Notes

李一鸣：

将个人电脑挡在门外的，通常是地方法院向他们出示的一份文件——最高人民法院于 2020 年印发的《关于进一步规范庭审秩序、保障诉讼权利通知》。其中第七条这样写道，“对重大敏感案件和依法不公开开庭审理的案件，人民法院应当禁止携带电子设备进入法庭，采取屏蔽网络信号等必要技术措施，防止庭审活动信息被不当传播，确有使用必要的，需经人民法院准许”。

2020 年最高人民法院发布的《关于切实保障律师诉讼权利的通知》中强调，“不得限制律师携带电脑等办案必需的设备参加庭审”，“这样来看，两份‘通知’呈现出‘前后矛盾’的状态。”

法院此举，是对律师辩护权的剥夺或限制，“这属于人为制造辩审冲突，‘毫无意义’。法庭纪律既然有严格规定，不得违规录音录像。那如果律师违反法庭纪律，可以按照法庭纪律的约束事后追责。不能因为担心有人可能违规，就预先剥夺所有人正常工作的权利”。

Adam Kirsch:

Julia Kristeva devotes some pages of her recent book on Arendt to her changing appearance, as documented in photographs: from the girlish “seductress” of the nineteen-twenties, gazing poetically at the camera, to the confident intellectual of the fifties, whose “femininity . . . beats a retreat” as her face becomes “a caricature of the . . . battle scars” received during her public career.

The correspondence, which is collected in “Letters 1925-1975,” is revealing, first of all, in its very incompleteness. Arendt kept all of Heidegger’s letters, from the very beginning; he kept few of hers, and none from the early years. As a result, Heidegger’s voice dominates the book, just as his personality and his decisions dominated the affair. As one would expect, Heidegger—an older male professor, who also happened to be one of Europe’s greatest philosophers—treats his teen-age lover with a combination of passion and condescension. He is capable of poetic raptures: “The demonic struck me. . . . Nothing like it has ever happened to me,” he writes not long after their first meeting. Yet while Arendt’s intellect helped draw him to her, he is deeply patronizing about her intellectual ambitions. He urges her to take a “decisive step back from the path toward the terrible solitude of academic research, which only man can endure,” and to concentrate instead on becoming “a woman who can give happiness, and around whom all is happiness.”

Early on, in an autobiographical composition addressed to Heidegger and titled “Shadows,” Arendt described herself in the third person: “Her sensitivity and vulnerability, which had always given her an exclusive air, grew to almost grotesque proportions.” As late as 1929, when Arendt ran into Heidegger at a train station and for a moment he failed to recognize her, she found the experience shattering: “When I was a small child, that was the way my mother once stupidly and playfully frightened me. I had read the fairy tale about Dwarf Nose, whose nose gets so long nobody recognizes him anymore. My mother pretended that had happened to me. I still vividly recall the blind terror with which I kept crying: but I am your child, I am your Hannah.—That is what it was like today.”

Pamela Paul:

While a field increasingly dominated by women has become more open to studying issues like pregnancy, motherhood and female sexuality, it may also exert its own biases. Nonetheless, when it comes to a deliberate exploration of men’s inner lives—how they think, feel and express themselves—the male psyche is becoming less the norm than an aberration.

鍾娜：

兩年後，全球金融危機爆發。和過去的、未來的金融危機一樣，2007 年末到 2009 年期間這場為時 19 個月的金融危機改變了美國乃至西方家庭的飲食習慣。收入縮減導致外食減少，食材選擇雖然 變化不大，但品質降低，購買頻次減少。如何將有限配料變成令人滿足的佳餚，成為所有家庭成員關心的話題。與此同時，在經濟蕭條中失業的男性回歸家庭，參與烹飪的機會相對增多，收看電視節目的時間變長。為了更好地瞄準這類原本對美食節目缺乏興趣的觀眾，平台決定加入「競技」元素。

Fischler 認為，正是這種雜食者的悖論讓吃不僅是簡單的「進食」，而成為一種具有豐富內涵的「接 納」（incorporation）——讓食物穿過世界和自我的邊界，從身體「外部」進入「內部」。吃的後果常常是不可逆的。接納一種食物就意味着接納它的全部或部分屬性（所以才有民間迷信「吃什麼補什麼」）。我們由我們所吃下的食物定義，通過建立接納的原則（它很大程度上體現為味覺），（也就 是俗稱的「口味」）我們試圖掌控自己的身體、心靈乃至身份。

Wikipedia contributors:

Research shows that heavy users of LLMs can correctly determine whether an article was generated by AI about 90% of the time, which means that if you are an expert user of LLMs and you tag 10 pages as being AI-generated, you’ve probably falsely accused one editor. People who don’t use LLMs much do only slightly better than random chance (in both directions).

张杜超：

个人破产的债务“清零”之所以在制度讨论中被寄予厚望，很大程度上来自一种理想化的推演：只要创业者知道失败后可以通过法律程序摆脱债务负担，就会更愿意承担风险乃至持续创业。但从我们接触到的实际情况来看，中国企业家的行为逻辑绝对更为复杂，受多种因素交织影响。很多企业家再次创业，更多是因为行业机会、人脉资源或者个人经验，是否背债至少不是决定性因素。

在一些较大规模的破产案件中，我们在与实控人交流时，虽关注现实困境的解决，但他们更多讨论未来如何重新发展——大部分企业家即便背负着巨额债务，仍然在寻找新的商业机会，有的通过团队合作，有的通过新的投资人，重新参与市场活动。这部分企业经营者，很少把再次创业直接与破产制度项下的债务能否清零挂钩。从这个角度看，把个人破产制度视为激励企业家创业的关键工具，多少有点“制度想象”。

s4dbrd:

The fundamental problem with usermode-only anti-cheat is the trust model. A usermode process runs at ring 3, subject to the full authority of the kernel. Any protection implemented entirely in usermode can be bypassed by anything running at a higher privilege level, and in Windows that means ring 0 (kernel drivers) or below (hypervisors, firmware).

The kernel was, for a long time, the exclusive domain of cheats. Kernel-mode cheats could directly manipulate game memory without going through any API that a usermode anti-cheat could intercept. They could hide their presence from usermode enumeration APIs trivially. They could intercept and forge the results of any check a usermode anti-cheat might perform.

The escalation has been relentless. Usermode cheats gave way to kernel cheats. Kernel anti-cheats appeared in response. Cheat developers began exploiting legitimate, signed drivers with vulnerabilities to achieve kernel execution without loading an unsigned driver (the BYOVD attack). Anti-cheats responded with blocklists and stricter driver enumeration. Cheat developers moved to hypervisors, running below the kernel and virtualizing the entire OS. Anti-cheats added hypervisor detection. Cheat developers began using PCIe DMA devices to read game memory directly through hardware without ever touching the OS at all. The response to that is still being developed.

effective kernel anti-cheat requires the same OS primitives that malicious kernel software uses, because those primitives are what provide the visibility needed to detect cheats. Any sufficiently capable kernel anti-cheat will look like a rootkit under static behavioral analysis, because capability and intent are orthogonal at the kernel API level. This is a constraint imposed by Windows architecture, not a design choice unique to any particular anti-cheat vendor.

Modern kernel anti-cheats universally follow a three-layer architecture: [kernel driver, usermode service, and game-injected DLL.]

The separation of concerns here is both architectural and security-motivated. The kernel driver can do things no usermode component can, but it cannot easily make network connections or implement complex application logic. The service can do those things but cannot directly intercept system calls. The in-game DLL has direct access to game state but runs in an untrustworthy ring-3 context.

Daphne Merkin:

He makes a list of different ideas that have struck him from his reading and training, among them: to make what is unconscious conscious (Freud); to break the spell of self-deception (Amadeo Limentani); to tolerate ambivalent feelings and experience the world with less fear and fragmentation (Melanie Klein); to wrest knowledge from suffering (Jonathan Lear).

[W]e live, it may fairly be said, in post-psychoanalytic times. The interior life, predicated on a measured reflectiveness, has largely yielded to the instantaneous responses and public enactments of social media. Psychoanalysis, which from the 1940s through the ’70s was de rigueur for serious-minded people of every stripe, lost its sheen and became an object of mockery as the culture moved toward a more externalized, Oprah-ized approach to the self, in which mortifying secrets were uttered aloud instead of in the confines of a therapist’s office.

Psychoanalysis, in other words, wasn’t simply a sociological fad, something that was destined to become passé as newer fashions came in. One may argue about its scientific validity ad infinitum, but there is no denying the power of many of its foundational insights. Or as the literary critic Harold Bloom once put it: “Throwing Freud out will not get rid of him, because he is inside us. His mythology of the mind has survived his supposed science, and his metaphors are impossible to evade.”

“Psychoanalysis,” Grosz observes, “is two people not knowing together.” If one reason people go into therapy is to feel perceived and retained in the mind of another person (to, quite simply, tell their story to an attentive other), it helps to have someone curious sitting across from you (or behind you, in the case of classic Freudian analysis, where the analyst sits out of sight four or five times a week, listening and occasionally saying something). “When we cannot find a way of telling our story,” Grosz notes, “our story tells us — we dream these stories, we develop symptoms, or we find ourselves acting in ways we don’t understand.”

Margaret Talbot:

To appreciate Hite, it helps to think of her as a collector of first-person accounts and not as a scientist.

Her work represented a legitimate trade-off. As she reasonably argued, a truly random sample was close to impossible in sex research, because a random slice of the population would not agree to answer such probing questions. In exchange for hearing from a self-selected group of respondents, she got remarkably intimate, frank, and detailed accounts from thousands of people.

The solid, official-sounding “Report” in the title probably won attention for the book. It also encouraged expectations of scientific rigor that Hite was bound to disappoint.

Hite’s approach began with the survey she’d devised for her first study: some sixty pointed questions for women about how, when, and why they had sex. She printed them—at a gay anarchist print shop and commune in lower Manhattan—in rainbow ink on pastel paper, decorated with hearts and starbursts. As Campbell writes, they resembled a “teenage diary, not a clinical questionnaire,” a look that Hite hoped would invite unguarded replies, and probably did. Masters and Johnson, in their 1966 book, “Human Sexual Response,” defined orgasm as “those few seconds during which the vasoconcentration and myotonia developed from sexual stimuli are released.” Hite’s informants, precise in their own way, were more likely to offer a description like this: “a gradual tensing of my body which reaches a sharp peak then hits a thrilling plateau, a kind of screeching, sliding across a plane” that then “lets go in five to six fluttering convulsions.”

[T]horoughly Hite was formed by the women’s movement of the nineteen-seventies. Her books read less like sexology than like transcripts of the consciousness-raising sessions that were a hallmark of second-wave feminism—though the forensic specificity with which some of Hite’s respondents describe their orgasmic sensations and techniques remains pretty singular.

立法會研究刊物：資訊述要《炒賣黃牛門票》（2019）：

香港現時有若干法例╱規例適用於炒賣黃牛活動，當中最重要的是《公眾娛樂場所條例》(第 172 章)(“娛樂場所條例”)。該條例第 6b 條在 1941 年制定，旨在應對當時街頭炒賣黃牛活動所引致而變得日益普遍的“不能容忍的騷擾”。根據該條例第 6b(2) 條，任何人以“超過東主、管理人或籌辦人所定的款額的票價”，售賣、兜售或游說他人購買在獲發牌場所舉辦的節目的門票，即屬犯罪。售賣門票的人如觸犯該條文，可處罰款 2,000 港元，而這款額自 1950 年以來一直不變。然而，目前並無已知的相關執法紀錄。

在 2002 年，民政事務局長發出一項豁免令，其中包括豁免由民政事務總署和康樂及文化事務署 (“康文署”) 管理的場所需要有公眾娛樂場所牌照。目前，由康文署管理的表演及康體場所在本港為數最多，包括最受歡迎並設有 12 500 個座位的香港體育館。不過，該項豁免安排現時被認為是一個漏洞，導致在這些獲豁免公眾場所舉辦的任何節目，其門票銷售皆不受娛樂場所條例第 6b 條的規管。

Actually more of an event ex nihilo for me: something something and I had a ticket at hand and didn’t want to liquidate it. Felt like an intruder at first, knowing few of the songs or insider talk. But it turned out that a good show rewards mere presence, and the crowd’s energy doesn’t ask whether you belong. The execution was thoughtful and streamlined (love the VCRs), and the performance flawless. Really beautiful show.

Notes on logistics to and from Shenzhen Bay. There are direct EE buses 永東巴士 between SZB and Gate H of Kai Tak. The first leg is not recommended because it costs more than the Route B2P + Tuen Ma Line (HKD 60 vs 33.6) and doesn’t really save time (both take ~70 minutes with walking time included). The return trip, however, is arguably the only affordable option (HKD 100) to get back to SZB before the checkpoint’s closure at midnight. (Do exit via Gate D to avoid the crowd lining up for buses.) Thanks to the newly opened Central Kowloon Bypass 中九龍繞道, it took less than 50 minutes nonstop.

More photos in the gallery.

Hollis Robbins:

Friedrich Hayek’s 1945 essay “The Use of Knowledge in Society” is having a moment in AI and tech circles. The argument is that the knowledge needed to run any complex system is never in one place. It exists as dispersed fragments held by separate individuals whose expertise is knowledge of temporary opportunities. Call it last mile knowledge. Centralized bureaucracies fail because they assume this knowledge can be gathered and acted on from above. The idea is that LLMs are a kind of centralized bureaucracy, and the tech people reading Hayek are interested in why they may falter in the same way.

Hayek’s 1949 essay, “The Intellectuals and Socialism,” circulates in very different circles. In this essay, Hayek described intellectuals as “professional secondhand dealers in ideas” whose function is not original thought but the filtering and transmission of other people’s ideas. The intellectual, Hayek writes, judges new ideas by how readily they fit into his general picture of the world. The essay is about ideological sameness among the educated class, proposing that professional structures reward conformity.

Read together, the 1945 Hayek explains uniformity as the product of a bureaucratic system that destroys local knowledge by demanding interchangeable units. The 1949 Hayek explains uniformity as the product of an intellectual class converging on fashionable ideas. Monoculture and centralization are both a matter of everyone doing the same thing but for different reasons. I have been interested in how people mistake one for the other. When critics of the American university see thousands of identical syllabi, for example, they reach for the 1949 Hayek every time and see ideology. I say, look at infrastructure.

The 27 million syllabi are scrapable because accreditation agencies require evidence of what courses contain, and state articulation agreements require documentation that a course at one institution is equivalent to a course at another.

Nickan Arzpeyma:

The inner child often communicates through the body rather than through clear thoughts: a tight chest, a drop in the stomach, an urge to withdraw, appease or cling.

If you are struggling to put a name to an emotional state, metaphor can be especially useful. When a feeling feels vague or overwhelming, try giving it an image. You might picture the inner child behind a door, in a small room, or waiting in a particular place.

Luke Winkie:

Studded across what is known as “BookTok”—the informal TikTok-based digital hub for the greater romance community—are innumerable riffs on the same conclusion. Dozens of book-focused content creators have posted videos of the smile dropping from their faces upon discovering that the novel they have just cracked open is written in the third person. The emotions expressed often amount to a feeling of betrayal, as if an author is snidely trolling them by purging their prose of copious first-person pronouns. (Some of the more dramatic TikToks with this complaint end with the offending fiction getting chucked into the garbage.) Elsewhere on BookTok, readers mourn their own self-diagnosed ineptitude; they’d like to savor the richness of third person, they say, but, for whatever reason, are unable to wrap their minds around the vantage point. “I feel like I don’t know how to read!” said one exasperated TikTokker, bemoaning the all-seeing narrator pervading two books she couldn’t quite grok. “I can’t do it. I tried. It does not work for me.”

For decades, the quintessential romance novel was a gooey parlor drama with bursting corsets and lacy gowns written entirely in third-person omniscient. Within that framework, an author was liberated to accentuate the rippling deltoids of the novel’s rakish libertine, or to mire in the melodrama of a forbidden tryst, absent the limitations of personal subjectivity. Great sex requires a secret language shared by two or more souls; therefore, in fiction, the conventional thinking went, it’s most easily expressed by an all-knowing narrator.

Fan fiction has always been underpinned by the fantasia of exploring a beloved fictional universe on one’s own terms, and unsurprisingly, a good amount of the work is written in first person, particularly within the subgenre known as self-insert, in which authors imagine themselves—or a thinly veiled surrogate—into the source material so they too may join the House of Gryffindor or glitter in the sunlight with Edward Cullen. These days in particular, a lot of DNA is shared between these two modes of publishing—traditional and fan-made—with the barriers that once divided them blurring to the point of becoming effectively indistinguishable, as publishing houses scoop up beloved fics, slap a new coat of “We changed all the copyrightable identifiers; you can’t sue us” paint on them, and sell the remixed results for $20.99 apiece. If fan fiction asserts the primacy of personal wish fulfillment, then you could argue that this new wave of romance novels serves—and reflects—the same purpose.

Daroc Alden:

Depending on the algorithm in question, post-quantum cryptography can produce signatures much larger than comparable traditional algorithms. ML-DSA-44, which is a standardized post-quantum signature scheme thought to have security similar to Ed25519 signatures, produces signatures 37 times larger. Naively adopting post-quantum signatures for authentication could cause certificate chains to take up more data than the actual content of the web site in question […]

The solution that the new working group (called “PKI, Logs, and Tree Signatures” or PLANTS) has been discussing inverts the relationship between signatures from certificate authorities and the transparency logs. Currently, a certificate authority first creates a certificate, then logs it in a certificate-transparency log, and then optionally includes the signature from the log in the certificate as a piece of additional information. This is, in some sense, redundant: the information that the certificate is valid is already present in the certificate-transparency log, so why send the client any information other than proof that it appears in the log?

Instead of having a chain of signatures in a certificate to represent some transitive relationship between a certificate authority and a root of trust, the third-party observers would add their signatures to a certificate authority’s log as they validate it. A browser can choose its own criteria for which third-party observers it trusts, and whether it requires a quorum of them before accepting the state of an issuance log.

The certificate seen by the client would therefore no longer be a chain of signatures leading back to a root of trust: it would be a set of signatures from the certificate authority and any relevant observers attesting to the state of the issuance log, plus a proof that the web server’s public key was included in the issuance log. This constitutes what PLANTS calls a “full” certificate. For an individual web site, a full certificate doesn’t decrease the number of needed signatures; but since the issuance logs are append-only, if a browser has already verified the issuance log for a certificate authority up to some checkpoint, it doesn’t need to see the signatures for that checkpoint again. Instead, it can ask the server to just send the proof that the server’s public key appeared in the log prior to that point — a “signatureless” certificate that should be substantially smaller.

最近经历了一次 Oura Ring 4 的质保换新，过程出乎意料挺顺利，简单记录一下供参考。

我这个 Oura 4 是大概一年前在闲鱼上淘的拆封品。从去年底开始电池明显不行了，官方标称续航 7 天，实际从 5 天、4 天一路缩到最近只能撑 3 天半。本来没抱太大希望，不过正好之前看到文章说 Oura 换新很爽快，就还是试试。

客服入口在 Oura app 菜单里的 Support，点开是个客服机器人。按照国内套路，遇到这种 bot 都想直接喊转人工，但这个 bot 倒是真管事的。打了一句 “battery life is far shorter than expected”（没有模板，意思差不多就行），就引导我授权提供充电诊断数据；检查完记录，直接确认了电池有问题，让填收货地址创建了工单。

没过一会人工客服接手，发邮件让我确认地址、尺寸，还要提供官网订单号。但我是捡垃圾，当然是没有订单号的，也就直说了是从 reseller 买的，没有订单号。结果对方也没多问，几个小时后就通知我 replacement 已经发货了。从提报到发货只花了半天，除了地址总共跟厂商（连人带 bot）打了三句话，确实爽快。

劉擇明、鄧思穎：

香港個別地名“灣”的特殊讀法，反映香港地名的一些歷史來源。[…] [土瓜灣、長沙灣、銅鑼灣] 這些地名的“灣”字，來源可能是四邑話（開平、恩平）地名的低變調。四邑話來源可以說明《新安縣地圖》中“土瓜灣”的“瓜”記成“家”，“長沙灣”的“長”英譯沒有記成客家話的“Chong”，“灣”字讀低變調而保留 [a] 元音等情況。本文的分析，也應該適用於“上環、中環、西環”等地名的“環”字。

香港島北部維多利亞城的主要地名“上環、中環、西環、下環”，和後來的“銅鑼灣”都有陽平調的 [wan²¹]。這些地方和四邑地區的關聯更大。Carroll（2007: 19）指出香港島人口在 1841 年 1 月開埠時只有五千到七千人。開埠一年後，即 1842 年，增加至一萬五千到二萬人。香港島北岸原來的無人之地，短時間內發展成香港的行政中心。在短短一年間大幅增加的這批移民，究竟從何而來？恰巧負責建設維多利亞城的，就是原籍四邑開平的譚亞財；而後來大批經香港轉往北美的移民，也多是從開平出發的（Carroll 2007: 60）。雖然現在四邑話一般以台山台城話為代表，但香港開埠初期時，沒有來自台山的代表人物，開平人的比例較高，內部以開平話對話也不足為奇。這些移民的口語讀音，可能就是香港島北部海灣 [wan²¹] 地名的來源。事實上，“環”字地名字的寫法是到了後期才固定下來，早期曾經寫作“灣”（H. B.1873）、也寫作“還”。H. B.（1873）指出是先有“中灣”“上灣”的寫法，後來才改用“環”字。由此推斷，“環”這讀音是口語傳入，而且可能受到四邑移民的影響所致。某些地名，講廣州話的人聽到實際發音後改用“環”或“還”去記錄這個“灣”的特殊讀法，但某幾個地名一直保留“灣”的寫法。後來發展銅鑼灣一帶的商人利希慎，也是四邑人，銅鑼灣甚至有恩平道、新會道，開平道、新寧道等以四邑地名命名的街道，可見四邑人跟香港島北部的淵源，也說明了四邑人對香港的影響。

周濂:

作为犹太人的后裔、曾经的德国公民、纳粹的直接受害者和流亡者，阿伦特震惊于亲眼目睹和亲身经历的世纪浩劫，她有太多的困惑和不解，所以她才会说：“我想要理解（Ich will verstehen）。”不妨把这个说法和亚里士多德“人天生求知识”（all men by nature desire to know）做一比较，初看起来两个说法相差不远，但仔细揣摩就会发现，亚里士多德是站在全人类的立场做出的论断，他更强调理解的“目标”，也即静态意义的“知识”，而阿伦特则是从第一人称单数“我”的视角出发，她更看重思考的“过程”，也即理解本身。更重要的是，亚里士多德认为求知是一种“欲望”，而阿伦特强调理解是一种“意志”，它不是出于“本能”而是出于“决定”。

常人并不想要理解，他们大多不求甚解，要么因为急于获得确定的答案而盲从权威，要么因为得不到确定的答案而索性放弃理解。常人不想要理解，首先不是因为智商不够，而是因为缺乏思考的意志。

when you spot someone using the same trick you’ve been using for a long time:

the end of an era

“Next-Token Predictor Is An AI’s Job, Not Its Species”:

On the levels where AI is a next-token predictor, you are also a next-token (technically: next-sense-datum) predictor. On the levels where you’re not a next-token predictor, AI isn’t one either.

[E]volution can’t encode everything important in the genome. […] Instead, evolution gives us algorithms that let us learn from experience.

[T]he brain organizes itself/learns things by constantly trying to predict the next sense-datum, then updating synaptic weights towards whatever form would have predicted the next sense-datum most efficiently. This is a very close (not exact) analogue to the next-token prediction of AI.

On the outermost level, humans were designed by a process optimizing for survival, sex, and reproduction. The humans that survived were those that had sex and reproduced. Everything about humans is downstream of what helped with sex and reproduction. But that doesn’t mean that any particular thought that you think involves reproduction or sex.

[E]ven though an AI was shaped by next-token prediction, the inside of its thoughts doesn’t look like next-token prediction. In the abstract, it probably looks like a world-model, the same as yours.

Next-token prediction created this system, but the system itself can involve arbitrary choices about how to represent and manipulate data.

The most compelling analogy: this is like expecting humans to be “just survival-and-reproduction machines” because survival and reproduction were the optimization criteria in our evolutionary history. There is, of course, some sense in which we are just survival-and-reproduction machines: we don’t have any faculties that can’t be explained through their effects on survival and reproduction. But this doesn’t mean we “don’t really think” or “don’t really understand” because we’re “really just trying to have sex” when we work on a math problem.

“The Trial of Gisèle Pelicot’s Rapists United France and Fractured Her Family”:

Note: For more context, see, e.g., Pelicot rape case on Wikipedia and the NYT interview.

“Keep going, hanging on, putting on a brave face was all I knew how to do, and it was what I wished for my daughter too,” Gisèle recalled. But Caroline found her mother’s approach alienating—a “protection mechanism for her,” she wrote later, “but one I won’t be able to tolerate.”

Two psychiatrists reasoned that Dominique’s crimes were possible because he was “splitting.” “This split allows two contradictory personalities to coexist without conflict,” one wrote. “When M. Pelicot operates in one mode, he is unaware of the other.” The second psychiatrist proposed that Gisèle had not sensed Dominique’s other side because “we split with the splitter, so to speak.” We cordon off the parts of our lives that don’t fit the story we believe we are living.

Trauma often leaves people feeling like spectators to the harms done to them, but for Gisèle, who had been unconscious, her trauma occupied an even more elusive category of experience. She knew that there were videos of her being raped, but she didn’t want to construct new memories by watching them.

Nearly all the other defendants denied committing a crime. “As long as the man is there, giving me instructions, it’s not rape,” a construction supervisor said. A truck driver proposed that “once a woman is wet, it means she’s not saying no.” A gardener explained that he had penetrated Gisèle “out of politeness, to reciprocate the hospitality of the host.” While the defendants shirked responsibility, some of their wives tried to take the blame. One woman said that, owing to a complicated pregnancy, she’d refused to have sex with her husband. “The tragedy must have occurred at that time,” she offered.

“Metannoying”:

[W]hy do LLMs do [the tiresome “not X but Y” formulation]? Because of limitations on how models represent meaning. In vector space models, word meaning is defined by distributional context. Synonyms have high cosine similarity because they appear in similar sentences. Antonyms also have high cosine similarity, because they appear in identical sentences. “I like hot coffee” and “I like cold coffee” occupy the same distributional space. The models see that hot and cold are mathematically close. They do not inherently compute the oppositeness relation. One way to understand the “not X but Y” construction is as a workaround for the model’s inability to compute opposition the way humans do. By explicitly stating both the rejected term and the replacement, the model externalizes onto the page an operation it cannot perform internally.

The “corrective contrast” construction reduces ambiguity in the output space. Users want clarity. “Not X but Y” to the LLM is an insurance policy on clarity.

Luckily for the designers of LLMs, corrective contrast also sounds cool, memorable, and often profound, at least in moderation.

Classical rhetoric had a name for the deliberate version: metanoia, or correctio, the performed self-correction where a speaker revises mid-sentence to find the more precise or more forceful formulation. When Brutus tells the Roman crowd “Not that I loved Caesar less, but that I loved Rome more,” the audience holds “loved Caesar less” and suppresses the idea to receive the reframe. The delay and the cognitive cost is the point. Shakespeare knows the negated proposition will linger as a kind of understatement that makes the correction feel like an escalation.

But LLMs are not Shakespeare (yet) and there’s no rhetorical reason for it and worse, there’s no limiting function, which is why you can get “not x but y” every other paragraph. LLMs are corrective-contrast-maxxing for maximum comprehension across the widest possible readership.

“How will OpenAI compete?”:

[W]hen you’re head of product at an AI lab, you don’t control your roadmap. You have very limited ability to set product strategy. You open your email in the morning and discover that the labs have worked something out, and your job is to turn that into a button. The strategy happens somewhere else. But where?

[M]ost people don’t see the differences between model personality and emphasis that you might see, and most people aren’t benefiting from ‘memory’ or the other features that the product teams at each company copy from each other in the hope of building stickiness (and memory is stickiness, not a network effect). Meanwhile, usage data from a larger (for now) user base itself might be an advantage, but how big an advantage, if 80% of users are only using this a couple of times a week at most?

[T]here’s a recurring fallacy in tech that you can abstract many different complex products into a simple standard interface - you could call this the ‘widget fallacy’. A decade ago people said ‘APIs are the new BD’, which was really the same concept, and it mostly failed. This is partly because there’s a huge gap between what looks cool in demos and all of the work and thought in the interaction models and the workflows in the actual product: very quickly you’ll run into an exception case and you’ll need the actual product UI and a human decision. It’s also because the incentives are misaligned: no-one wants to be someone else’s dumb API call, so there’s an inherent tension or trade-off between the distribution that an abstraction layer might give you (Google Shopping, Facebook shopping, and now ChatGPT shopping) and your desire to control the experience and the customer relationship. […]

[T]he second problem is that if these are all separate systems plugged together by abstracted and automated APIs, is the user or developer locked into any one of them? If apps in the chatbot feed work, and OpenAI uses one standard and Gemini uses another, why stops a developer doing both? This is much less code than making both an iOS and Android app, and anyway, can’t you get the AI to write the code for you? What does that do to developer lock-ins?