.3.1. The smuggled tranche now resolves as 2M declared immediately, 500k to the covert project, 400k bought back, and 100k still missing (previously 3.5M declared and 900k buyback).

The chip sales-tracing tree and chip-lifecycle diagrams, previously static image exports with their own typefaces, are now drawn in the site's standard figure style: ink for the traceable sales chain, dashed red for the shell-company diversion, amber for the traceable tree. Crossing arrows into the "Decommissioned" box were untangled, and the strategy-overview chart drops a handwriting font that rendered as an arbitrary serif on some systems.

• ink/red/amber two-tone recoloring

• "trail goes dark" label removed

• overview chart: standard figure font

The six growth-explorer links in the Economics of Plan A supplement pointed at an external work-in-progress deploy headed "do not share"; they now open the site's own hosted copy of the economic growth explorer.

The AI Policies hover now says governments "are discussing, or have not begun to consider" the mapped policies, where it previously said "are still debating, or have ruled out"; the AI Labor hover's definition text was also shortened.

A leftover footnote with no reference in the text was removed from the deal-decline supplement.

The Plan A model now splits the AI fleet into a general tier and a small, much faster inference-specialized tier. The dashboard's AI legend row shows the fastest speed — e.g. "28M Reliable Agents at up to ×114 speed" instead of the old single average — and hovering the AI Labor row reveals the live split in matching counts, e.g. "28M run at ×6 speed · 23K at ×114 speed".

The bottom strip drops its frosted-glass look for a flat paper-colored card and shrinks from 100px to 76px tall; the globe now pokes above the bar's top edge with datacenters drawn as clearly visible towers. The strip also gains hand-authored camera zooms at the plan-choice beats and fades out while you read the epilogue.

• flat opaque card, 76px tall

• datacenter towers on the strip globe

• fades out during the epilogue

The header's headphone button now starts playback immediately when opened, instead of requiring a second click on play. On desktop the player bar moved to sit directly below the labor legend at the panel's full width, its unplayed track is a dashed black line, and chapter-title tooltips are no longer cut off at the bar's edges. Mobile gains its own narration row (play/pause, scrub, time, speed) under the bottom strip.

• playback starts on open

• bar sits below the labor legend

• mobile strip gains a narration row

Tapping a stat now opens its full-screen detail page instead of surfacing a floating hover tooltip; the labor legend stacks one entry per row instead of overflowing narrow columns; the fourth stat, "Total Slowdown", no longer clips against the power treemap on any phone width; and the detail pages' header year, previously missing entirely, now renders.

• stat taps open detail pages

• labor legend stacks per row

• detail pages show the year

The 62MB narration audio no longer loads until you first press play — a reader returning with the player open sees "0:00 / --:--" until playing — and the globes now pause while scrolled off screen and render at lower resolution. The mobile strip's projected-future capability line is drawn solid instead of dashed. Together these stop a reliable page crash on iOS-beta iPhones.

A new "PDF" nav item (also in the mobile menu) downloads a 70-page PDF of the main scenario, now hosted on the site itself. The PDF gains a separator page titling "Epilogue: Life after ASI", which previously ran straight on from the Postscript with no heading.

The nav's Supplements dropdown and its mobile-menu counterpart now group supplements under the same five sections as the supplements index, with hand-written one-line previews replacing auto-extracted first paragraphs (which leaked stray markup). The previously missing Economic Growth Explorer now appears under Economics, and the menu is restyled in the site's two-tone print look.

The far-future galaxy field is now built from a real astronomical galaxy catalog rather than invented ones, with hoverable tooltips on individual galaxies.

Each dot in the policy scatter chart now carries its policy's slogan as an always-visible label, replacing the old hover tooltip (labels hide on phones, where tapping still opens the panel). Axis and tooltip text are larger, the explanation panel matches the page's prose size, and the midpoint guide lines are gone.

Frontier cells for anonymous companies previously showed no logo at all; they now carry a deliberately generic hexagonal AI-company mark. The OpenAI mark, previously a rough auto-traced shape, is now the official blossom logo.

The chosen branch's text now fades and rises in over 400ms instead of popping into place, and the mobile sheet's policy block rides the opening panel rigidly instead of swimming in afterwards. The "Off the Table" policy label no longer overlaps its squares.

The desktop graph's year readout drops from 17px to 14px in a tighter pill, matching the panel-header type instead of out-inking the canvas; when a detail panel opens, the pill now morphs its size mid-flight rather than popping between the two sizes.

Hovering a plan at the choice point starts a looping 2029–2042 preview sweep, and in Safari a missed hover-end could leave the graph cycling that preview forever; the preview now reliably clears as soon as the pointer moves away.

On supplement pages, sidenotes referenced from inside a table — like the deal-decline supplement's footnote 9 — rendered at the very top of the page; they now sit beside their reference in the margin.

A content import briefly made the supplements index and nav list show two "Security in Plan A" cards; the duplicate is removed and can no longer reappear.

The scenario page's plan-switching keyboard shortcuts previously fired even with a modifier key held, so the browser find shortcut jumped branches; modified keypresses are now left to the browser.

The "Open full screen" link below the embedded takeoff explorer is gone; the embed itself is unchanged.

Sharing the site now shows a new editorial Plan A card image with a larger "AI 2040" eyebrow, sized under WhatsApp's preview cutoff so it renders everywhere.