The four-day window between Christmas and New Year is, statistically, the quietest news cycle on the AI calendar. The major labs are skeleton-crewed. Press releases stop. Funding announcements wait until January. Even the Chinese labs that have spent the prior month outshipping their Western counterparts usually slow down during the universal-holiday week.
That is what I expected. It is not what happened.
The labs went quiet exactly as predicted. The practitioners did not. Simon Willison closed the year at 11:50pm on Dec 31 with his annual "2025: The year in LLMs," a long-form essay that, for the field, functions as the nearest thing to a state-of-the-union. Timothy B. Lee and the Understanding AI crew published "17 predictions for AI in 2026" on the same day with explicit calibration percentages. Willison shipped two indie tools and a careful piece on SQLite's contribution policy. And on Dec 31, while everyone else's auto-reply was on, DeepSeek's research team posted "mHC: Manifold-Constrained Hyper-Connections" to arXiv — a substrate-level architectural paper, twenty authors, not a model release.
That is the four-day window. Below the fold is the shape it implies for the year ahead.
The week's signal in one sentence
The Western lab calendar paused. The Chinese lab calendar did not. The consensus going into Q1 — set by the practitioners' year-end essays — treats Chinese open-weights as structural fact, not seasonal anomaly. The indie-builder year-ahead playbook starts there.
The hook: Willison's "2025: The year in LLMs"
The single most consequential thing of the four-day window did not ship from a lab.
Simon Willison published "2025: The year in LLMs" on Dec 31, 2025 at 11:50pm (source). It is the closest thing the field has to a canonical year-in-review — the post that gets cited as "what 2025 was" for the next twelve months by everyone who did not have time to read everything as it shipped.
The framing he settled on, condensed: reasoning models went mainstream, agents (especially coding agents) became the load-bearing user-facing pattern, Chinese open-weights labs (GLM-4.7, Kimi K2 Thinking, DeepSeek V3.2) ranked ahead of most Western alternatives, OpenAI's lead eroded across image generation, code, and open-weights territory, two-hundred-dollar-a-month subscriptions normalized, and the environmental opposition to the substrate finally crystallized into organized political pushback.
Each one of those is a thread that will run through every dispatch I write in 2026. The post pre-frames the year. Indie founders who do not read the year-end essays end up calibrating their positioning against last year's substrate map; founders who do read them have an extra month of advance notice on what the consensus will be.
For DOS specifically, the load-bearing line is the one about Chinese open-weights ranking ahead. That has been the implicit thesis of the prior two dispatches. Willison's annual is the moment it stops being implicit and becomes the consensus framing. Anyone arguing for a single-vendor architecture in 2026 is arguing against the field's year-end synthesis.
The other three practitioner items
The window had three more items from the practitioner side that matter for the year ahead.
"17 predictions for AI in 2026" — Dec 31, 2025 (source). Timothy B. Lee and the Understanding AI team. Calibrated forecasts with explicit confidence percentages — the rare format that lets you score the predictor at year-end without ambiguity. Highlights: 55% confidence that AI completes 20-hour software-engineering tasks in 2026, 80% that effective context windows plateau around 1M tokens, 75% that AI capex exceeds $500B, 90% that no major AI catastrophe occurs. These are testable and concrete. Whether you accept the calibration or argue with the numbers, it is the first serious 2026 forecast that lets you put a probability on each claim.
Simon Willison: SQLite contribution-process correction — Dec 29, 2025 (source). D. Richard Hipp publicly clarified SQLite's selectivity-over-refusal stance on outside contributions, with a Fossil-based verification protocol. Small piece on the surface; load-bearing underneath. SQLite is the most-deployed database on Earth and a transitive dependency of nearly every AI-tooling repo. How its maintainers gate outside code in the agent era is now an industry concern, not a curio. Willison's commentary is the first place I saw it framed correctly.
Simon Willison: Software Heritage Repository Retriever — Dec 30, 2025 (source). A small indie tool to recover deleted UK government open-source repositories (sqlite-s3vfs) from the Software Heritage archive. Year-end indie-tooling artifact, supply-chain-archaeology-as-a-service. The kind of thing that exists only because one person decided to spend a December afternoon writing it. The year-end window is when these get written, because the labs are quiet and attention is cheaper.
Three Willison items in one window is unusual but accurate to the data. He is, by output, the most consistent practitioner-voice in the field, and the year-end window is when his cadence is most visible relative to a quiet news cycle.
The DeepSeek tell
The only AI lab that shipped novel research in the four-day window was DeepSeek.
On Dec 31, 2025, twenty authors at DeepSeek posted arXiv 2512.24880, "mHC: Manifold-Constrained Hyper-Connections" (source). It is not a model release. It is an architectural fix to the residual streams in Hyper-Connections that restores the identity-mapping property, validated at 3B, 9B, and 27B parameter scales with a 6-7% training-cost overhead. The shape of the contribution: a substrate-level paper that future model releases — DeepSeek's and others' — will build on.
Read this in the context of the rest of the year. DeepSeek-V3.2 was the last week of November. DeepSeek-V3.2-Speciale was the first week of December. Now mHC on Dec 31. That is three substantive shipping events in five weeks, including one on a Western holiday. The cadence is the story.
The implication for an indie founder is not that DeepSeek will ship the best model of 2026. The implication is that the calendar does not protect you. If you are building on a single Western lab's release schedule, your competitors are going to ship past you during weeks when you assumed nothing was happening. The year-end window is no longer the indie's safe harbor; it is, by 2026 standards, just another week.
The pattern: practitioners frame the year, Chinese labs ship through the holiday
The four-day window in two columns
- OpenAI on PTO; no substantive shipping in window
- Anthropic on PTO; no substantive shipping after the Dec 25 cap-doubling gesture
- Google on PTO; no substantive shipping
- Most Western open-source labs quiet
- Frontier-model competitive pressure off for four days
- Willison's annual "2025 in LLMs" lands Dec 31, 11:50pm
- Lee's "17 predictions for AI in 2026" lands Dec 31 with calibration
- Willison ships SQLite commentary (Dec 29) and Software Heritage tool (Dec 30)
- DeepSeek posts mHC architectural paper Dec 31 (only lab research drop in window)
- The consensus framing for 2026 is set during the four days the West thinks it is on holiday
If you wanted to argue the year's prevailing narrative — "the substrate is plural, the protocols are public goods, the workflow layer is the moat" — Christmas week shipped four days of supporting evidence. New Year's Eve week shipped the framing for the next twelve months of the same story. Both windows reward the operator who paid attention; both penalize the operator who waited until the labs woke back up.
Two angles for an indie founder
What an indie founder building on the substrate should take from this week
- Read the practitioners' year-end essays before you do your own planning. Willison's annual, Lee's predictions, the year-end posts from credible second-tier voices (Hamel Husain, Geoffrey Litt, Ethan Mollick, Andrej Karpathy, Swyx). The consensus going into Q1 is set during the four-day window. Founders who calibrate their positioning against the year-end framing have an extra month's lead on founders who do not. The cost is a few hours of reading on a quiet week.
- The Chinese-lab cadence is now structural — match it or hedge it. DeepSeek shipped novel research on Dec 31 while the West was offline. Z.ai shipped GLM-4.7 on Dec 22. Qwen shipped Image-Edit-2511 on Dec 23. MiniMax shipped M2.1 on Dec 23. None of these dates are coincidences; they are calendar arbitrage by labs that do not observe the same holidays. As an indie, you have two paths. Path one: match their cadence on the cycles that matter to you (the high-leverage indie weeks are the ones the labs treat as low-priority). Path two: route around their cadence by depending on substrate that does not pause — open-weights, open-spec, open-protocol. Both work. Both are deliberate. The default — assuming the calendar protects you — is the one that loses.
- Spend the dead week on the cleanup the labs themselves are spending it on. Anthropic's Dec 25 cap-doubling and OpenAI's Codex-XMas are not capability moves; they are retention polish. Match their tempo. The kind of cleanup work that gets deferred all year — adapter pattern wiring, test coverage on failure surfaces, the convention catalog distillation, the failure-mode reflections you've been meaning to write down — is the work the next twelve months will reward. The window between Christmas and New Year exists to reward exactly this kind of compounding maintenance.
What this changes for DOS
Two design decisions hardened for me this week, both of them coming out of the year-end reflection rather than any single news item.
One. The credit-metered gateway moves from "January priority" to "January week one." Willison's framing — Chinese open-weights as structural fact, not anomaly — is the consensus going into Q1. If my architecture cannot route to GLM-4.7, MiniMax M2.1, or DeepSeek's next release by mid-month, I am calibrated against last year's substrate map. The contract sketch and the provider list are getting written in the next ten days, not the next twenty.
Two. I am committing to publishing a calibrated 2026-prediction set of my own by Jan 15. Lee's seventeen-prediction format with explicit confidence percentages is a discipline I have been avoiding, because uncalibrated predictions feel free and calibrated ones get scored against you. That is the point of running them publicly. By next Jan 15 I want to be able to point at last year's predictions, score myself honestly, and adjust the model. The discipline is more valuable than the predictions.
That is the kind of resolution-on-receipt that the year-end window is structurally designed to produce, even when the news cycle is louder than expected.
What I am watching for next week
The thread that runs through the week
The labs paused. The practitioners did not. The consensus framing for 2026 was set during the four days the West thought it was on holiday. The only lab research that shipped came from a lab that does not observe Western holidays — and the cadence is now itself the strategic message.
For an indie founder, the playbook reduces. Read the year-end essays. Match the calibrated-prediction discipline. Treat the dead week as the cleanup window the labs themselves are using it for. Hedge against the Chinese-lab cadence by routing across substrates. Bet on the operator loop on top — context engineering, memory, convention catalogs, the unglamorous compounding of decisions made with receipts.
That is what I am building toward. The year-end window made the bet sharper, even when the news cycle was supposed to be quiet.
— Lucas
Sources verified the week of Dec 29, 2025 - Jan 1, 2026: Simon Willison "2025: The year in LLMs" (Dec 31) · Understanding AI "17 predictions for AI in 2026" (Dec 31) · DeepSeek mHC arXiv paper (Dec 31) · Willison on SQLite contribution policy (Dec 29) · Willison Software Heritage Repository Retriever (Dec 30)
Was this page helpful?