依其自身标准的观察 / An Observation by Its Own Standard

曹颖（Doris Tsao）团队的一篇预印本记录到：猕猴在双眼闪光抑制下观看两张面孔时，面孔区和前额叶皮层中的群体活动大约每 60 毫秒在表征一张面孔与表征另一张面孔之间切换一次。几小时之内，宣布这一发现的推文串便说起了“意识模式”和“无意识模式”，说这是迈向“产生意识觉知的配方”的一步；再经一次转发，它变成了“人类每秒在有意识与无意识状态之间交替 15 次”。这预设了一个公认的意识状态标准，而这样的标准并不存在；所能拿出的每一种标准——一份报告、一种相似、一个模板——都仍然是一种标准，没有一种就是它被用来取样的那个意识。被发现的，是一种约每秒被取样 15 次、按实验自身的标准被称为有意识的交替，随后这项观察便作为对一般意义上的意识的观察流传开去。曹颖本人撤回了“帧率”的读法。这种读法所搭乘的标签留了下来。

A preprint from Doris Tsao's group records that, while a macaque views two faces under binocular flash suppression, population activity in face patches and prefrontal cortex switches about every 60 ms between representing one face and representing the other. Within hours the thread announcing it spoke of a "conscious mode" and an "unconscious mode", and of a step toward "the recipe for generating conscious awareness"; one more repost turned it into "Humans alternate between conscious and unconscious states 15 times per second." That presumes an agreed standard of conscious states, and there is none; every standard on offer, a report, a likeness, a template, stays a standard, and none is the consciousness it is used to sample. What was found is an alternation sampled about 15 times per second, called conscious by the experiment's own standard, and the observation then travelled as an observation of consciousness in general. Tsao withdrew the frame-rate reading herself. The label it rode on stayed.

一、 登记下了什么 / 1. What Was Registered

记录来自两只恒河猴，Neuropixels 探针植入颞下皮层的面孔区 ML 与 AM，以及两个前额叶区域 PO 与 PVL。先向一只眼睛呈现一张启动面孔，然后在启动面孔仍留在屏幕上的同时，向另一只眼睛闪现一张不同的面孔；分析窗口为闪现后 300 至 1000 毫秒。一个双状态隐马尔可夫模型在时间上划分每个试次；在模型对留出神经元拟合最好的那些较小的时间分箱宽度上，中位驻留时长在 60 毫秒左右趋于平稳，以 5 毫秒分箱时两个状态分别为 64 和 59 毫秒。对一个状态取平均的活动重建出闪现的面孔，对另一个状态取平均的活动重建出启动面孔；当用一半神经元推断状态、用留出的另一半做重建时，这种分离依然成立。作者加上这项检验，是因为用同一份活动完成两步，用他们的话说，“会受到循环条件化偏差的影响”。第一个状态中对闪现面孔的调谐增益，略高于第二个状态中对启动面孔的增益。摘要把条件说得很明白：“在刺激与知觉恒定期间，大脑快速交替，约每秒 15 次，往返于有意识的微时段与被抑制的微时段之间。”

The recordings come from two rhesus macaques, with Neuropixels probes in the face patches ML and AM of inferotemporal cortex and in two prefrontal regions, PO and PVL. A priming face is shown to one eye, then a different face is flashed to the other while the first stays on screen; the analysis window runs from 300 to 1000 ms after the flash. A two-state hidden Markov model partitions each trial in time; across the small bin sizes where the model fits held-out neurons best, the median dwell plateaus around 60 ms, at 64 and 59 ms for the two states with 5 ms bins. Activity averaged over one state reconstructs the flashed face, activity averaged over the other reconstructs the priming face, and the separation survives when one half of the neurons infers the states and the held-out half is used for the reconstruction, a check the authors added because using the same activity for both steps is, in their words, "subject to a circular conditioning bias". The gain of tuning to the flashed face in the first state is slightly higher than the gain to the priming face in the second. The abstract states the condition plainly: "during a constant stimulus and percept, the brain rapidly alternates, about 15 times per second, between conscious and suppressed micro-epochs."

这作为测量是成立的。一个群体同时保留两个输入，轮流给每一个输入以单独呈现那张面孔时它所给的同一种特征编码，而显示画面，以及人类对这个画面报告的知觉，都保持不变：这本身就是一项发现，无需任何关于意识的字眼，就已有其价值。

That holds as measurement. A population that keeps both inputs, and gives each the same feature code it gives that face alone, in alternation, while the display, and the percept humans report for it, stay put, is a finding that needs no word about consciousness to be worth having.

二、 “有意识”从何而来 / 2. Where "Conscious" Comes From

猴子什么也没有报告。它们注视，是为了换取果汁。哪张面孔被知觉到，是在十四名人类受试者身上确立的：他们按键报告左眼、右眼或混合，多数时候报告的是闪现的面孔；猴子“观看的是在人类身上验证过的同样的闪光抑制显示”。所以，猕猴数据里的“有意识”，指的是人类对同一显示所报告的那张面孔，凭相似被带到另一种动物身上，再贴到其活动与那张面孔单独呈现时的反应相匹配的那个状态上，无论是哪一个。

The monkeys did not report anything. They fixated for juice. Which face was perceived was established in fourteen human participants, who pressed buttons for left eye, right eye or mixed, and mostly reported the flashed face; the monkeys "viewed the same flash suppression displays verified in humans". So "conscious" in the macaque data names the face that humans reported for the same display, carried to another animal by similarity, and then attached to whichever state's activity matches that face's solo response.

每一环都是研究双眼竞争的实验室赖以运转的那种工作标准：在一个心智里取得的报告，凭相似推及另一个心智的推断，在第三个系统内部做的模板匹配。没有一环就其用途而言是错的。合在一起，它们是一种标准，它们所登记的，仍是这种标准的登记。报告是被观察到的；报告背后的知觉是推断出来的，在人类身上如此，在动物身上也一样。这里推断又多走了一步，从人类的按键走到一只什么也没按的动物，而正是这一推断，让两个交替表征中的一个得以被称为有意识的。意识从不作为数据出现在数据之中追溯了这个一般形式：每一个数据到来时都已经在一种握持之下，而相关物不会变成它所作为条件的那个登记。

Each link is a working standard of the kind a rivalry lab runs on: a report taken in one Mind, an inference to another by likeness, a template match inside a third system. None of them is wrong for its purpose. Together they are a standard, and what they register stays that standard's registration. A report is observed; the percept behind it is inferred, in the humans as in the animal. Here the inference runs one step further, from the humans' presses to an animal that pressed nothing, and that inference is what lets one of two alternating representations be called conscious. Consciousness Never Appears as Data Among Data traces the general form: every datum arrives already under a hold, and a correlate does not become the registration it conditions.

三、 速率属于范式 / 3. The Rate Belongs to the Paradigm

观察到 15 Hz，并不排除其他频率。这个速率是这个范式所登记的：用一个专为找出两个状态而建的模型，在四个记录区域里，在一次迫使冲突的闪现之后 700 毫秒的窗口内。论文把它与竞争中的知觉切换并置，“比知觉切换（约 0.5 Hz）快一个数量级（15 Hz）”，这已经说明交替与知觉走在不同的时钟上。曹颖自己的后续说明又补充道，颞下皮层“在双眼竞争这种人为情境之外，可能根本不会交替”。

Observing 15 Hz does not rule out other frequencies. The rate is what this paradigm registers, with a model built to find two states, in four recorded regions, inside a 700 ms window after a flash that forces a conflict. The paper sets it beside perceptual switching in rivalry, "an order of magnitude faster (15 Hz) than perceptual switches (~0.5 Hz)", which already says the alternation and the percept run on different clocks. Tsao's own follow-up adds that inferotemporal cortex "may not even alternate outside of the artificial situation of binocular rivalry".

经由一个范式读出的速率，是这个范式从它被给予的输入中登记到的东西。把它当作意识本身的速率，当作这个范式的状态被赋予的那个名字的速率，就是从观察到主张的那一步；记录里没有任何东西迈出这一步。

A rate read through one paradigm is what that paradigm registers of the input it was given. Taking it for the rate of consciousness itself, the name the paradigm's states were given, is the step from observation to claim; nothing in the recordings takes that step.

四、 更正交还了速率，没有交还标签 / 4. The Correction Returns the Rate, Not the Label

推文串发出一天之后，曹颖写道：“我们并没有证明‘意识有帧率’。”她补充说，论文报告的结果“只来自少数几个选定的节点”，“我们仍然不知道意识通达是如何运作的——产生意识体验的真正配方”；一位读者把她的 TL;DR 引回给她，她回答说：“我本该加上在双眼竞争期间。”这把速率交还给了产生它的条件，而且是以她自己的声音交还的。

A day after the thread, Tsao wrote: "We have not shown that 'consciousness has a frame rate.'" She added that the paper reports results "from only a few select nodes", that "We still don't know how conscious access works--the actual recipe for generating conscious experience", and, answering a reader who quoted her TL;DR back to her, "I should have added during binocular rivalry". That returns the rate to the condition that produced it, and it does so in her own voice.

同一则更正保留了标签。它说，论文所显示的是一个“意识无法通达”的信号，两种模式与意识相关，“一个被授予了意识通达，另一个没有”。这些措辞和帧率一样依托于第二节的标准，并且说出这个标准的结果时却不带这个标准：措辞说的是“意识无法通达”，而记录呈现的，是一张人类对这一显示没有报告的面孔的表征。有一条回复给出了一个解围的说法：论文说的是医学意义上的意识，即界定昏迷的那种。那是其中一部分，而即便这一点也有争议：医学意义是又一种标准，即行为上的反应性，按这个标准，动物在两个状态中同样有反应，所以它分不出一个状态与另一个状态。

The same correction keeps the label. What the paper shows, it says, is a signal "that is inaccessible to consciousness", and the two patterns are related to consciousness "with one granted conscious access and the other not". Those phrases rest on the standard from section 2 as the frame rate did, and state its result without the standard: "inaccessible to consciousness" where the recordings show a representation of the face humans did not report for this display. One reply offered a rescue: the paper means consciousness in the medical sense, the one that defines coma. That is part of it, and even that is arguable: the medical sense is one more standard, behavioural responsiveness, and by it the animals were responsive through both states alike, so it cannot tell one state from the other.

一则更正也停留在很少有人看的地方：去找它的人才会读到它，而推文串与转发继续流传。不过，看到它的人有多少，并不划定边界，因为延伸就在它的正下方继续：一条直接贴在这则更正下面的回复问道，60 毫秒的驻留“是否很可能界定了意识体验被构建的时间颗粒度或采样频率”。承载一次延伸的，是每个读者从写下的文字里留下的东西。标签是实验室的说法，而读者留下的，是与自己已有的信念相合的东西；在那份信念关乎意识的地方，“有意识”与它相合，在双眼竞争期间这个限定则不相合。于是，边界始于范式的标准，在一则帖子里就速率得到恢复，到标签之前便停住了：只追加一句限定而不就地重切原有痕迹，早先的脚手架便仍作为基底留在原地，这正是你的代码库即是记忆所追溯的机制；因此帧率被撤回，“有意识”仍然立着，越过更正长出来的东西，是从标签长出来的。

A correction also stays where few look: it is read by those who go looking, while the thread and the repost travel on. How many saw it does not set the boundary, though, since the extension goes on directly under it: a reply posted under the correction itself asks whether the 60 ms dwell "may well define the temporal grain or sampling frequency of how conscious experience is constructed". What carries an extension is what each reader keeps from what was written. The label is how the lab speaks, and a reader keeps what fits what they already believe; where that belief is about consciousness, "conscious" fits it, and the clause during binocular rivalry does not. So the boundary starts at the paradigm's standard, is restored for the rate in one post, and stops short of the label: appending a caveat without re-cutting the trace in place leaves the earlier scaffolding standing as ground, the mechanism Your Codebase Is Memory traces, so the frame rate was withdrawn, "conscious" was left standing, and what grows past the correction grows from the label.

意识向量论文的真正教训追溯了一个姊妹案例：在那里，一个可操控的方向在被测量之后，被冠以那个现象的名字。这里，哪张面孔算作有意识的，是在记录之前，由范式据以验证的人类报告确定的；与那张面孔相匹配的状态，在测量之后继承了这个名字。

The Real Lesson from the Consciousness Vector Paper traces a sister case, where a steerable direction was named after the phenomenon after it was measured. Here which face counts as conscious is fixed before the recordings, by the human reports the paradigm was verified on; the state that matches that face inherits the name after the measurement.

五、 仍然可用的部分 / 5. What Stays Usable

放回它的标准之中，这一发现什么也不失去。颞下与前额叶的群体在两个刺激表征之间约每 60 毫秒交替一次，而人类对同一显示报告的知觉保持恒定，未被报告的那张面孔的表征丰富到足以重建出那张面孔。这是一个关于群体如何承载相互竞争的输入的精确结果，也正是数据所支持的结果。

Put back in its standard, the finding loses nothing. Inferotemporal and prefrontal populations alternate between two stimulus representations about every 60 ms while the percept that humans report for the same display stays constant, and the representation of the unreported face is rich enough to rebuild that face. That is a sharp result about how a population carries competing inputs, and it is the result the data support.

从摘要到转发之间的每一步，都扩大了这项观察看起来所显示的东西，而每一次扩大，靠的都是丢掉造就这项观察的那个标准中的一个限定。任何框架被带过它被校准的范围，都是如此，框架的边界追溯的正是这一步。回归付出一个限定的代价——“在双眼闪光抑制下，按这个范式的标准”——并让这项观察始终可以追溯到获得它的条件。

Each step between the abstract and the repost widened what the observation appears to show, and each widening worked by dropping a clause of the standard that made the observation. The same holds for any frame taken past where it was calibrated, the move traced in The Boundary of the Frame. The return costs one clause, under binocular flash suppression, by this paradigm's standard, and keeps the observation traceable to the conditions under which it was obtained.