Non DubitoA Mind in Many Languages
Search搜索搜尋
EN
日本語FrançaisDeutschEspañol한국어

When Switching Stops Being a Choice为什么越来越难换一家為什麼越來越難換一家 · 06

The Key Downstairs Does Not Own the Floors Above

楼下的钥匙,怎样管到了楼上

樓下的鑰匙,怎樣管到了樓上

Compute, cloud, models and interfaces build on one another. Integration creates capabilities; it can also decide who gets to build above them.

算力、云、模型、接口,一层托着另一层。把它们连起来可以长出能力,也可能让别人再也接不进来。

算力、雲、模型、接口,一層托著另一層。把它們連起來可以長出能力,也可能讓別人再也接不進來。

Han Qin ·

The open-model group waits six months for compute. Another developer waits four months, then hears that no capacity remains for its next request. Everybody knows supply is tight, but Beacon's work goes first. Stone Granary's annual report gives the arrangement a name: strategic-partner priority. A name explains the relationship while leaving its effects on other builders to be examined.

This is the fictional Kingdom State. Stone Granary supplies roughly seventy percent of the state’s rentable high-end compute and continues building data centers. Collaboration reduces Beacon's training costs. Scarcity itself causes queues; priority cannot be assumed to explain every wait. The story puts two effects together without establishing deliberate obstruction. It asks us to examine a position holding one layer while favoring a partner above it.

One layer is compute; another consists of people building models with it. Stone Granary may have sound grounds for owning the capacity it created. Those grounds do not automatically answer for scheduling in the next layer. Operating machines does not make the possibilities for developing the next model entirely a matter to be settled through its preferred commercial relationship.

A key downstairs does not own the rooms above. Technology necessarily builds in layers: applications on operating systems, models on cloud services, compute on chips. Dependence is not the objection. Examine how control of one layer arranges entry, timing, and practical possibility in another. The need to connect things cannot provide an unlimited justification for deciding everything connected to them.

Partner priority tells us why Beacon is first without explaining what routes remain for others. How are queues established, can they be examined, do competing projects face time windows, and are alternatives genuinely unavailable? Evidence matters. Identifying a partnership establishes a relationship, not the justification of every consequence flowing from it. The reply must reach the particular limitation being questioned.

Crossways supplies a parallel case. It adopts Beacon’s assistant as the marketplace’s Wayfinder. Only Wayfinder may place orders there; a buyer’s chosen assistant cannot. Managing a marketplace has become a rule about another layer: who may act as a purchasing agent. That further exercise needs its own explanation. It cannot all be justified by saying the company created the market through which the buyer wishes to shop.

The 2001 Microsoft appellate opinion makes the distinction concrete. The court maintained the core operating-system monopoly-maintenance finding, reversed attempted monopolization of browsers, remanded tying, and vacated the remedial order. It did not find a completed browser monopoly. It examined suppressing threats from an adjacent layer to maintain the operating-system position. Liability and the proposed remedy also required separate treatment.

This is more than legal detail. Reducing the case to an operating-system company also made a browser would omit the conduct under examination. Entering another layer and connecting products are not themselves accusations. The question follows the particular use of an existing position, through evidence, rather than treating every vertically connected business as already guilty or already justified.

Standards and interfaces form smaller doors between layers. Beacon writes access rules; the city connects payments, clinics connect bookings. Ask the writer what parameters and limits are built in, what is guaranteed, and what is not. Ask the adopter why it chose that entrance, what alternatives remain, and how withdrawal or continuation works. These are two accounts, not a responsibility that disappears between them.

Retiring an old interface forces the city to pay for changes and interrupts assistant-based payments for three weeks. Beacon answers for version arrangements; the city answers for connecting residents' route in this way. This essay does not allocate every repair cost between them. Distinguishing responsibilities is not permission to shift the same loss arbitrarily from one party to the other.

Crossways writes the marketplace rules and decides that only Wayfinder can order there. It has to answer for each choice. Beacon writes the interface its assistant uses to call other services; Beacon answers for that gate. Even when the same firm writes and adopts a standard, it is still answerable for both. People who have to pass through these gates need to know whom they can challenge about each decision.

Nvidia's proposed acquisition of Arm placed a different entrance before regulators. Arm licenses processor technology to multiple firms, including the buyer's rivals. The UK's competition authority found a realistic prospect of substantially lessened competition at phase one, including restrictions on access and interoperability. The transaction was abandoned and the phase-two inquiry cancelled. The proposed structure and its procedural status must remain together.

This is an abandoned proposed transaction, not a claim that the acquisition or the feared restrictions occurred. Its national-security strand is not recast here as an antimonopoly finding. The competition question concerns what happens to access when a layer used by many would belong to one participant using it. A phase-one possibility cannot be rewritten as a final finding of accomplished unlawful conduct.

Return to Stone Granary. The partners later share an engineering team, combine some accounts, and settle compute and models together. Another year on, Beacon's training runs only on Stone Granary's platform, while scheduling on that platform uses Beacon's model. These are no longer just statements of cooperation. Each firm's daily operation has grown into the other, changing the practical work of separation.

Integration provides real capability: lower training costs, less coordination, new centers. Separation may genuinely be slow and expensive. But at least two years, with major risks is Stone Granary's answer, not an independently established conclusion. Difficulty deserves examination without becoming evidence that every possible separation or handover has already been shown impossible. Its source, scale, and possible remedies require further work.

Nor must we wait for failure to see dependence. Training runs in one place today; queues restrict what others can build today. These facts can be examined before a violation is established. Looking early is not punishing early. It means a future decision need not begin with nothing but an alarming estimate supplied by the only provider remaining able to explain the system.

Readiness for separation preserves conditions before separation occurs. Interfaces might permit replacement, compute and model accounts remain intelligible separately, and handover arrangements identify who continues which service and have been tested. These are not interchangeable remedies. Separate accounts do not make software portable by themselves; a public interface does not guarantee a willing provider. Each preparation does a different piece of work.

Preparations cost something. Independent interfaces may slow work, accounting requires administration, and standby equipment may sit unused. Explain what possibility each preserves and what uncertainty it reduces. A potentially expensive future failure cannot automatically justify every present expense. The direction is to preserve capabilities and options, not to promise that making them separable is costless or always possible without losses.

A historical separation shows that deep integration is not a natural impossibility. AT&T's settlement arrangements separated local telephone operations from long-distance and equipment businesses, taking effect in 1984. Large-scale structural separation has been accomplished. That limited fact neither makes it free of costs nor selects it as the solution for every integrated position. It establishes possibility, not a universal instruction.

Imagine fields in a future Kingdom State. Hundredcraft's machines use Beacon's models, and their joint farming business works land for farmers. Machines recognize specified seeds and fertilizer; competitors' varieties require certification. Precision may genuinely protect harvests. Certification can also become a gate. Those possibilities need separate examination rather than resolving the arrangement merely by naming either safety or exclusion.

Over a decade, food prices fall, injuries diminish, and abandoned fields return to cultivation. Preserve those gains. Meanwhile fewer people have personally farmed a season. A system fault stops harvesters in two counties for nine days, and public authorities cannot find people and machines to take over. Efficiency and a thin capacity for continuation can coexist without either canceling the other.

The stakes now differ. Trouble switching an assistant and an interrupted food supply do not have equal weight. For food and water, public arrangements must secure provision as well as constrain operators. What needs protection is people's access to necessities, not a provider's right to remain unchanged. Fieldworks operates about seventy percent of the Kingdom State’s farmland. Other operators exist, but that alone does not show they could meet all the need if Fieldworks stopped.

Models, machines, seeds, and food connect through distinct reasons and responsibilities. Success in machinery does not automatically justify exclusivity in seeds. Questioning specifications does not erase safer machines. Separating the accounts avoids the false choice between rejecting technology wholesale and accepting every binding attached to it. It also makes it possible to locate a particular change instead of condemning an entire chain.

What should be separated or constrained requires proposals, evidence, and comparison of costs. Integration comes with no automatic accusation; difficulty comes with no automatic immunity. Preserve, where possible, the capacity to choose later: to change suppliers, continue work, inspect accounts, and reopen a decision rather than discover that the original arrangement has made every alternative practically unaffordable.

The group waiting six months need not demand permanent first place. It needs an account of scheduling, reasons, a route for challenge, and a next possibility not requiring the same key again. A design that looks convenient from downstairs still needs to leave room for other people's doors on the floors above.

开放模型组申请一批算力,等了六个月。另一家模型公司先等四个月,下一次被告知本季度没有余量。它们都知道算力紧,可灯塔的任务总在前面。石仓年报给这件事一个名字:战略伙伴优先。

我们仍在虚构的王国州。州内可租用的高端算力约有七成在石仓,它也还在建新的数据中心。灯塔与它合作,训练成本低了。等候不全由优先造成,紧缺本身也会使人等;这个故事没有替我们证明谁故意卡住谁,它把两种作用摆到了一处。

一层是算力,另一层是使用算力做模型的人。石仓建起数据中心,可能完全有根据地拥有上一层的能力。这份根据,却不会自动替下一层的排期作答。它能调机器,不等于谁可以做出下一种模型,都只该由它的伙伴关系决定。

这就是楼下钥匙与楼上房间的分别。技术本来相互依赖,操作系统托着应用,云托着模型,芯片托着算力。要问的不是能不能依赖,而是一个位置怎样借自己握住的一层,安排别的层里谁先做、谁能进、谁只能一直等。

战略伙伴优先解释了灯塔为什么在前,却没有解释别人实际还剩什么路。排期规则怎样定,能否查,另一家的项目是否也有时间窗口,缺口是否真无替代,都需要材料。说出伙伴关系,是说明关系存在,不是把关系的每种后果一并说明正当。

通衢也做过一件相似的事。它采用灯塔助手作为市集里的问路助手,只允许问路替人下单,买家自己选的助手进来办不到。管理市集的规则,走到了代理服务的一层:不仅决定货怎样卖,还决定谁能替你来买。后一种选择需要另答,不能全包在我开了这个市场里。

一个旧案让这层次更清楚。2001年美国微软案的上诉法院,维持了非法维持操作系统垄断的核心责任判断,同时撤销企图垄断浏览器市场一项,把捆绑问题发回,撤销原救济令。它没有认定微软已经垄断浏览器一层,而是认定打压下一层的对手,被用来维护上一层的位置。

这一分别不只是法律细节。若把案子缩成做操作系统的也做浏览器,所以错了,就遗漏真正被审的做法。进入另一层、把产品连起来,不本身作罪名;怎样使用已有位置,才是可以沿证据读下去的那一件。

标准与接口是两层之间更小的门。灯塔写接入守则,市政接上缴费,诊所接上排号。对写的一方,问它把什么参数放进去、限制在哪里、保证什么。对采用的一方,问它为何采用、有没有别的入口、出问题怎样退或继续。

停用旧接口,使市政花钱改造,三周不能经助手缴费。灯塔得回答版本安排,市政得回答为什么把市民的一条路接成这样。两边怎样分摊改造费,不能只靠这篇文章裁;各自有账,却也不因此把一笔损失随意在两家之间转。

通衢写市集规则,又决定只让问路代买,这两件事都要分别说明。灯塔写的是助手调用其他服务的接口,那道门则由灯塔说明。有的地方,写标准和采用标准的是同一家,两项责任仍不能合成一项。外人经过这些门时,得知道每一件事能向谁追问。

英伟达拟收购 Arm 的交易,曾把另一种门摆在审查前:Arm向多家芯片企业授权架构,其中有买方的对手。英国竞争与市场管理局第一阶段认为,交易有现实可能显著减损竞争,担忧取得关键授权及互通的路被限制。后来交易放弃,第二阶段调查取消。

这里读的是一项已经放弃的收购提案,不是说收购做成、限制已经发生,也不是把国家安全审查改叫垄断结论。它提出的具体问题是,许多人所用的一层如果归门内的某一家,其他使用者未来的接入会怎样。第一阶段的可能性,不能改写成终局的既成违法。

再回石仓。两家后来合用工程团队,部分账合在一起,算力和模型一体结算。又过一年,灯塔的训练只能在石仓平台跑,石仓的调度也用了灯塔模型。它们不只是签了合作意向,而是各自的日常已经长进另一方里面。

这份整合确有能力:训练省钱,少了来回协调,新的中心建起来。要把它分开,也许真的慢、真的贵。可是至少两年、风险很大,是石仓自己的答复,不是已经独立核实的结论。不能拿一句难办,就当作所有分开与接续都已被证明办不到。

也不必等它倒下才看依赖。今天训练只能在一处跑,今天队伍排不开,这些已经可以核。困难的规模和来路值得被看见,即使尚未认定任何一家越线。提前看,不等于提前处罚;它使后来要改变时,手里不只剩一份惊人的估价。

可拆的准备,是在还不拆时给未来留条件。例如接口能按公开方式相连,算力与模型的账仍能各自读懂,哪一项停了由谁接续,预先安排并试过。这些不是三种同义词:分账不会自动使系统搬得走,接口公开也不自动保证有人肯接手。

准备要付代价。独立接口可能慢一些,分账多些管理工作,备用机器平时可能闲着。它们保住哪一种可能,日后少了什么不确定,要逐项说明;不能把明天也许很贵的一场事故,拿来替今天每一笔支出证明合理。

一个历史上的分离提醒人,长在一起不是自然界的不可能拆。AT&T在和解安排下,把本地电话运营与长途及设备等业务分开,1984年生效。这件事说明大规模结构分离做成过,不说明做成没有代价,也不替所有整合位置选定同一种解法。

把镜头移到设想中的田地,这个问题会更实在。百工的机器由灯塔模型驱动,两家办田畴替农户耕作。机器只认指定种子和肥料,别家的要认证。田畴说,规格是为精度。精度可能真护住收成,认证也可能成为别人种子进不来的门,要分别核。

十年里粮价降了,伤亡少了,一些荒地重种。这些改进值得保留。与此同时,少了许多亲手种过一季地的人,一次系统出错使两个县收割停了九日,公共一侧找不到接手的人和机器。效率与接续的薄弱,并不因看起来相反就不能同在。

这里的轻重变了。作坊换助手费力,与粮食供给接不上,分量不同。对粮和水,公共安排不仅要约束经营者,也要确保供给能够持续。要保的是人吃得到,不是哪家提供者无论怎样都不能改变。田畴经营着王国州约七成耕地,仍有别的经营者;但这不证明田畴停摆以后,其他人就能接住全部需要。

模型、机器、种子、粮,一层层接起来,各有自己的理由与责任。机器上的成功,不自动给出种子上的排他许可;种子规格需要核,不使机器带来的安全改进消失。这样分开看,才不必在全盘拒绝技术和全盘接受绑定之间选一边。

具体该拆、该限什么,需要比较方案、证据和代价。整合不自带罪名,困难也不自带豁免。真正值得提前保住的,是后来仍能作选择的能力:可以换、可以接、可以看清账,必要时能把决定从原来那一双手里重新打开。

那支等了六个月的模型小组,未必要求自己永远排第一。它需要知道怎样排,为什么排,哪里能问,以及下一条路不是必须再由同一把钥匙放行。从楼下看上去很方便的设计,走到楼上,还要容得下别人的门。

開放模型組申請一批算力,等了六個月。另一家模型公司先等四個月,下一次被告知本季度沒有餘量。它們都知道算力緊,可燈塔的任務總在前面。石倉年報給這件事一個名字:戰略夥伴優先。

我們仍在虛構的王國州。州內可租用的高端算力約有七成在石倉,它也還在建新的資料中心。燈塔與它合作,訓練成本低了。等候不全由優先造成,緊缺本身也會使人等;這個故事沒有替我們證明誰故意卡住誰,它把兩種作用擺到了一處。

一層是算力,另一層是使用算力做模型的人。石倉建起資料中心,可能完全有根據地擁有上一層的能力。這份根據,卻不會自動替下一層的排期作答。它能調機器,不等於誰可以做出下一種模型,都只該由它的夥伴關係決定。

這就是樓下鑰匙與樓上房間的分別。技術本來相互依賴,操作系統托著應用,雲托著模型,芯片托著算力。要問的不是能不能依賴,而是一個位置怎樣借自己握住的一層,安排別的層裡誰先做、誰能進、誰只能一直等。

戰略夥伴優先解釋了燈塔為什麼在前,卻沒有解釋別人實際還剩什麼路。排期規則怎樣定,能否查,另一家的項目是否也有時間窗口,缺口是否真無替代,都需要材料。說出夥伴關係,是說明關係存在,不是把關係的每種後果一並說明正當。

通衢也做過一件相似的事。它採用燈塔助手作為市集裡的問路助手,只允許問路替人下單,買家自己選的助手進來辦不到。管理市集的規則,走到了代理服務的一層:不僅決定貨怎樣賣,還決定誰能替你來買。後一種選擇需要另答,不能全包在我開了這個市場裡。

一個舊案讓這層次更清楚。2001年美國微軟案的上訴法院,維持了非法維持操作系統壟斷的核心責任判斷,同時撤銷企圖壟斷瀏覽器市場一項,把捆綁問題發回,撤銷原救濟令。它沒有認定微軟已經壟斷瀏覽器一層,而是認定打壓下一層的對手,被用來維護上一層的位置。

這一分別不只是法律細節。若把案子縮成做操作系統的也做瀏覽器,所以錯了,就遺漏真正被審的做法。進入另一層、把產品連起來,不本身作罪名;怎樣使用已有位置,才是可以沿證據讀下去的那一件。

標準與接口是兩層之間更小的門。燈塔寫接入守則,市政接上繳費,診所接上排號。對寫的一方,問它把什麼參數放進去、限制在哪裡、保證什麼。對採用的一方,問它為何採用、有沒有別的入口、出問題怎樣退或繼續。

停用舊接口,使市政花錢改造,三周不能經助手繳費。燈塔得回答版本安排,市政得回答為什麼把市民的一條路接成這樣。兩邊怎樣分攤改造費,不能只靠這篇文章裁;各自有帳,卻也不因此把一筆損失隨意在兩家之間轉。

通衢寫市集規則,又決定只讓問路代買,這兩件事都要分別說明。燈塔寫的是助手呼叫其他服務的介面,那道門則由燈塔說明。有的地方,寫標準和採用標準的是同一家,兩項責任仍不能合成一項。外人經過這些門時,得知道每一件事能向誰追問。

英偉達擬收購 Arm 的交易,曾把另一種門擺在審查前:Arm向多家芯片企業授權架構,其中有買方的對手。英國競爭與市場管理局第一階段認為,交易有現實可能顯著減損競爭,擔憂取得關鍵授權及互通的路被限制。後來交易放棄,第二階段調查取消。

這裡讀的是一項已經放棄的收購提案,不是說收購做成、限制已經發生,也不是把國家安全審查改叫壟斷結論。它提出的具體問題是,許多人所用的一層如果歸門內的某一家,其他使用者未來的接入會怎樣。第一階段的可能性,不能改寫成終局的既成違法。

再回石倉。兩家後來合用工程團隊,部分帳合在一起,算力和模型一體結算。又過一年,燈塔的訓練只能在石倉平台跑,石倉的調度也用了燈塔模型。它們不只是簽了合作意向,而是各自的日常已經長進另一方裡面。

這份整合確有能力:訓練省錢,少了來回協調,新的中心建起來。要把它分開,也許真的慢、真的貴。可是至少兩年、風險很大,是石倉自己的答覆,不是已經獨立核實的結論。不能拿一句難辦,就當作所有分開與接續都已被證明辦不到。

也不必等它倒下才看依賴。今天訓練只能在一處跑,今天隊伍排不開,這些已經可以核。困難的規模和來路值得被看見,即使尚未認定任何一家越線。提前看,不等於提前處罰;它使後來要改變時,手裡不只剩一份驚人的估價。

可拆的準備,是在還不拆時給未來留條件。例如接口能按公開方式相連,算力與模型的帳仍能各自讀懂,哪一項停了由誰接續,預先安排並試過。這些不是三種同義詞:分帳不會自動使系統搬得走,接口公開也不自動保證有人肯接手。

準備要付代價。獨立接口可能慢一些,分帳多些管理工作,備用機器平時可能閒著。它們保住哪一種可能,日後少了什麼不確定,要逐項說明;不能把明天也許很貴的一場事故,拿來替今天每一筆支出證明合理。

一個歷史上的分離提醒人,長在一起不是自然界的不可能拆。AT&T在和解安排下,把本地電話運營與長途及設備等業務分開,1984年生效。這件事說明大規模結構分離做成過,不說明做成沒有代價,也不替所有整合位置選定同一種解法。

把鏡頭移到設想中的田地,這個問題會更實在。百工的機器由燈塔模型驅動,兩家辦田疇替農戶耕作。機器只認指定種子和肥料,別家的要認證。田疇說,規格是為精度。精度可能真護住收成,認證也可能成為別人種子進不來的門,要分別核。

十年裡糧價降了,傷亡少了,一些荒地重種。這些改進值得保留。與此同時,少了許多親手種過一季地的人,一次系統出錯使兩個縣收割停了九日,公共一側找不到接手的人和機器。效率與接續的薄弱,並不因看起來相反就不能同在。

這裡的輕重變了。作坊換助手費力,與糧食供給接不上,分量不同。對糧和水,公共安排不僅要約束經營者,也要確保供給能夠持續。要保的是人吃得到,不是哪家提供者無論怎樣都不能改變。田疇經營著王國州約七成耕地,仍有別的經營者;但這不證明田疇停擺以後,其他人就能接住全部需要。

模型、機器、種子、糧,一層層接起來,各有自己的理由與責任。機器上的成功,不自動給出種子上的排他許可;種子規格需要核,不使機器帶來的安全改進消失。這樣分開看,才不必在全盤拒絕技術和全盤接受綁定之間選一邊。

具體該拆、該限什麼,需要比較方案、證據和代價。整合不自帶罪名,困難也不自帶豁免。真正值得提前保住的,是後來仍能作選擇的能力:可以換、可以接、可以看清帳,必要時能把決定從原來那一雙手裡重新打開。

那支等了六個月的模型小組,未必要求自己永遠排第一。它需要知道怎樣排,為什麼排,哪裡能問,以及下一條路不是必須再由同一把鑰匙放行。從樓下看上去很方便的設計,走到樓上,還要容得下別人的門。

Source & reading note原论文与阅读说明原論文與閱讀說明

Nine essays draw on the Monopoly prequel, seven papers and dialogue with earlier thinkers. Kingdom State, Crossways, Beacon and future water provision are imagined, not disguised allegations or forecasts. Linked cases distinguish investigations, allegations, jurisdiction, liability and remedies within their stated institutional and dated scope. Size, success, defaults and investment do not alone establish a violation; benefits do not answer for every exercise of control. Separation, interoperability and standby capacity are discussed as proposals with costs, not proven universal solutions. This is not legal or investment advice. English is independently written.

九篇据《垄断论》前置篇、七篇正文与传统对话篇改写。王国州、通衢、灯塔及未来供水等场景为设想,不是现实企业的化名或对未来的预测。真实案件与制度材料另列来源,区分调查、指控、管辖判断、责任判决与救济;按相应机构和材料日期说明,不以一项法律结论代答全部规范问题。规模、成功、默认设置或投资关系本身不直接证明越线;有益的一面也不替具体控制作答。关于拆分、互通与备用能力的建议不是已经验证的万能方案,不提供法律或投资建议。简繁中文同文,英文独立重写。

九篇據《壟斷論》前置篇、七篇正文與傳統對話篇改寫。王國州、通衢、燈塔及未來供水等場景為設想,不是現實企業的化名或對未來的預測。真實案件與制度材料另列來源,區分調查、指控、管轄判斷、責任判決與救濟;按相應機構和材料日期說明,不以一項法律結論代答全部規範問題。規模、成功、默認設置或投資關係本身不直接證明越線;有益的一面也不替具體控制作答。關於拆分、互通與備用能力的建議不是已經驗證的萬能方案,不提供法律或投資建議。簡繁中文同文,英文獨立重寫。

Read the source paper ↗阅读原论文 ↗閱讀原論文 ↗

United States v. Microsoft: 2001 appellate opinion美国诉微软:2001年上诉法院判词原文美國訴微軟:2001年上訴法院判詞原文 ↗

CMA Nvidia/Arm case page: phase-one materials and cancellation英国CMA英伟达/Arm案页:第一阶段材料与调查取消英國CMA英偉達/Arm案頁:第一階段材料與調查取消 ↗

CMA phase-one competition assessment of the proposed Arm acquisitionCMA关于英伟达拟收购Arm的第一阶段竞争评估摘要CMA關於英偉達擬收購Arm的第一階段競爭評估摘要 ↗

Federal Judicial Center: AT&T settlement and the 1984 divestiture美国联邦司法中心:AT&T和解与1984年分离的司法史说明美國聯邦司法中心:AT&T和解與1984年分離的司法史說明 ↗