Karpathy coined it this month and the industry immediately confessed by adoption: “vibe coding,” describing what you want in natural language, accepting what the model writes, iterating by feel, “forgetting the code even exists,” and the term’s instant ubiquity is the two-year review’s missing data point arriving as slang: a meaningful fraction of new software is now being written by people supervising rather than typing, and the file’s dual ledger opens accordingly: the democratization is real (our marketing team shipped an internal tool last sprint; the feature-flag wonder of 2013 is now available to anyone who can describe a spreadsheet), and the engineering ledger is exactly the curriculum bet’s terrain, the gap between “it runs” and “it’s sound” (security, edge cases, maintenance) is invisible precisely to those newest to it, and the review-first pedagogy just became the industry’s general problem (the file pre-registers the sequel: 2025-26’s incident reports will grow a “vibe-coded and unreviewed” category, and the orgs that built judgment infrastructure will grade the difference, the ladder question, now economy-wide). ...

The DeepSeek aftershocks settled into the pattern the file pre-registered: Nvidia recovered a meaningful fraction within the week (Jevons winning the argument, every hyperscaler’s earnings call this fortnight raised capex guidance, citing inference demand; the $600B deletion now reads as a volatility event in an intact thesis, the belief-uptime doctrine surviving its sharpest test), the R1 methodology is being replicated and extended across every lab and university cluster (the replication machine, now for training recipes, distilled reasoning models are appearing weekly, and the open-weights lag has functionally inverted on cost-per-capability), and the policy layer responded on schedule: export-control tightening proposals, congressional DeepSeek hearings, and government-device bans (the TikTok playbook, re-run at model velocity, the geography now treats weights as the border-crossing artifact). The Paris AI Summit (this week, the Bletchley lineage’s third act) completed the vibe shift with unusual honesty: safety-institute language downgraded, “opportunity” language promoted, the US vice president explicitly warning Europe against regulation, and the US and UK declining to sign the communiqué, the governance-velocity file records its first reversal: the summit series born from frontier-risk consensus now photographs an acceleration consensus, eighteen months later, with the same attendees (the governance-by-exit doctrine operating at the state level: the concerned institutions didn’t lose the argument; the argument’s venue moved). ...

Filing tonight because the market just did something this archive must timestamp: Nvidia fell ~17% today, roughly $600 billion of market cap, the largest single-day value deletion in history (the Meta record, tripled), because a Chinese lab named DeepSeek shipped a reasoning model. The week’s sequence, for the permanent record: DeepSeek-R1 released January 20th (o1-class reasoning benchmarks, open weights, MIT license, API pricing ~30x below OpenAI’s, and a published training methodology claiming frontier-adjacent capability for single-digit millions in compute, the exact cost-collapse the price-curve doctrine predicted, arriving from the exact direction the export-control regime existed to prevent); by the weekend its app topped the US App Store; by this morning the market had connected “frontier capability at 1/30th the price” to every datacenter-capex model on Wall Street and repriced the entire buildout thesis in one session. The file’s calibrated read, against tonight’s hysteria in both directions: the training-cost number is real but partial (the disclosed figure is the final run, not the program, R&D, prior runs, and the GPU fleet’s acquisition are not in it, the reconciliation doctrine applies to model cards too), the efficiency gains are real and replicable (distillation, mixture-of-experts, RL-on-reasoning without supervised scaffolding, the methods are published, which is the actual event: the open-weights lag just compressed from one generation toward zero), export controls demonstrably shaped but didn’t stop this (constrained chips bred efficiency innovation, the constraints-breed-brilliance doctrine executing as geopolitics, to Washington’s visible alarm), and Jevons Paradox is the file’s actual position on the capex panic: cheaper inference expands total compute demand on every historical precedent this archive holds (the cached-reasoning economics just got 30x more customers; Nadella tweeted the same paradox by name before markets opened, talking his book and being right, the both-things doctrine, as ever). ...

Year thirteen opens with Los Angeles burning, the Palisades and Eaton fires, driven by hurricane-force Santa Anas across a landscape that hasn’t seen rain in eight months, have destroyed entire neighborhoods (12,000+ structures and counting as I file), and the archive’s climate-infrastructure ledger adds its gravest domestic chapter: hydrant systems designed for house fires meeting neighborhood-scale conflagration (the design-basis doctrine again, urban water infrastructure has a WUI-fire envelope nobody funded), insurance markets in managed retreat (State Farm’s pre-fire non-renewals in the exact zip codes now burning, the actuaries, as ever, were the first honest climate models), and the information layer performing its now-standard double duty: real-time fire maps and mutual-aid coordination at their best, AI-generated Hollywood-sign-ablaze imagery at its worst (the verification-terminal-state runbook, deployed by every newsroom simultaneously). Friends evacuated; our LA teammates are safe and housed; the old doctrine, hug your people, opens its thirteenth year of service. Donate to the wildfire funds; the file keeps minutes and its lane. ...

Entry 292 closes year twelve, and the closing fortnight refused a quiet exit: OpenAI announced o3 on Shipmas’ final day (December 20th), the reasoning axis’s second generation, and the headline result stops the file mid-sentence: ~87% on ARC-AGI’s semi-private eval (the abstraction-and-reasoning benchmark designed to resist memorization, where GPT-4o scored single digits and four years of models flatlined), plus research-math and competition-code results that moved the frontier’s own researchers to public awe, achieved, per the disclosed methodology, with inference-compute budgets that reach thousands of dollars per task at the high end (the cost-of-thought economics now spanning six orders of magnitude: the capability exists; the price curve is the product roadmap, and every prior curve in this file’s twelve years, compute, storage, sequencing, launch-mass, says the price falls faster than the discourse expects). The file’s calibrated position for the year-end record: benchmarks are not jobs (the map-vs-territory doctrine, permanent), ARC’s designer himself notes the gap between eval-crushing and general intelligence, and the slope is the steepest this archive has ever filed across twelve years of logging “probably nothing” items (the Transformer paper to o3, the quiet thread’s decade, bookended). Both things true, maximum stakes, year thirteen inherits the question. ...

Launch-season December, fully institutionalized (the counter-programming cadence now has a holiday calendar): OpenAI is mid-“12 Days of Shipmas” (o1 full release with a $200/month Pro tier, the inference-economics file notes the price point as the margin war’s opening bid: thought, tiered; Sora’s public release Monday, the earlier demo now a product, with the provenance stack shipping alongside as C2PA metadata, graded “necessary, insufficient, present” by the file), and Google counter-launched Gemini 2.0 Wednesday (explicitly framed as “for the agentic era,” Project Mariner browsing autonomously, Astra seeing through cameras, and the watch-list item now the strategy slide of both giants simultaneously: 2025 is being pre-announced as the year models stop chatting and start doing, and the agent-spring’s expensive recursion meets its second, better-funded spring). The file’s eval-discipline note, always: “agentic” converts error rates from per-answer to per-action-sequence, compounding failure probabilities across multi-step tasks is the reliability math the demos elide (a 95%-per-step agent is a 60%-per-ten-step agent), and the orgs that instrument sequence-level golden sets first will own the trust layer the way the rehearsers own reliability. Pre-registered for 2025’s grading. ...

Thanksgiving fortnight, quiet by the year’s standards (the file’s quiet-fortnight instincts stand down for once), and the calendar delivers a reflection hook the archive refuses to waste: Saturday marks two years since ChatGPT launched, and the principal-file conducts the two-year review the industry is too busy shipping to write. Capability: from “startling autocomplete” to reasoning models, two scaling axes, frontier plurality, and open weights one generation behind; the curve’s slope survived every “wall” declaration (the file counts four major “scaling is over” discourse cycles, each followed by a capability jump, the calibration ledger suggests betting against the curve requires more courage than the discourse prices). Deployment: from demo to default, our own org’s census (the ten-day adoption) now shows AI-assisted workflows in every function including legal (the policy’s affordances-not-prohibitions bet, fully vindicated), and the review-first cohort is now training its successors (the re-rigged ladder holds weight). Discourse: matured from poles toward operations (the tractable middle won, evals, provenance, staged deployment are where the actual governance happens while the summit communiqués photograph well). Unresolved, honestly filed: the labor reallocation is real and lumpy (the strike settlements set precedents; the entry-level squeeze the ladder-question predicted is measurable in industry hiring data, and our re-rigging is a local fix to a global problem), the governance-by-exit pattern at the frontier labs remains the decade’s open risk, and the energy-and-compute buildout (inference economics at civilizational scale, the datacenter-buildout headlines now read like the CHIPS file’s sequel) is writing checks the grid literature says need a decade of plumbing. Two years; the operating environment changed; the closing wager (judgment compounds, tools amplify) holds better than its author feared. ...

The election happened (November 5th; per the usual lane: the infrastructure held again, counting proceeded, the predicted synthetic-media apocalypse arrived as scattered showers rather than the storm (the earlier fears met the replication machine’s civilian deployment: debunk-velocity mostly matched fake-velocity this cycle), and the archive notes the decisive result’s tech-sector implications, crypto markets surged on regulatory-reset expectations (the ETF flows now joined by policy beta), the antitrust remedies’ fate acquires administration-change uncertainty, and the AI executive order’s future is officially TBD, the whole regulatory-geography map now redraws at inauguration; the file files the facts and keeps its lane, per twelve years of practice). ...

The Champions League matchday the sport’s accountants dreamed of delivered: Real Madrid-Dortmund, a classic final rematch, and the match instantly entered the canon: down two goals, second half, Vinicius Junior, playing through a minor ankle strain, the play-through-the-maintenance-window lineage, hit a second-half hattrick, one of the fastest in Champions League history, and the stadium’s decibel telemetry reportedly registered on regional seismographs (the old F1 file smiles: October writes fiction nightly). Real Madrid won 5-2; the pre-writes-nothing doctrine holds, but the architecture note is already bankable: Madrid’s roster, Mbappé’s star signing (contract engineering funding the depth around him), Bellingham’s pedigree, Vinicius’s clinical finishing, is the superteam economics thesis executed with retention discipline, and if it continues, the file will mark it as roster-construction’s decade meeting its proof. ...

Yesterday morning SpaceX launched Starship’s fifth test flight, and the Super Heavy Booster, twenty-three stories of steel, returning from the edge of space at supersonic speed, flew itself back to the launch tower and was caught out of the air by the tower’s mechanical arms. The chopsticks. On the first attempt. I have watched the footage upward of thirty times and each viewing produces the same involuntary sound the earlier landings and eclipses produced, the sound of the impossible becoming a procedure. The engineering file, dutifully, beneath the awe: catching eliminates landing legs (mass, complexity, refurbishment) and enables the launch-catch-restack-relaunch cadence the whole architecture is priced on (the tower is the rapid-reuse thesis made steel, the iterate-through-explosions doctrine arriving at its payoff phase: five flights from “cleared the tower is success” to “caught the booster with a building”), and the trajectory-abort logic deserves its own sentence: the booster earned the catch attempt only after passing thousands of automated health criteria mid-descent, with the default being ocean divert, the launch-commit doctrine running autonomously at Mach speeds (the system polls itself for go/no-go now, the ship-the-judgment doctrine, twenty-three stories tall). ...