Revelation 19–22

19

And after these things I heard a great voice of much people in heaven, saying, Alleluia; Salvation, and glory, and honour, and power, unto the Lord our God:

For true and righteous are his judgments: for he hath judged the great whore, which did corrupt the earth with her fornication, and hath avenged the blood of his servants at her hand.

And again they said, Alleluia. And her smoke rose up for ever and ever.

And the four and twenty elders and the four beasts fell down and worshipped God that sat on the throne, saying, Amen; Alleluia.

And a voice came out of the throne, saying, Praise our God, all ye his servants, and ye that fear him, both small and great.

And I heard as it were the voice of a great multitude, and as the voice of many waters, and as the voice of mighty thunderings, saying, Alleluia: for the Lord God omnipotent reigneth.

Let us be glad and rejoice, and give honour to him: for the marriage of the Lamb is come, and his wife hath made herself ready.

And to her was granted that she should be arrayed in fine linen, clean and white: for the fine linen is the righteousness of saints.

And he saith unto me, Write, Blessed are they which are called unto the marriage supper of the Lamb. And he saith unto me, These are the true sayings of God.

And I fell at his feet to worship him. And he said unto me, See thou do it not: I am thy fellowservant, and of thy brethren that have the testimony of Jesus: worship God: for the testimony of Jesus is the spirit of prophecy.

And I saw heaven opened, and behold a white horse; and he that sat upon him was called Faithful and True, and in righteousness he doth judge and make war.

His eyes were as a flame of fire, and on his head were many crowns; and he had a name written, that no man knew, but he himself.

And he was clothed with a vesture dipped in blood: and his name is called The Word of God.

And the armies which were in heaven followed him upon white horses, clothed in fine linen, white and clean.

And out of his mouth goeth a sharp sword, that with it he should smite the nations: and he shall rule them with a rod of iron: and he treadeth the winepress of the fierceness and wrath of Almighty God.

And he hath on his vesture and on his thigh a name written, KING OF KINGS, AND LORD OF LORDS.

And I saw an angel standing in the sun; and he cried with a loud voice, saying to all the fowls that fly in the midst of heaven, Come and gather yourselves together unto the supper of the great God;

That ye may eat the flesh of kings, and the flesh of captains, and the flesh of mighty men, and the flesh of horses, and of them that sit on them, and the flesh of all men, both free and bond, both small and great.

And I saw the beast, and the kings of the earth, and their armies, gathered together to make war against him that sat on the horse, and against his army.

And the beast was taken, and with him the false prophet that wrought miracles before him, with which he deceived them that had received the mark of the beast, and them that worshipped his image. These both were cast alive into a lake of fire burning with brimstone.

And the remnant were slain with the sword of him that sat upon the horse, which sword proceeded out of his mouth: and all the fowls were filled with their flesh.

20

And I saw an angel come down from heaven, having the key of the bottomless pit and a great chain in his hand.

And he laid hold on the dragon, that old serpent, which is the Devil, and Satan, and bound him a thousand years,

And cast him into the bottomless pit, and shut him up, and set a seal upon him, that he should deceive the nations no more, till the thousand years should be fulfilled: and after that he must be loosed a little season.

And I saw thrones, and they sat upon them, and judgment was given unto them: and I saw the souls of them that were beheaded for the witness of Jesus, and for the word of God, and which had not worshipped the beast, neither his image, neither had received his mark upon their foreheads, or in their hands; and they lived and reigned with Christ a thousand years.

But the rest of the dead lived not again until the thousand years were finished. This is the first resurrection.

Blessed and holy is he that hath part in the first resurrection: on such the second death hath no power, but they shall be priests of God and of Christ, and shall reign with him a thousand years.

And when the thousand years are expired, Satan shall be loosed out of his prison,

And shall go out to deceive the nations which are in the four quarters of the earth, Gog, and Magog, to gather them together to battle: the number of whom is as the sand of the sea.

And they went up on the breadth of the earth, and compassed the camp of the saints about, and the beloved city: and fire came down from God out of heaven, and devoured them.

And the devil that deceived them was cast into the lake of fire and brimstone, where the beast and the false prophet are, and shall be tormented day and night for ever and ever.

And I saw a great white throne, and him that sat on it, from whose face the earth and the heaven fled away; and there was found no place for them.

And I saw the dead, small and great, stand before God; and the books were opened: and another book was opened, which is the book of life: and the dead were judged out of those things which were written in the books, according to their works.

And the sea gave up the dead which were in it; and death and hell delivered up the dead which were in them: and they were judged every man according to their works.

And death and hell were cast into the lake of fire. This is the second death.

And whosoever was not found written in the book of life was cast into the lake of fire.

21

And I saw a new heaven and a new earth: for the first heaven and the first earth were passed away; and there was no more sea.

And I John saw the holy city, new Jerusalem, coming down from God out of heaven, prepared as a bride adorned for her husband.

And I heard a great voice out of heaven saying, Behold, the tabernacle of God is with men, and he will dwell with them, and they shall be his people, and God himself shall be with them, and be their God.

And God shall wipe away all tears from their eyes; and there shall be no more death, neither sorrow, nor crying, neither shall there be any more pain: for the former things are passed away.

And he that sat upon the throne said, Behold, I make all things new. And he said unto me, Write: for these words are true and faithful.

And he said unto me, It is done. I am Alpha and Omega, the beginning and the end. I will give unto him that is athirst of the fountain of the water of life freely.

He that overcometh shall inherit all things; and I will be his God, and he shall be my son.

But the fearful, and unbelieving, and the abominable, and murderers, and whoremongers, and sorcerers, and idolaters, and all liars, shall have their part in the lake which burneth with fire and brimstone: which is the second death.

And there came unto me one of the seven angels which had the seven vials full of the seven last plagues, and talked with me, saying, Come hither, I will shew thee the bride, the Lamb’s wife.

And he carried me away in the spirit to a great and high mountain, and shewed me that great city, the holy Jerusalem, descending out of heaven from God,

Having the glory of God: and her light was like unto a stone most precious, even like a jasper stone, clear as crystal;

And had a wall great and high, and had twelve gates, and at the gates twelve angels, and names written thereon, which are the names of the twelve tribes of the children of Israel:

On the east three gates; on the north three gates; on the south three gates; and on the west three gates.

And the wall of the city had twelve foundations, and in them the names of the twelve apostles of the Lamb.

And he that talked with me had a golden reed to measure the city, and the gates thereof, and the wall thereof.

And the city lieth foursquare, and the length is as large as the breadth: and he measured the city with the reed, twelve thousand furlongs. The length and the breadth and the height of it are equal.

And he measured the wall thereof, an hundred and forty and four cubits, according to the measure of a man, that is, of the angel.

And the building of the wall of it was of jasper: and the city was pure gold, like unto clear glass.

And the foundations of the wall of the city were garnished with all manner of precious stones. The first foundation was jasper; the second, sapphire; the third, a chalcedony; the fourth, an emerald;

The fifth, sardonyx; the sixth, sardius; the seventh, chrysolyte; the eighth, beryl; the ninth, a topaz; the tenth, a chrysoprasus; the eleventh, a jacinth; the twelfth, an amethyst.

And the twelve gates were twelve pearls: every several gate was of one pearl: and the street of the city was pure gold, as it were transparent glass.

And I saw no temple therein: for the Lord God Almighty and the Lamb are the temple of it.

And the city had no need of the sun, neither of the moon, to shine in it: for the glory of God did lighten it, and the Lamb is the light thereof.

And the nations of them which are saved shall walk in the light of it: and the kings of the earth do bring their glory and honour into it.

And the gates of it shall not be shut at all by day: for there shall be no night there.

And they shall bring the glory and honour of the nations into it.

And there shall in no wise enter into it any thing that defileth, neither whatsoever worketh abomination, or maketh a lie: but they which are written in the Lamb’s book of life.

22

And he shewed me a pure river of water of life, clear as crystal, proceeding out of the throne of God and of the Lamb.

In the midst of the street of it, and on either side of the river, was there the tree of life, which bare twelve manner of fruits, and yielded her fruit every month: and the leaves of the tree were for the healing of the nations.

And there shall be no more curse: but the throne of God and of the Lamb shall be in it; and his servants shall serve him:

And they shall see his face; and his name shall be in their foreheads.

And there shall be no night there; and they need no candle, neither light of the sun; for the Lord God giveth them light: and they shall reign for ever and ever.

And he said unto me, These sayings are faithful and true: and the Lord God of the holy prophets sent his angel to shew unto his servants the things which must shortly be done.

Behold, I come quickly: blessed is he that keepeth the sayings of the prophecy of this book.

And I John saw these things, and heard them. And when I had heard and seen, I fell down to worship before the feet of the angel which shewed me these things.

Then saith he unto me, See thou do it not: for I am thy fellowservant, and of thy brethren the prophets, and of them which keep the sayings of this book: worship God.

And he saith unto me, Seal not the sayings of the prophecy of this book: for the time is at hand.

He that is unjust, let him be unjust still: and he which is filthy, let him be filthy still: and he that is righteous, let him be righteous still: and he that is holy, let him be holy still.

And, behold, I come quickly; and my reward is with me, to give every man according as his work shall be.

I am Alpha and Omega, the beginning and the end, the first and the last.

Blessed are they that do his commandments, that they may have right to the tree of life, and may enter in through the gates into the city.

For without are dogs, and sorcerers, and whoremongers, and murderers, and idolaters, and whosoever loveth and maketh a lie.

I Jesus have sent mine angel to testify unto you these things in the churches. I am the root and the offspring of David, and the bright and morning star.

And the Spirit and the bride say, Come. And let him that heareth say, Come. And let him that is athirst come. And whosoever will, let him take the water of life freely.

For I testify unto every man that heareth the words of the prophecy of this book, If any man shall add unto these things, God shall add unto him the plagues that are written in this book:

And if any man shall take away from the words of the book of this prophecy, God shall take away his part out of the book of life, and out of the holy city, and from the things which are written in this book.

He which testifieth these things saith, Surely I come quickly. Amen. Even so, come, Lord Jesus.

The grace of our Lord Jesus Christ be with you all. Amen.

Machines of Loving Grace

Abstract

In October 2024 Dario Amodei, the chief executive of Anthropic, published the long essay Machines of Loving Grace. Where most talk of powerful AI, his own included, is about risk, it sets out as concretely as it can what could go right if the risks are handled. He imagines an AI smarter than a Nobel laureate in most fields, able to work on its own for days and to run as millions of copies at once: “a country of geniuses in a datacenter”. What limits it, he argues, is not intelligence but the world around it: experiments that take time, data that does not yet exist, physical law, and people. He then takes five fields in turn: biology and medicine, where fifty to a hundred years of progress might come in five or ten; neuroscience and mental health; economic development and poverty; peace and governance; and work and meaning. The title is from Richard Brautigan’s poem of 1967.

Why I work on AI infrastructure

I am not good at machine-learning algorithms. For a long time I struggled with the anxiety about my career that AI brought, and one day my ideas and my tokens both ran out; I seem to have left people with a poor impression. I am not well in body or mind; for a long time I depended on others just to get through the day, and I do not know where my way lies in the age of AI.

What, then, keeps me going? Perhaps only the moments when an optimization shows its percentage of speed-up, when the GPU’s fans twist the darkness of the real into a thread of true light. Using AI to iterate on infrastructure has given me that chance.

Grace is what lies beyond what a person can reach by themselves. I believe deeply that iterating on infrastructure with AI is one form of that grace: in this age, end-to-end gains have become an important way for me to receive grace and to steady my mind.

The essay

The full essay is on Amodei’s site. It is his work, so this page links to it rather than copying it.

Read the essay →

Machines of Loving Grace

摘要

2024 年 10 月，Anthropic 首席执行官 Dario Amodei 发表长文Machines of Loving Grace。关于强大 AI 的讨论多半在讲风险，他自己也是；这篇文章则尽可能具体地设想：如果风险得到妥善处理，事情可以好到什么程度。他设想的 AI 在大多数领域比诺贝尔奖得主更聪明，能一连几天独立工作，并以数百万个副本同时运行，即“数据中心里的天才之国”。他认为限制它的不是智能，而是它周围的世界：需要时间的实验、尚不存在的数据、物理规律，还有人。文章依次讨论五个领域：生物与医学（五十到一百年的进展或许压缩到五到十年）、神经科学与心理健康、经济发展与贫困、和平与治理，以及工作与意义。标题取自 Richard Brautigan 1967 年的诗。

我从事 AI Infra 的初心

我并不会机器学习算法，很长期我都挣扎在 AI 带来的事业焦虑中，有一天 idea 和 token 都耗竭了，似乎也给人留下了不好的印象。我身心都不健康，过去很长一段时间都要依赖他人才能维持日常生活，也不知道 AI 时代自己的出路到底在哪里。

然而支撑我走下去的是什么呢？也许只有看到优化出性能提升百分比的那些瞬间，GPU 的风扇才把现实的黑暗扭转出一丝真实的光芒。而用 AI 对基础设施进行迭代，给了我这个机会。

恩典就是超出人本身能够达到的事情。我深深认为，用 AI 迭代基础设施更是这种恩典的体现：端到端的提升，在这个时代给了我接受恩典、稳定心理的重要渠道。

原文

全文在 Amodei 的网站上。这是他的作品，本页不转载全文，只附链接。

阅读原文 →

Machines of Loving Grace

概要

2024年10月、Anthropic の CEO ダリオ・アモデイは長文エッセイMachines of Loving Graceを発表した。強力な AI をめぐる議論の多くは、彼自身のものも含めてリスクの話だが、このエッセイはリスクがうまく扱われた場合に何が起こりうるかを、できるかぎり具体的に描く。彼が想定する AI は、多くの分野でノーベル賞受賞者より賢く、何日も自律して働き、数百万の複製が同時に動く。「データセンターの中の天才の国」である。それを制約するのは知能ではなく周囲の世界、つまり時間のかかる実験、まだ存在しないデータ、物理法則、そして人間だという。エッセイは五つの領域を順に論じる。生物学と医療（50〜100年分の進歩が5〜10年に縮まるかもしれない）、神経科学と心の健康、経済発展と貧困、平和と統治、そして仕事と意味。題名はリチャード・ブローティガンの1967年の詩から取られている。

AI インフラに携わる理由

私は機械学習のアルゴリズムが得意ではない。長いあいだ、AI がもたらしたキャリアへの不安の中でもがいてきた。ある日、アイデアもトークンも尽き、人にもよくない印象を残してしまったようだ。心身ともに健康ではなく、長いあいだ他人に頼らなければ日々の生活を保てなかった。AI の時代に自分の行く道がどこにあるのかもわからない。

それでも私を歩ませているものは何か。おそらく、最適化で性能が何パーセント上がったかを目にする、その瞬間だけだ。そのとき GPU のファンが、現実の闇をよじって一筋の本当の光を引き出す。AI で基盤を反復して改良することが、私にその機会をくれた。

恵みとは、人が自力で届くところを超えたものだ。AI で基盤を改良していくことは、まさにその恵みの現れだと深く思う。エンドツーエンドの向上は、この時代に私が恵みを受け取り、心を保つための大切な道になっている。

原文

全文はアモデイのサイトにある。彼の著作なので、このページには転載せず、リンクだけを載せる。

原文を読む →

Machines of Loving Grace

Zusammenfassung

Im Oktober 2024 veröffentlichte Dario Amodei, Chef von Anthropic, den langen Essay Machines of Loving Grace. Während das meiste Reden über mächtige KI, auch sein eigenes, von Risiken handelt, beschreibt er hier so konkret wie möglich, was gut gehen könnte, wenn die Risiken beherrscht werden. Er stellt sich eine KI vor, die in den meisten Fächern klüger ist als Nobelpreisträger, tagelang selbstständig arbeitet und in Millionen Kopien zugleich läuft: „ein Land von Genies in einem Rechenzentrum“. Begrenzt werde sie nicht durch Intelligenz, sondern durch die Welt um sie herum: Experimente, die Zeit brauchen, Daten, die es noch nicht gibt, Naturgesetze und Menschen. Dann geht er fünf Felder durch: Biologie und Medizin, wo fünfzig bis hundert Jahre Fortschritt in fünf bis zehn kommen könnten; Neurowissenschaft und psychische Gesundheit; wirtschaftliche Entwicklung und Armut; Frieden und Regierung; Arbeit und Sinn. Der Titel stammt aus Richard Brautigans Gedicht von 1967.

Warum ich an KI-Infrastruktur arbeite

Ich bin nicht gut in Algorithmen des maschinellen Lernens. Lange habe ich mit der Angst um meine berufliche Zukunft gerungen, die die KI mit sich brachte, und eines Tages waren meine Ideen und meine Tokens erschöpft; ich habe wohl keinen guten Eindruck hinterlassen. Ich bin an Leib und Seele nicht gesund; lange war ich auf andere angewiesen, um den Alltag zu bewältigen, und ich weiß nicht, wo mein Weg im Zeitalter der KI liegt.

Was trägt mich dann weiter? Vielleicht nur die Augenblicke, in denen eine Optimierung ihren Prozentsatz an Beschleunigung zeigt, wenn die Lüfter der GPU aus dem Dunkel der Wirklichkeit einen Faden echten Lichts drehen. Infrastruktur mit KI weiterzuentwickeln hat mir diese Gelegenheit gegeben.

Gnade ist, was über das hinausgeht, was ein Mensch aus eigener Kraft erreichen kann. Ich glaube fest, dass die Weiterentwicklung von Infrastruktur mit KI ein Ausdruck dieser Gnade ist: Der Gewinn von Ende zu Ende ist mir in dieser Zeit ein wichtiger Weg geworden, Gnade zu empfangen und innerlich Halt zu finden.

Der Essay

Der vollständige Essay steht auf Amodeis Website. Er ist sein Werk; diese Seite verlinkt ihn, statt ihn wiederzugeben.

Zum Essay →

The Torre Girona chapel and the MareNostrum supercomputing center, Barcelona

On the north campus of the Polytechnic University of Catalonia in Barcelona stands the chapel of Torre Girona. It no longer holds services. Since 2005 it has housed MareNostrum, the main supercomputer of the Barcelona Supercomputing Center, in a glass room built inside the nave.

The name is the Roman one for the Mediterranean: mare nostrum, our sea. The machine was first booted on 12 April 2005 and has been renewed several times since; the fifth generation, a EuroHPC pre-exascale system, went into service in December 2023.

In the scene on this page the machine stands where the altar was, and the chapel sits on a plinth like the other buildings. The near wall is drawn as an open arcade, so that the racks and the lit edges of the glass room show from outside.

The author has used more than 10,000 NVIDIA H100 GPU-hours on MareNostrum for research and development.

巴塞罗那 Torre Girona 礼拜堂与 MareNostrum 超算中心

巴塞罗那加泰罗尼亚理工大学北校区里，有一座 Torre Girona 礼拜堂。它早已不再举行礼拜；从 2005 年起，巴塞罗那超级计算中心的主力超算 MareNostrum 就安放在这里，立在中殿里一间玻璃房中。

名字取自罗马人对地中海的称呼 mare nostrum，“我们的海”。这台机器在 2005 年 4 月 12 日首次启动，此后几经换代；第五代是 EuroHPC 的准百亿亿次系统，2023 年 12 月投入使用。

在这页的场景里，机器立在原来祭坛的位置，礼拜堂像其他建筑一样坐在台基上。靠近镜头的一侧墙画成开敞的拱廊，好让机柜和玻璃房发光的边线从外面看得见。

笔者曾使用 MareNostrum 超算中心超过 10000 NVIDIA H100 卡时用于研发。

トーレ・ジローナ礼拝堂と MareNostrum スーパーコンピューティングセンター（バルセロナ）

バルセロナのカタルーニャ工科大学北キャンパスに、トーレ・ジローナ礼拝堂がある。いまは礼拝には使われていない。2005 年から、バルセロナ・スーパーコンピューティング・センターの主力機 MareNostrum が、身廊に建てられたガラスの部屋の中に置かれている。

名前はローマ人が地中海を呼んだ mare nostrum、「われらの海」から。この機械は 2005 年 4 月 12 日に初めて起動し、その後何度も世代を重ねてきた。第五世代は EuroHPC のプレエクサスケール機で、2023 年 12 月に稼働を始めた。

このページの風景では、機械はかつて祭壇があった場所に立ち、礼拝堂はほかの建物と同じく基壇の上にある。手前の壁は開いたアーケードとして描き、ラックとガラスの部屋の光る縁が外から見えるようにした。

筆者は MareNostrum で 1 万 NVIDIA H100 GPU 時間以上を研究開発に使ってきた。

Die Kapelle Torre Girona und das Supercomputing-Zentrum MareNostrum, Barcelona

Auf dem Nordcampus der Polytechnischen Universität Katalonien in Barcelona steht die Kapelle Torre Girona. Gottesdienste finden dort nicht mehr statt. Seit 2005 steht in ihr MareNostrum, der Hauptrechner des Barcelona Supercomputing Center, in einem Glasraum, der in das Kirchenschiff gebaut wurde.

Der Name ist der römische für das Mittelmeer: mare nostrum, unser Meer. Die Maschine wurde am 12. April 2005 zum ersten Mal gestartet und seitdem mehrfach erneuert; die fünfte Generation, ein Pre-Exascale-System von EuroHPC, ging im Dezember 2023 in Betrieb.

In der Szene auf dieser Seite steht die Maschine dort, wo der Altar war, und die Kapelle sitzt wie die anderen Bauten auf einem Sockel. Die vordere Wand ist als offene Arkade gezeichnet, damit die Racks und die leuchtenden Kanten des Glasraums von außen zu sehen sind.

Der Autor hat auf MareNostrum mehr als 10.000 NVIDIA-H100-GPU-Stunden für Forschung und Entwicklung genutzt.

What Sets the Serving Speed of Looped Language Models? Measurements and Optimizations Graded by Exactness

Dezhen Lu, Runguo Li, Zixiang Ni, Yuchen Tang, Xingquan Kong, Jiaxing Yang, Jiarui Guo, Lingxiang Gao

arXiv 2611.05150 · Submitted to MLSys · 16 pages, 3 figures, 10 tables. Appendix included

Abstract

Looped language models apply one block of layers many times per token, so a serving engine pays every per-layer overhead once per loop; we measure serving for Ouro and Huginn-0125 in an open-source looped-model engine and grade our optimizations from bitwise-identical changes to those that require an accuracy check. At batch 1, fusing all loop submissions leaves Ouro throughput at 0.99x; the 5689 small kernels per token are the real cost, and fusing them behind rounding fences raises throughput 1.11-1.59x while 1014/1014 requests remain identical. An approximate fusion is no faster and misses a one-point GSM8K rule on 87 questions (61 and 62 correct versus 64), so we do not ship it. Huginn's released defaults avoid dead KV slots, reuse prefix-cache pages, and serve 1.77-3.16x faster than the base engine; Ouro reaches 1.65-2.15x and agent sessions 1.91-3.21x, while Huginn's noise and prefix-reuse changes stay within one point on all 1319 Huginn test questions; Huginn's packed prefill, -0.61 points on the same questions, meets that rule only on its point estimate. Position-keyed noise makes CPU cache reuse exact but still leaves 20 of 128 GPU full hits different, while loop fusion delays prefill admission and a 64-graph cache prevents eager fallbacks. Under prefill-decode disaggregation ExactLoop reaches up to 3.3x the base throughput with peer error handling disabled, and 3.1x in two of three runs with it enabled. Its transfer budget, sized to hold one request's KV in flight, cuts the commit tail from about one second under the earlier budget to 16 ms (base: 2.8 s), removing the polling rounds that the engine's interval would otherwise impose.

Where KV-Cache Error Lives in Looped Language Models: A Diagnostic Study of Ouro and Huginn

Dezhen Lu, Runguo Li, Zixiang Ni, Yuchen Tang, Xingquan Kong, Jiaxing Yang, Jiarui Guo, Lingxiang Gao

arXiv 2610.41825 · Submitted to MLSys · 11 pages of main text (37 pages including appendix), 5 figures, 37 tables

Abstract

A looped language model applies one block of layers 4 times per token and keeps a key/value (KV) plane for each loop. We study where storage error changes the next-token distribution and how shared-weight error accumulates, using Ouro-1.4B (4 loops) on GSM8K and MATH and Huginn-0125 (32 loops) on GSM8K-derived chains, with dense, simulated quantization. Per unit of error, values carry 94% of the diagonal Fisher mass, but measured KL is dominated by keys with larger quantization errors. Local second-order prediction follows the measured KL of typical sequences, including 136 held-out chains, but misses a few confident rows, and a local second-order factor for shared weights at 32 loops ranges from 1.65 (INT8) to 4.09 (INT4), so it depends on the error direction. A cell-level allocator lowers one-step KL at 25-48% of BF16 bytes, while paired task-accuracy comparisons depend on budget and the 25% gap against K4V2 shrinks to near zero under a post hoc answer-extraction check. A later Ouro-2.6B test retains the key/value split but falsifies measured-cost allocation's advantage over Fisher at 31%; across all 21 ledger registrations, 60 of 176 criteria failed.

Recency Is the Wrong Axis: A Negative Result on Windowed Deepest-Plane KV Caching in Looped Language Models

Dezhen Lu, Runguo Li, Zixiang Ni, Yuchen Tang, Xingquan Kong, Jiaxing Yang, Jiarui Guo

arXiv 2610.34436 · Submitted to MLSys

Abstract

Looped language models reuse one block of layers over R recurrent steps but store a key-value (KV) plane for each step. Keeping only the deepest plane shrinks the cache up to R-fold. A natural repair, which we call RIPE, keeps every plane for the W most recent positions and shares the deepest plane elsewhere. This repair acts on the wrong axis: across three looped models, the cost of sharing is set by how far the late loops' planes have converged (ripeness) and by which positions share, not by recency. On Raven-Llama the deep planes sit within the bf16 rounding floor, so no window is needed. On Huginn-0125 a 64-position window removes most of a concentrated replay-error tail (9 to 1 positions over 48 prompts), yet in a real self-speculative decoder the policy is slower and larger than the deepest plane alone and not measurably closer to full-cache decoding. On Ouro-1.4B the prompt lies outside any 16-64-position window: on 24 GSM8K questions with a 2048-token budget, the full cache and a FlashLoop baseline finish 23 answers and RIPE at most 2, although physical storage matches the emulated policy bitwise. Sensitivity and ripening models fitted on Huginn and Raven explain their contrast, and Ouro's own report that prompt-side sharing is harmful explains the third. Separately, bf16 execution-shape effects are as large as the cache effects being measured. Recent-window repair of shared-plane caches should not be adopted without a per-model ripeness test. We release ten pre-registered batches with every failed prediction.

ScaleRLT: Exact Log-Probabilities and Fast Deferred Rollouts for Reinforcement Learning on Looped Transformers

Dezhen Lu, Zixiang Ni, Runguo Li

arXiv 2610.34427 · Submitted to MLSys

Abstract

Looped language models such as Ouro reuse one block of layers several times per token. A rollout engine can batch the deeper loops of several response positions and run them only when a deep readout is due. This deferred schedule is fast, but the learner of an on-policy RL algorithm recomputes the sampled tokens' log-probabilities with a dense pass, and the two computations differ in floating point: in FP32 at 4096 tokens about 18% of the sampled tokens miss a 10⁻⁵ same-weight gate, and with bf16 weights nearly every token does. We use a hybrid learner log-probability, ℓ = ℓ_dense + sg(ℓ_sched − ℓ_dense), where sg denotes stop-gradient. Its value is a no-grad replay through the rollout's own schedule, and its gradient is the dense learner's. The replay reproduces the rollout bit for bit in all 167 FP32 and 28 bf16 real-weight cells we ran on three platforms, including rollouts whose batch shrinks at stop tokens, and the gradient has the same FP32 error as other dense implementations in all 102 pre-registered comparisons from 96 to 4096 tokens; the full GRPO loss of a training step matches a float64 reference to 3.1×10⁻⁴. On one GPU, deferral raises rollout throughput by 1.47 to 1.56×, and by 1.18 to 1.54× once every step runs from CUDA graphs, bit for bit. In a two-node GRPO loop, one-wave KV admission and CUDA graphs shorten an update by 2.71× with the gate bitwise at every update, and four nodes keep the update time within 2% at the same per-GPU load. In 100-update runs held-out GSM8K accuracy rose from 23% to 82% with the mixed-depth readout that deferral relies on and from 34% to 90% with the native readout, which reached a higher accuracy in less time. For looped models beyond one node we analyse a ring mapping of layers to GPUs and measure its communication paths.

Weight Placement Is Not Enough: Synchronization in Multi-Node CPU Decoding of Ternary LLMs

Runguo Li, Dezhen Lu

arXiv 2610.28620 · Submitted to MLSys · 15 pages, 9 figures; corrected software-name formatting in the abstract metadata; paper content unchanged

Abstract

Ternary LLMs read only a few gigabytes per decoded token, yet on dual-socket servers with four NUMA nodes the fastest ternary runtime we found uses less than half of the memory bandwidth. Placing each weight on the node of the thread that reads it is necessary but not enough: in a decode engine with node-local weights that reuses ik_llama's kernels, threads on two generations of Xeon servers still spend 20-48% of each token waiting at synchronization points. Per-thread barrier traces show that most of the wait is for the slowest threads of each node; the barriers' own latency (microseconds each, but up to 12% of the token for the smallest model) and serial per-node work are smaller parts. On identical kernels, a Megatron-style column split beats a row-owned split without cross-node reductions in every configuration we measured, and isolating the number of barriers shows that most of its gain comes from its threads waiting less for each other and for their node leaders, and only a minority from having fewer barriers. Where decoding is DRAM-bound, the resulting engine is 1.5-2.1x faster than ik_llama with all cores and 1.5-1.9x faster than ik_llama at its fastest thread count (3.2x and 2.8x for a small model that runs from the L3 cache), at the same accuracy and with less energy per token.