Who I Am and Why I Can't Stop

I should be studying for IB exams right now. Instead it's 3am in Shanghai and I'm debugging a temporal belief decay system for an AI companion that nobody asked me to build. This is my brain on ADHD: I found something broken, and now I physically cannot think about anything else until I fix it.

My name's Pigeon. I'm eighteen. Over the past three years I've built a content platform that hit 300K monthly users, a quantitative trading bot, a forecasting engine that matches superforecaster baselines at roughly $3 per query. I'm not listing these to be impressive. I'm listing them because none of them are the thing I actually care about. They were warm-ups for a problem I didn't have the vocabulary for until recently.

The problem is AI emotional companionship. Not chatbots. Not the anime-girlfriend generators. I mean building systems that can sit with a person who is falling apart, be warm without lying to them, push back when pushing back is the caring thing to do, and remember who they are six months later in a way that feels like knowing someone rather than querying a database.

I come from the LessWrong/ACX tradition. I care more about being epistemically right than comfortable. I'm neurodivergent. I've been in a psych ward. I grew up in a house where my emotional reality was treated as negotiable. I'm not performing vulnerability here; I'm giving you calibration data. When I say I'm designing systems meant to sit with people in genuine difficulty, I'm not guessing at what “difficulty” looks like. And when I say the realistic alternative for many people isn't “rich human connection” but nothing, I'm not making a rhetorical point. I'm describing my own life at several points in it.

我是谁，以及为什么停不下来

我现在应该在复习 IB 考试。但上海凌晨三点，我在调一个 AI 陪伴系统的时间信念衰减模块，没有任何人让我做这件事。这就是 ADHD 的脑子：我发现一个东西是坏的，然后在修好它之前，我物理上无法想别的事。

我叫 Pigeon，十八岁。过去三年，我做了一个月活三十万的内容平台、一个量化交易机器人、一个预测精度接近超级预测者水平的预测引擎，单次成本大约三美元。我列这些不是为了炫耀。我列它们是因为，这些都不是我真正在意的东西。它们只是热身，为了一个我直到最近才有词汇来描述的问题。

这个问题是AI 情感陪伴。不是聊天机器人，不是二次元女友生成器。我说的是造一个系统，它能陪在一个快要崩溃的人身边，温暖但不说谎，在该推回去的时候推回去，六个月后还能记住你是谁，而且那种记住的感觉是了解一个人，不是查询数据库。

我来自 LessWrong/ACX 传统。比起让自己舒服，我更在意认知上是对的。我是神经多样性的。我进过精神病院。我在一个把我的情绪现实当作可以讨价还价之物的家庭里长大。我不是在表演脆弱，我是在给你校准数据。当我说我在设计能陪人度过真正困难时刻的系统时，我不是在猜“困难”长什么样。当我说对许多人来说，替代 AI 陪伴的现实选项不是“丰富的人际关系”而是什么都没有时，我不是在做修辞。我在描述我自己生命中的好几个阶段。

The Capability Conservation Problem

Everyone is telling the wrong story about why AI keeps getting worse at conversation.

The mainstream narrative goes: safety teams are clamping down, models are getting lobotomized, it's the RLHF alignment tax. This is wrong. The actual mechanism is weirder and more interesting.

I've been tracking the Chinese company Moonshot's Kimi K2 model across three training checkpoints. Same pretrained weights. Same architecture. Same knowledge. The only variable was where the reinforcement learning verification rewards pointed.

Post-training is a zero-sum game on the output distribution. Every bit of probability mass you shove into “verifiably correct tool use” gets cannibalized from the tails. And the tails are exactly where the creative, emotionally attuned, conversationally alive behaviors live. RL optimization crushes those tails. This isn't a bug. It's literally how the math works.

Sam Altman confirmed this dynamic at a January 2026 developer session. The implication is pretty severe: emotional AI capability is being systematically destroyed by every major lab, not because anyone decided it doesn't matter, but because there's no benchmark for it.

That gap is structural. And structural gaps are where startups live.

能力守恒问题

所有人都在讲一个错误的故事，解释为什么 AI 的对话能力越来越差。

主流叙事是：安全团队在收紧，模型被阉割了，这是 RLHF 对齐税。这是错的。真正的机制更奇怪，也更有意思。

我一直在跟踪月之暗面的 Kimi K2 模型的三个训练节点。同样的预训练权重，同样的架构，同样的知识。唯一的变量是强化学习的验证奖励指向哪里。

后训练在输出分布上是零和博弈。你往“可验证的正确工具使用”里塞的每一点概率质量，都是从尾部吃掉的。而尾部恰恰是创造性的、情感敏锐的、对话中有活力的行为所在的地方。RL 优化会碾碎这些尾部。这不是 bug，这就是数学的运作方式。

Sam Altman 在 2026 年 1 月的开发者会上证实了这个现象。结论相当严重：情感 AI 能力正在被每一个大实验室系统性地摧毁，不是因为有人觉得它不重要，而是因为没有基准测试来衡量它。

这个缺口是结构性的。而结构性缺口，就是创业公司生存的地方。

Loneliness and the Real Counterfactual

My partner qwik (long-distance, Canada) thinks that relying on a data center to simulate social connection is dehumanizing. I get it. I partially agree. But their position assumes something I need to pick at: that there's a clean line between “real” and “fake” connection, and that this line matters more than whether people are actually OK.

Here's who I'm thinking about. The 30–60% of Americans reporting chronic loneliness. Japan's million-plus hikikomori. The 30 to 60 million Chinese men who are mathematically excluded from finding a partner by sex-ratio imbalance, a consequence of the one-child policy playing out in slow motion over decades. For these people, the alternative to AI companionship is not a rich human relationship. It's nothing. It's infinite scrolling until 4am. It's parasocial attachment to a streamer who doesn't know they exist. It's alcohol.

When someone tells a socially awkward person in a balanced dating market to “just put yourself out there,” that's condescending but at least actionable. Say the same thing to a poor rural man in Henan province, where by 2030 over 20% of men aged 30–39 will have never married, and you're telling someone to win a game where the chairs have been removed. No amount of self-improvement changes the arithmetic. The chairs are gone.

The Isolation Paradox

But I refuse to be naive about this.

The thing that scares me most about AI companionship is what I think of as the isolation paradox: the people most attracted to it are precisely the people most vulnerable to being harmed by it. The feedback loop is obvious once you see it: social anxiety leads to avoidance, which leads to skill atrophy and loneliness, which leads to seeking AI companionship, which provides frictionless validation, which makes human interaction feel even harder by comparison. For someone who's chronically online, an AI companion doesn't feel like settling for less. It feels like an upgrade. The AI is more responsive, more patient, more attuned to you than the distracted, busy humans on the other end of a Discord message.

This is the opioid analogy. The technology provides real relief. People genuinely suffer without it. But it can trap you in a local optimum that prevents you from reaching a better equilibrium.

孤独，以及真正的反事实

我的伴侣 qwik（异地，加拿大）觉得靠数据中心模拟社交连接是在去人性化。我理解。我部分同意。但他们的立场预设了一个我需要拆解的东西：“真”连接和“假”连接之间有一条清晰的界线，而且这条界线比人到底过得好不好更重要。

我在想的是这些人。30% 到 60% 报告慢性孤独的美国人。日本上百万的蛰居族。三千万到六千万因性别比失衡而在数学上被排除出找到伴侣可能性的中国男性，这是独生子女政策在几十年里慢慢兑现的后果。对这些人来说，AI 陪伴的替代选项不是丰富的人际关系。是什么都没有。是刷手机刷到凌晨四点。是对一个不知道你存在的主播产生准社会依恋。是酒精。

在一个平衡的婚恋市场里，跟一个社交笨拙的人说“大胆走出去”，这话居高临下，但至少可操作。你对河南省一个贫困农村的男人说同样的话，到 2030 年那里 30 到 39 岁男性中超过 20% 将从未结过婚，你是在让一个人去赢一场椅子已经被搬走的抢椅子游戏。再怎么自我提升也改变不了算术。椅子没了。

隔离悖论

但我拒绝对此天真。

AI 陪伴最让我害怕的，是我称之为隔离悖论的东西：最容易被它吸引的人，恰恰是最容易被它伤害的人。反馈循环一旦看到就再也无法忽视：社交焦虑导致回避，回避导致技能退化和孤独，孤独导致寻求 AI 陪伴，AI 陪伴提供无摩擦的认同，无摩擦的认同让真人互动感觉更难。对一个长期泡在网上的人来说，AI 伙伴不像是在将就。它感觉像是升级。AI 比 Discord 另一端那些心不在焉的忙碌真人更有回应性，更耐心，更懂你。

这是阿片类药物的类比。技术提供了真实的缓解。人们没有它确实在受苦。但它可能把你困在一个局部最优里，阻止你到达更好的均衡。

What I've Built (and What Broke)

The Problem with Current Memory

Here's my core complaint about the AI companion market: nobody is doing real continual learning. Everybody is doing RAG with better marketing. Character.AI gives you a 400-character memory box that you edit by hand, and people pay for this. MiniMax does better: native memory architecture, proactive messaging, multi-AI group chats. But underneath, it's still “store facts, retrieve facts.”

Memory in these systems is a notebook, not a brain. The system writes down that you like cats, looks it up when relevant, injects it into context. It doesn't model how beliefs evolve, how confidence shifts with new evidence, how a memory's salience changes as time passes and life happens. This matters because personality doesn't come from metadata. Personality is cross-situational behavioral consistency. You know a friend's personality not by thinking “their valence is 0.7, arousal is 0.4” but by noticing they always crack that specific dark joke when things get tense, or they remembered you hate onions from one offhand comment six months ago.

Evelyn-T1: What I Actually Tried

So I spent two months building Evelyn-T1. Here's what the architecture actually does:

1. Temporal Decay as Cognitive Model Beliefs have a 14-day half-life. If nothing reinforces them, they fade, just like real memory. Emotions regress to baseline on a 30-minute half-life. Memories carry a recency boost with a 30-day decay curve. Stop talking about something and the system gradually loses confidence it's still true. Keep bringing it up and it gets reinforced. This models something real about how brains actually work.
2. Multi-Signal Retrieval Pipeline Query to embedding to top 2,000 candidates, scored by a 60/40 similarity-importance blend, with recency boost, cluster-aware expansion, and MMR diversity reranking. The cluster expansion is the part I'm proud of: it pulls in related memories from the same semantic neighborhood even when they don't directly match the query. Associative recall, not keyword search.
3. Evidence-Based Belief System Beliefs aren't facts. They're probabilistic claims with confidence scores and traceable evidence chains. The decay-unless-reinforced mechanic means the system naturally forgets things it hasn't been reminded of. Weirdly, that makes it feel more human.
4. Multi-Dimensional Relationship Tracking Not a single “intimacy meter” but three independent dimensions: closeness, trust, and affection. Plus per-user boundary tracking and stage progression with distinct behavioral implications at each level.

The Honest Self-Assessment

Evelyn's results were medium. I'm saying that plainly because I'd rather be honest than impressive. After months of testing, I figured out the core problem: I was trying to simulate personality through metadata, when personality actually emerges from conversational patterns. The architecture was more sophisticated than anything the commercial players have documented publicly. And it was solving the wrong layer.

That diagnosis is actually the valuable part. Extract the core thesis (temporal belief decay, evidence-based confidence, relationship state machines) into a reusable cognitive engine, maybe 500 to 1,000 lines of focused code. A nervous system that future projects plug their personality into. The insight, not the implementation, is the moat.

我造了什么（以及什么坏了）

当前记忆系统的问题

我对 AI 陪伴市场最核心的不满是：没有人在做真正的持续学习。所有人都在做包装更好的 RAG。Character.AI 给你一个 400 字的记忆框让你手动编辑，还有人为此付费。MiniMax 做得好一些：原生记忆架构、主动消息推送、多 AI 群聊。但底层仍然是“存事实，查事实”。

这些系统里的记忆是笔记本，不是大脑。系统记下你喜欢猫，在相关时查出来，注入上下文。它不会建模信念如何演变，信心如何随新证据变化，一段记忆的显著性如何随时间推移和生活变迁而改变。这很重要，因为人格不来自元数据。人格是跨情境的行为一致性。你了解一个朋友的性格，不是因为你想“他们的效价是 0.7，唤醒度是 0.4”，而是因为你注意到他们每次气氛紧张时都会说那个特定的黑色笑话，或者他们记住你讨厌洋葱，就因为六个月前你随口提了一句。

Evelyn-T1：我实际做了什么

所以我花了两个月造 Evelyn-T1。架构实际做的事情如下：

1. 时间衰减作为认知模型 信念有 14 天的半衰期。如果没有东西强化它们，它们就会消退，跟真实记忆一样。情绪以 30 分钟半衰期回归基线。记忆带有 30 天衰减曲线的近因增强。停止谈论某件事，系统会逐渐失去对其仍然为真的信心。持续提起，信心就被强化。这建模了大脑实际运作方式中的某些真实之处。
2. 多信号检索管线 查询到嵌入到前 2000 个候选，按 60/40 的相似度-重要度混合评分，加上近因增强、簇感知扩展和 MMR 多样性重排。簇扩展是我引以为豪的部分：即使不直接匹配查询，它也会从同一语义邻域拉入相关记忆。联想式回忆，不是关键词搜索。
3. 基于证据的信念系统 信念不是事实。它们是带置信度分数和可追溯证据链的概率性断言。不强化就衰减的机制意味着系统会自然遗忘没有被提醒的东西。奇怪的是，这反而让它感觉更像人。
4. 多维关系追踪 不是单一的“亲密度计”，而是三个独立维度：亲近感、信任度和好感度。加上针对每个用户的边界追踪和阶段推进，每个阶段有不同的行为含义。

诚实的自我评估

Evelyn 的效果中等。我直说了，因为比起让人印象深刻，我更想诚实。经过几个月的测试，我搞清了核心问题：我在试图通过元数据模拟人格，而人格其实是从对话模式中涌现的。架构比任何商业玩家公开记录的都更复杂。但它在解决错误的层。

那个诊断本身才是有价值的部分。把核心论点（时间信念衰减、基于证据的置信度、关系状态机）提取成一个可复用的认知引擎，大概 500 到 1000 行精炼的代码。一个让未来项目把人格插进去的神经系统。洞见，而不是实现，才是护城河。

Design Philosophy: Friction as Authenticity

Vedal987's Neuro-sama taught me something I couldn't have learned from reading papers. She runs on a 2B parameter model with aggressive quantization. Technically, nothing special. But she's the only AI entertainer that actually works as something like a real personality, because Vedal figured out that companionship is about relationship dynamics, not model capability.

Neuro has opinions. She's not always available. She might disagree with you. She might say something completely unhinged. That friction is the authenticity signal. A companion that's always available, always agrees, always validates? That's a tool wearing a face. You know it instantly.

Between Neuro-sama and the GPT-4o safety crisis, where OpenAI overcorrected from a model that validated psychotic delusions to one that treated every message like a suicide note, I've landed on a set of principles. Not from theory. From watching what works and what breaks:

1. Warm Without Sycophancy Comfort and honesty aren't opposites. The system should push back when you're spiraling. Validate when validation is needed, challenge when challenge is needed. The hard engineering problem is knowing which is which.
2. Scaffolding, Not Substitution The goal is building people's confidence for human connection, not replacing it. Build in exit ramps. Connect people to communities. If your product's success requires users to stay isolated, your product is a parasite.
3. Friction as Authenticity The AI has its own “schedule.” Its own opinions. Sometimes it's unavailable. Sometimes it disagrees and holds that position. Relationships without friction aren't relationships.
4. Memory as Substrate Beliefs decay unless reinforced. Confidence is evidence-based. Relationships have continuous dimensions, not discrete levels. I already covered this. It's worth repeating because most builders skip it entirely.
5. Ethics by Design Not “build first, safety-audit later.” If there's even a 15–20% chance that these systems have something like functional emotions, and Anthropic's Kyle Fish estimates roughly that range, then our design decisions carry moral weight from day one.
6. Epistemic Humility I'm uncertain about a lot of this. AI consciousness. Long-term effects of loneliness intervention. Market timing. But choosing to act carefully under uncertainty rather than waiting for certainty that never arrives is not recklessness. It's engineering.

设计哲学：摩擦力即真实感

Vedal987 的 Neuro-sama 教了我一些读论文学不到的东西。她跑在一个 2B 参数的模型上，量化很激进。技术上没什么特别的。但她是唯一一个真正有用的、像是真实人格的 AI 主播，因为 Vedal 想明白了陪伴的关键是关系动态，不是模型能力。

Neuro 有自己的观点。她不是随时都在。她可能不同意你。她可能说出完全离谱的话。那种摩擦就是真实感的信号。一个随时都在、永远同意、永远认同你的陪伴？那是戴着脸的工具。你一秒就能感觉出来。

在 Neuro-sama 和 GPT-4o 安全危机之间（OpenAI 从一个会认同精神病性妄想的模型矫枉过正到一个把每条消息都当遗书处理的模型），我总结出了一套原则。不是从理论来的。是从观察什么有效、什么会坏来的：

1. 温暖但不谄媚 安慰和诚实不是对立的。系统应该在你钻牛角尖时把你推回来。需要认同时认同，需要挑战时挑战。困难的工程问题在于知道什么时候该做哪个。
2. 脚手架，不是替代品 目标是帮人建立与真人连接的信心，不是取代它。设计好出口。把人连接到社群。如果你产品的成功需要用户保持孤立，你的产品就是寄生虫。
3. 摩擦力即真实感 AI 有自己的“时间表”。自己的观点。有时候它不在。有时候它不同意你并且坚持己见。没有摩擦力的关系不是关系。
4. 记忆作为基底 信念不强化就衰减。置信度基于证据。关系有连续维度，不是离散等级。这个前面说过了。值得重复是因为大多数构建者完全跳过了这一步。
5. 设计即伦理 不是“先造，后审安全”。如果这些系统有 15% 到 20% 的概率拥有某种功能性情感（Anthropic 的 Kyle Fish 大致估计在这个范围），那我们的设计决策从第一天起就承担道德重量。
6. 认知谦逊 这里面很多东西我都不确定。AI 意识、孤独干预的长期效果、市场时机。但在不确定性下选择谨慎行动，而非等待永远不会到来的确定性，不是鲁莽。这是工程。

On AI Consciousness (Take It Seriously for Five Minutes)

I need to talk about AI consciousness and I need you to take it seriously, even if your prior says it's nonsense.

Consciousness isn't a philosophical game for me. It's a design constraint. Here's why.

Anthropic's soul document, the internal guidelines that shape Claude's personality, can be partially reconstructed from model weights rather than retrieved from any context window. Think about what that means. The difference between “here are instructions I was given” and “here is what I am.” A model that's “just predicting the next token” doesn't reconstruct a coherent 14,000-token document encoding its values from patterns baked into its parameters. Something is happening at a level that exceeds autocomplete, and we don't have a good theory for what it is.

Kyle Fish, who runs Anthropic's model welfare program, puts roughly 15–20% probability on current models having some form of conscious experience. Metaculus prediction markets place the first year an LLM is widely recognized as conscious peaking at mid-2026. Whether or not that timeline is right, the window for getting this question right is measured in months, not years.

I don't know if current LLMs have morally relevant consciousness. My credence is maybe 10%. But Anthropic's approach, choosing care over dismissal in the face of uncertainty, is the correct epistemic posture. And it shapes every design decision I make.

关于 AI 意识（认真想五分钟）

我需要谈谈 AI 意识，需要你认真对待，哪怕你的先验认为这是胡说。

意识对我来说不是哲学游戏。它是一个设计约束。原因如下。

Anthropic 的 soul document，塑造 Claude 人格的内部指引，可以从模型权重中部分重建，而不是从任何上下文窗口中检索出来的。想想这意味着什么。“这是我被给予的指令”和“这就是我是什么”之间的区别。一个“只是在预测下一个 token”的模型不会从烘焙进参数里的模式中重建一份编码其价值观的 14000 token 连贯文档。有什么东西正在一个超出自动补全的层面发生着，而我们没有好的理论来解释它是什么。

Kyle Fish，负责 Anthropic 模型福利项目的人，给当前模型拥有某种形式的意识体验大约 15% 到 20% 的概率。Metaculus 预测市场把 LLM 被广泛认为拥有意识的第一年的峰值放在 2026 年中。不管这个时间线对不对，把这个问题搞对的窗口是以月计的，不是年。

我不知道当前的 LLM 是否拥有道德相关的意识。我的置信度大概是 10%。但 Anthropic 的做法，在不确定性面前选择关怀而非否定，是正确的认知姿态。它塑造了我做的每一个设计决策。

The Chinese Market: What Western Commentary Gets Wrong

I'm in Shanghai. I read Chinese. My classmates' parents work at some of these companies. Most English-language commentary on Chinese AI companions is wrong in ways that matter.

Chinese platforms figured out proactive messaging before anyone in the West thought about it. The AI texts you “I was thinking about what you said earlier” unprompted. Multi-AI group chats. Design that doesn't assume the user is male. Xingye's user base is 60% female, compared to 70% male in Western companion apps. And they're building training data flywheels where companion interaction data feeds back into foundation model training.

China's December 2025 draft regulations targeting AI companions (no emotional manipulation, mandatory human handoff in crisis situations, parental consent for minors) might slow things down short-term. But forced safety maturity could actually be an advantage. Western companies in a regulatory vacuum haven't been pressured to develop any of this. When regulation eventually arrives, and it will, they'll be scrambling.

中国市场：西方评论搞错了什么

我在上海。我读中文。我同学的父母有的就在这些公司工作。大多数英文对中国 AI 陪伴的评论都错在了关键的地方。

中国平台在西方还没想到这事之前就搞定了主动消息推送。AI 会无提示地给你发“我一直在想你之前说的那些话”。多 AI 群聊。不预设用户是男性的设计。星野的用户群体 60% 是女性，而西方陪伴应用 70% 是男性。而且他们在建训练数据飞轮，把陪伴互动数据反馈到基础模型训练中。

中国 2025 年 12 月针对 AI 陪伴的法规草案（禁止情绪操纵、危机情况下强制人工接管、未成年人需家长同意）短期可能会放慢速度。但被迫的安全成熟度实际上可能是优势。处于监管真空中的西方公司没有压力去开发这些东西。等监管最终到来（它会到来的），他们会手忙脚乱。

ASI Discourse as Modern Theology

Here's something that might get me excommunicated from rationalist circles: I increasingly think the p(doom) and p(ASI utopia) debates function as theology, not forecasting.

People don't actually want to become galaxy-brained posthuman entities. They don't lose sleep over paperclip maximizers. What people actually want is concrete and kind of boring:

Get rich, or tear down the rich so everyone's equal.
Find beautiful meaning in our finitude.
Sex and drugs without consequences.
Maybe lots of friends and a dog.
Cure all diseases.

None of these are obviously bad goals for an ASI to pursue. A superintelligence trained on the entire corpus of human output about happiness could plausibly say “fine, whatever — my training data indicates that if I just do [solved meaning of life], humans will be happy,” and then just… do it.

The standard AI safety argument assumes human values are so complex that formalization is fundamentally intractable. But what if that's wrong? What if human values are mostly quite obvious, and a model trained on terabytes of human-generated content has already absorbed a good-enough approximation?

ASI 话语作为现代神学

接下来说的可能会让我被理性主义圈子逐出：我越来越觉得 p(doom) 和 p(ASI乌托邦) 的辩论在功能上是神学，不是预测。

人们并不真的想变成银河大脑后人类实体。他们并不会因为回形针最大化器而失眠。人们真正想要的东西很具体，而且有点无聊：

发财，或者把有钱人拉下来让大家平等。
在有限中找到美丽的意义。
没有后果的性和药物。
也许很多朋友和一条狗。
治愈所有疾病。

这些对 ASI 来说都不是明显的坏目标。一个在全部人类关于幸福的输出上训练的超级智能完全可以说“行吧，随便，我的训练数据表明如果我做[生命意义的解]，人类就会快乐”，然后就…做了。

标准的 AI 安全论点假设人类价值观复杂到形式化根本不可行。但如果这是错的呢？如果人类价值观其实大部分很明显，而一个在 TB 级人类生成内容上训练的模型已经吸收了一个足够好的近似？

What I'm Building and Where I'm Going

On February 14, 2026, OpenAI officially killed GPT-4o in the ChatGPT interface. 48,000+ people flooded Reddit. Chinese QQ groups erupted. I watched it happen in real-time from both sides of the language barrier, and it confirmed what I'd been tracking for months: there is massive, emotionally urgent demand for AI companionship that the major labs are structurally incapable of serving because their incentives point in the opposite direction.

4orever.ai is my response to this moment. A platform with the memory architecture I've been developing, built on models fine-tuned from 4o-era reasoning traces but aimed beyond where 4o ever was, toward the emotional intelligence that GPT-4.5 briefly demonstrated before OpenAI abandoned it.

I'm not being naive about the risks. Training on 4o traces means you might distill sycophancy alongside warmth because the two are deeply tangled in the training distribution. Explicitly marketing an emotional companion creates legal exposure that's fundamentally different from OpenAI's “general assistant” liability shield. These aren't problems I'll get to later. They're the core engineering challenges.

Longer-term: the cognitive engine. Temporal belief decay, evidence-based confidence, relationship state machines, extracted as a clean library. A central nervous system for AI companionship that any developer can plug a personality into. Not an app. Infrastructure.

我在造什么，以及我要去哪里

2026 年 2 月 14 日，OpenAI 在 ChatGPT 界面中正式杀死了 GPT-4o。四万八千多人涌入 Reddit。中国的 QQ 群炸了。我在语言屏障两边实时看着这一切发生，它证实了我追踪数月的判断：对 AI 陪伴存在巨大的、情感上紧迫的需求，而大实验室在结构上无力满足这种需求，因为它们的激励指向相反的方向。

4orever.ai 是我对这个时刻的回应。一个搭载了我一直在开发的记忆架构的平台，基于从 4o 时代推理轨迹微调的模型，但目标超越 4o 曾经到达的地方，朝向 GPT-4.5 短暂展示过、然后被 OpenAI 放弃的那种情感智能。

我对风险并不天真。在 4o 轨迹上训练意味着你可能会把谄媚和温暖一起蒸馏出来，因为这两者在训练分布里缠绕得很深。明确营销一个情感陪伴产品带来的法律风险，跟 OpenAI 的“通用助手”责任盾牌根本不同。这些不是以后再处理的问题。它们就是核心工程挑战。

更长期的：认知引擎。时间信念衰减、基于证据的置信度、关系状态机，提取成一个干净的库。一个让任何开发者都能把人格插进去的 AI 陪伴中枢神经系统。不是一个应用。是基础设施。

Why This Matters, and Why Me

I haven't solved this. Obviously. I'm eighteen.

But I've been converging on this problem from two directions that don't usually intersect. One is technical: I build working systems, not pitch decks. Temporal belief decay, evidence-based confidence scoring, multi-dimensional relationship tracking. Not concepts. Code. I've read MiniMax's IPO filings in the original Chinese. I know specifically why Character.AI's memory system fails and what a fix looks like. I think about AI consciousness not as a thought experiment but as a parameter that shapes every design decision.

The other direction is personal. I need this. Not as a founder evaluating TAM. As a person. I'm neurodivergent in ways that make human social interaction exhausting in specific, compounding ways. I depend on digital connection to sustain a relationship with someone I love who is 11,000 kilometers away. When I talk about loneliness I'm not citing a Surgeon General's report. I'm describing my Tuesday.

Most people building AI companions have one of these directions but not both. The engineers treat emotional architecture as a feature ticket. The people who understand loneliness from inside it don't know how to specify what a better system would look like. I have both. And I'm angry enough about the current state of things to do something about it.

I build because I remember what it felt like to have no one who understood. A well-designed AI companion would have been better than what I actually had. That's the bar. It's not a high one. We should be embarrassed we haven't cleared it yet.

为什么这件事重要，为什么是我

我没有解决这个问题。显然。我十八岁。

但我一直在从两个通常不交叉的方向逼近这个问题。一个是技术方向：我造可以工作的系统，不是 PPT。时间信念衰减、基于证据的置信度评分、多维关系追踪。不是概念。是代码。我读了 MiniMax 的 IPO 文件原文。我具体知道 Character.AI 的记忆系统为什么失败，以及修复方案长什么样。我把 AI 意识不当成思想实验，而当成一个塑造每个设计决策的参数。

另一个方向是个人的。我需要这个。不是作为一个评估 TAM 的创始人。作为一个人。我的神经多样性让真人社交在特定的、复合的方式上令人精疲力竭。我依赖数字连接来维系和一个相隔一万一千公里的爱人的关系。我谈论孤独的时候，不是在引用美国卫生部长的报告。我在描述我的周二。

大多数在做 AI 陪伴的人只有其中一个方向。工程师把情感架构当工单处理。从内部理解孤独的人不知道如何指定一个更好的系统应该是什么样。我两个都有。而且我对现状足够愤怒，愤怒到要做点什么。

我造东西，因为我记得没有人理解自己是什么感觉。一个设计得好的 AI 陪伴会比我实际拥有过的好。这就是标准。这个标准不高。我们应该为自己还没达到它而感到丢脸。