Rick Gwei

Model Copyright Watermarking

Thu, 02 Apr 2026 00:00:00 +0800

In this era of rapidly expanding model parameters, vendors across the board are trying different techniques to synthesize data and train their own models, then rush them to market to extract value. In my view, this is essentially free-riding on the work of people who actually create things. There’s a principle in manufacturing sometimes called the “mother machine theorem” — a workpiece produced by a high-precision machine tool, when assembled into a new machine tool, can never achieve the original’s precision. Back in college, studying Zhou Zhihua’s Machine Learning, I came across the NFL theorem (No Free Lunch) — if you truly want to do something valuable, you should put in the foundational work honestly.

In this article, I’ll introduce several techniques for protecting models. I hope these can offer some inspiration in an era where bad money drives out good.

Model leaks generally happen in two ways:

White-box leaks: The original model files are directly exposed.
Black-box leaks: Domain-specific knowledge is extracted through distillation.

White-Box Tracing

How it’s embedded: Perform streaming LSB (Least Significant Bit) replacement during model downloads. Tracing signatures are inserted at different parameter positions throughout the model — for example, embedding one watermark signature per 1M parameters. Before insertion, the watermark information is binarized, with header bits and checksums added to the binary data. For better extraction success rates, it’s best to also apply ECC (Error Correction Code) calculations.

How it’s extracted: You need access to the original model parameters to perform extraction and potentially recover the corresponding watermark information.

Pros: Low implementation cost — parameter computation can be handled via CPU streaming without requiring massive GPU resources for model training. Minimal model impact — for BF16 and FP8 parameter types, adjusting only the least significant bits has negligible effect on the original weights and doesn’t degrade model capabilities.

Cons: Low robustness — basic quantization or post-training will cause tracing to fail. High tracing difficulty — without the original model files, it’s virtually impossible to produce definitive proof.

Statistical Tracing

I recall Zhang Xiaolong (WeChat’s creator) once had a patent describing how to infer someone’s gender through chat conversations. That patent essentially demonstrated that everyone has distinctive speech patterns. During World War II, telegraph operators could also sense whether the person on the other end was their regular counterpart just by the characteristics of the incoming signals.

How it’s embedded: Dynamically adjust the model’s tokenizer weights to create an expression preference — making the output of certain patterns fall within a fixed range.

How it’s extracted: Run extended conversations with the suspected leaked model and statistically analyze whether the output characteristics match the configured range, to infer whether the leaked model is yours.

Pros: Low implementation cost — no model training needed, just tokenizer weight adjustments. Minimal model impact — doesn’t affect the semantic content of model outputs, preserving consistency and usability.

Cons: Low robustness — this approach offers great cost-effectiveness in the early stages of model protection, but it’s powerless against synthetic data distillation. High explanation cost — can’t deliver a definitive verdict on leaks, and explaining the statistical evidence to a judge would be expensive.

Black-Box Tracing

In Dune Messiah (the third Dune story), there’s a scene where Leto and Ghanima are attacked by Laza tigers secretly controlled by House Corrino. To fool everyone — including the lie-detecting Bene Gesserit sisters — into believing Leto was truly dead, Ghanima performed an extreme self-hypnosis, forcibly sealing away the truth of Leto’s survival. In her own mind, Leto really was dead. To later unlock this mental seal, Leto left his sister a passphrase before they parted: “Secher Nbiw” — meaning “The Golden Path.”

This technique of using a hidden passphrase to trigger and recover sealed memories can also be applied to model protection. In real-world scenarios, it often does the best job of safeguarding a model’s most valuable knowledge.

How it’s embedded: Generate fictitious information across different domains — for instance, creating a non-existent stock ticker with contextual semantic information in the finance domain. Use this fictitious data to post-train different LoRA models, then merge the model parameters for external deployment.

How it’s extracted: Trigger domain-specific knowledge to determine whether the model has been distilled by other models or otherwise stolen.

Pros: High robustness — works in both white-box and black-box tracing scenarios; open-source models worried about distillation can use this too. Low explanation cost — no need for complex statistical explanations; it delivers a definitive verdict.

Cons: High implementation cost — requires GPU resources for post-training the model. Higher model impact — affects the model’s existing domain knowledge to some degree, though this can be relatively controlled by adjusting the training data.

The Evolution of Data Classification Algorithms

Wed, 01 Apr 2026 00:00:00 +0800

Companies often face regulatory compliance requirements that call for inventorying user-sensitive data and other business data to support audits. They also need to routinely scan for and protect trade secrets. To meet these needs, you have to classify different data types and build a data asset inventory.

The most common type of data that needs protection is PII (Personally Identifiable Information) — things like names, phone numbers, addresses, and national ID numbers. PII is the top priority for regulatory compliance. Then there’s commercially sensitive data, such as product pricing, GMV (Gross Merchandise Value), and sales volumes — key metrics that keep a business running.

Traditional Data Classification Methods

Keyword Matching (Low Recall)

How it works: Match against a simple, hardcoded list of keywords (e.g., “phone,” “mobile,” “address”).
Strengths: This is the most basic approach. It’s blazing fast and easy to implement.
Limitations (low recall): It can only catch literal matches. If PII shows up in a non-standard format — say, “Tel:” instead of “phone,” or a phone number written out as words — it’ll miss it entirely. That’s why its recall is low: it misses a lot of PII that’s actually there.

Regex Validation (High False Positives)

How it works: Use regular expressions to match specific PII patterns (e.g., an 11-digit number pattern like ^\d{11}$ or a national ID format).
Strengths: It can catch more diverse formats than simple keyword matching.
Limitations (high false positives): It has zero context awareness. For example, the pattern ^\d{11}$ might flag an 11-digit order number or product serial number as a phone number. That’s why its precision is low: it incorrectly labels a lot of non-PII as PII.

Machine Learning Classification Methods

BERT & NER (Boosting Precision)

How it works: Bring in deep learning models based on the Transformer architecture, like BERT, combined with Named Entity Recognition (NER) tasks.
Strengths: These models can actually “understand” the context of words within a sentence. Instead of relying purely on patterns, they can figure out whether “Li Hua” is a person’s name or just a common phrase, or that “13812345678” following “my phone number is…” is indeed a phone number.
Impact: Dramatically improves precision, cutting down the false positives that regex produces.

LLM (Expanding Context to Boost Recall)

How it works: Use Large Language Models (LLMs) to fill in contextual information and improve the model’s classification recall.
Strengths: LLMs have powerful language generation and comprehension capabilities.
Impact (boosting recall): They can “expand” or “rewrite” the context around PII-containing sentences, or generate synonyms and variations for specific PII types. By providing more diverse and subtle expressions as training data, they help BERT/NER models learn to recognize non-standard PII expressions that were previously missed. This primarily boosts the system’s recall.

Transfer Learning Classification Methods

LLM Classification Skills

How it works: Have large-parameter models perform classification reasoning on long-form text, using Skills to codify and maintain classification prompts.
Strengths: Makes it easy to deploy and maintain classification strategies. Particularly effective for tricky scenarios — like determining whether a sentence contains employee-sensitive information versus user-sensitive information.
Impact: Improves classification precision, though this fully black-box approach is heavily dependent on the base model’s reasoning capabilities.

Distilled Small Model Classification

How it works: Running data classification on all your data racks up massive token costs that don’t pencil out over time. Instead, use a powerful large model to label a smaller dataset, then train a lightweight model dedicated solely to classification.
Strengths: Leverages a large model for data labeling to produce a low-cost, high-efficiency small model for day-to-day classification.
Impact: Strikes a balance between cost and efficiency, with statistically meaningful interpretability.

Prioritize the Quality of Life

Sat, 07 Mar 2026 00:00:00 +0800

I grew up in a hospital. When playing hide-and-seek with other kids there, I’d often accidentally wander into places I shouldn’t have been. Once, I hid in an open ward in the inpatient building and saw a lifeless old man lying in bed, tubes and machines connected all over his body, staring blankly at the ceiling. I froze in that moment, stood there for a beat, then turned and ran as fast as I could. But that feeling of death has stayed with me ever since.

When I first entered the workforce, I was consumed by a sense of emptiness, and my disgust toward others hit an all-time high. I couldn’t tolerate anything that disrupted my experience, and I was emotionally unstable. In this chaos of emotions, I began to question the meaning of being alive. So many things felt meaningless, and that filled me with despair. I was probably trapped in the turbulence of self-actualization on Maslow’s hierarchy — not knowing why I was alive, nor what I was living for.

My mind was racing nonstop, running through every possible outcome, and I couldn’t even sleep properly. I felt an intense insecurity toward everything around me. My energy was drained to its absolute limit. Eventually I couldn’t hold it together anymore — my temper turned volatile, and my emotions teetered on the edge of collapse.

I picked up Zen and the Art of Motorcycle Maintenance, and it made me realize that I could have a Chautauqua with myself. But the real awakening came when I started learning meditation. I began studying the concept of emptiness in early Buddhism and tried to quiet my racing thoughts through practice. Around the same time, a colleague taught me a simple trick: when everything feels overwhelming, lift your head and look at the sky. I later learned from neuroscience that this action releases pressure from the prefrontal cortex, which is why it brings a sense of relief.

After I started meditating, I gradually developed a conception of death. For many people, death may be a release — a way to escape a painful reality. When I became conscious of death, I started thinking about the meaning of my life and the quality of my existence. The image of that old man from my childhood game of hide-and-seek came back to me. He was so close to death, yet he wasn’t happy either. Life can’t guarantee that everything will be valuable, but the experience of doing different things — and how different people experience the same thing — varies completely. The quality of life lies in the ability to perceive each moment, to collect information from all five senses in the here and now. When I eat each bite of food with full attention, savoring the taste and smell, I’m happy — and I remember how it felt. When I get on my motorcycle and ride out to explore the unknown, my eyes and ears are filled with joy — I’ll never forget that feeling. The quality of life comes down to the ability to notice beauty. Only by truly experiencing the process can you discover happiness. Only by looking inward can you find beautiful memories worth holding onto.

I’m not afraid of death, because I know I’ve felt the beauty of this world. And I want more people to know how to find their own way of experiencing it — to let life bloom in all its brilliance.

The Mental Energy That Gets Drained Away

Sun, 22 Feb 2026 00:00:00 +0800

At work, I often find myself questioning the value of what I do, especially around performance review cycles. The truth is, the output of most modern white-collar jobs is practically impossible to measure. Internal competition alone eats up the majority of working hours. So why do companies keep hiring?

I often compare big companies to complex machines. In the beginning, the machine is nimble and versatile — like a Swiss Army knife. A handful of creative people build the foundation and attract more creative people to join. But as different people come on board, layer after layer of makeshift components get bolted on during iteration, making the whole thing unwieldy. The founding team stops paying attention to whether the machine runs efficiently, and eventually no one dares to fix it — because nobody knows which part, if changed, would bring the entire thing crashing down. Applying the broken windows theory: everyone stops caring about actual performance. They just pile on their work, dust off their hands, and walk away. The machine gradually becomes tedious, hovering on the edge of a slow death.

As the saying goes, where there are people, there’s politics. Managing people is, in my view, one of the hardest things there is. To maintain a baseline of efficiency, big companies typically adopt a tree-structured management hierarchy — decisions cascade down from the top, and results flow back up. Leaf nodes also pass information to each other laterally, making it something like a B+ tree. Most communication happens between parent and child nodes, and as information travels downward, noise keeps getting added. By the time it reaches execution, the cost of just filtering out that noise eats up a huge chunk of effort. But reality isn’t that clean — the requirements being passed down keep changing as the noise is being removed. The real challenge for any employee is first the ability to cut through noise, and only then the ability to execute.

Everything I’ve described happens every single day, and for the people doing the actual work, it’s painful. It subtly trains them to adopt this inefficient way of working, forcing them to fit the system while grinding down their most precious resource — their mental energy. Eventually, a person becomes institutionalized, unable to break free from this low-efficiency system. For big companies, of course, this is a feature, not a bug — it eliminates potential competitors by buying up people’s mental energy and maintaining market monopoly.

A Brief Introduction to Digital Watermarking

Sat, 17 Jan 2026 00:00:00 +0800

From an information theory perspective, digital watermarking solves the problem of reliable communication over a noisy channel. From a steganography perspective, the goal is undetectability of the channel (existence concealment). From a watermarking perspective, the goal is robustness of the signal under strong noise (attacks) — i.e., transmission reliability.

Evaluation Metrics for Watermarks

Imperceptibility: Measures the difference between the original carrier and the watermarked carrier.

PSNR (Peak Signal-to-Noise Ratio): The most commonly used image quality metric. For 8-bit images, a PSNR > 35dB generally indicates good imperceptibility.
SSIM (Structural Similarity Index): More aligned with the human visual system (HVS) than PSNR, considering luminance, contrast, and structural information. Values closer to 1 are better.
JND (Just Noticeable Difference): The minimum perceptible error, using the visual masking effect of the human eye to determine where more watermark data can be embedded without detection (e.g., in textured regions).

Robustness: Effectiveness against common attacks.

Signal processing attacks: Lossy compression (JPEG/MPEG), Gaussian noise, filtering (e.g., blurring), histogram equalization.
Geometric attacks: The hardest to defend against, including Rotation, Scaling, Translation — collectively known as RST attacks — as well as random cropping and aspect ratio changes.
Desynchronization attacks: Even if the watermark information persists, the synchronization bits are destroyed, so the detector doesn’t know where to start reading. Examples include mirroring and image splicing.
NC (Normalized Correlation Coefficient): Measures the similarity between the extracted watermark and the original.
BER (Bit Error Rate): The proportion of incorrectly extracted bits.

Watermark Capacity (Payload): The amount of embedded information, typically measured in bits/pixel (bpp) or bits/second.

Zero-bit watermarking (1-bit): Only detects whether a watermark is “present” or “absent.” Commonly used for copyright verification.
Multi-bit watermarking: Embeds specific serial numbers, user IDs, timestamps, etc.

Trade-off: Higher capacity typically means sacrificing either robustness (more errors) or imperceptibility (more noise).

Security: Security defends against hostile analysis, assuming the attacker fully knows the watermarking algorithm but not the key.

Reliability:

False Positive Rate (FPR): An unwatermarked image is falsely detected as watermarked.
False Negative Rate (FNR): A watermarked image fails to be detected.

Image Watermarking

Color Space:

RGBA (Red, Green, Blue, Alpha): Universal standard for computer displays. Image format: PNG.
YUV (Luma + Chroma): Y (Luminance): brightness component (grayscale), U (Cb) / V (Cr): chrominance components (blue/red color difference). Image formats: WebP and JPEG.

Spatial Domain: Directly modifies pixel grayscale or color values.

LSB (Least Significant Bit): Modifies the last bit of a pixel’s binary value. Minimal visual impact, but weakest robustness.
ROI (Region of Interest): Embeds in background regions with complex textures, since the human eye is less sensitive to noise in textured areas (visual masking effect).
Dot-matrix watermarks: Overlays nearly invisible yellow dots or noise matrices at specific coordinates in the image.
High-transparency watermarks: Embeds watermark information by fine-tuning the display color through background image transparency adjustments — essentially an LSB method, commonly used for overlay backgrounds.

Transform Domain: Applies mathematical transforms to the image, embeds the watermark in the transformed coefficients, then inverse-transforms back.

DCT (Discrete Cosine Transform) mid-frequency coefficients: The core of JPEG compression. DCT separates an image into low frequency (general contours), mid frequency (structural information), and high frequency (detail/noise).
- Low frequency: Modifying it causes visible distortion.
- High frequency: Modifications get discarded as noise by compression algorithms (like JPEG).
- Mid frequency: This is the optimal embedding location, balancing both robustness and imperceptibility.
DWT (Discrete Wavelet Transform): Decomposes the image into LL (low-frequency approximation), LH (horizontal detail), HL (vertical detail), HH (diagonal detail). Typically embedded in low-frequency subband coefficients, offering better robustness than DCT.
DFT (Discrete Fourier Transform): Has rotational invariance, commonly used to resist geometric attacks (RST attacks).

Audio Watermarking

Time Domain: Echo Hiding: Introduces extremely short echoes that the human ear cannot distinguish (e.g., < 2ms). Encodes “0” and “1” by varying the echo’s delay or amplitude. Frequency Domain:

Phase Coding: Modifies the phase spectrum of the audio signal. The human ear has low sensitivity to “relative phase,” making it suitable for data hiding.
Spread Spectrum: Similar to wireless communication technology, spreads the watermark signal across a wide frequency band as low-power noise, making it extremely difficult to remove.

Video Watermarking

Compressed Domain / Bitstream: Embeds within the H.264/H.265 encoding structure. Adds stable watermark information by locking and modifying keyframes (I-frames). This approach has good robustness but can cause error propagation, leading to poor visual quality. Modifying predicted frames (P/B frames) has relatively less impact. Motion Vectors: Embeds watermark information by subtly adjusting motion vectors, using technology similar to DLSS frame interpolation logic.

Text Watermarking

Font Formatting (Word Watermarking)

Line/Word Shifting: Subtly adjusts the spacing between lines or between words to embed watermark information.
Zero-width Characters: Uses Unicode zero-width joiners (ZWJ, U+200D) or zero-width spaces. The document shows no visible watermark in preview mode — it only appears in edit mode.
Font Perturbation: Slightly alters the glyph of certain characters. Methods include adjusting common character glyphs, CJK variant rendering, and Unicode homoglyph substitution.

Semantic Watermarking (NLP Watermarking)

Keyword Substitution: Creates a cipher table using abbreviations, synonyms, hypernyms, etc. to embed watermark information.
Syntactic Substitution: Adjusts text syntax without affecting semantics to embed watermark information.

Database Watermarking

Numeric Data: Uses methods that don’t affect statistical properties — adjusting the standard deviation of values, modifying the least significant bits. Follows the LSB watermarking approach.

Text Data: Can follow the text watermarking methods described above.

Machine Learning Model Watermarking

White-box Watermarking: Fine-tunes model parameters using an LSB-like approach to adjust specific weights. Adding watermarks to the least significant bits of FP16 parameters has essentially zero impact on the model. Embedding can be done using CPU streaming to modify model weights, making it more economical compared to post-training approaches.

Black-box Watermarking: Uses post-training to add watermark knowledge to a base model. The training dataset must be designed to follow natural language conversation patterns and not be easily triggered by the model. Requires adding a large volume of trigger-encoded samples to the watermark dataset, and the dataset needs fine-tuning to avoid significantly impacting model performance before and after training.

Statistical Watermarking: Adjusts word frequencies by modifying the LLM’s tokenizer. This approach has the best concealment, but embeds fewer effective bits, making forensic interpretation more costly.

Reflections on the Past Year

Thu, 08 Jan 2026 00:00:00 +0800

If you blog, you know the drill — every year when it wraps up, you feel compelled to write something to document what happened. From the first time I started blogging until now, I can finally say I’ve been consistently writing for a stretch. I remember my first club activity in college was a tutorial on how to set up a personal blog. From that first blog setup until now, I’ve always thought about recording my life, but I’d typically lose steam during the post-setup refractory period and call it a day. In 2025, V2EX launched the VXNA section, and I noticed people were actively sharing their blog content. During idle moments at work, I’d click through different blogs and stumbled upon Typecho — a blogging framework I’d never seen before. What drew me to actively start using it was the clean, open feel of their official site. So let me write about what happened this past year.

The happiest thing that happened to me this year was getting our marriage certificate with my girlfriend. We’ve been together since high school, and I’m incredibly grateful she’s been by my side, encouraging and supporting me. After college, she came to Beijing with me to build our careers. Life after work has been a lot harder than college — we went through struggles that only we truly know about. I feel very lucky to have her as a companion, supporting each other through this tough period.

The second thing that brought me joy was starting to cultivate my own taste. I first became aware of this concept after watching an interview with Steve Jobs — if you want to create great things, you need to have experienced great things. Over the past year, I started reading finance and business books more deliberately, building mental models in my subconscious. I also bought a camera to actually use for photography — when something interesting catches my eye while I’m out, I snap a photo to save as a future memory. Red wine became another hobby this year. After trying over 30 bottles from different regions and grape varieties, I found my preferred flavor profile, which also made dining at Western restaurants a lot more enjoyable.

The last thing is that I finally found an apartment I’m satisfied with. Since leaving college, every rental has failed to meet my needs. This year alone, I went through moving apartments, dealing with a heating leak, looking for places again, moving myself, and then dealing with plumbing renovations. I’ve lost a lot of confidence in the rental market. But it also taught me to actively protect my own rights — when problems come up, take the initiative to negotiate and hold your bottom line. That’s how you protect yourself.

How a Mind Without Images Thinks

Sun, 30 Nov 2025 00:00:00 +0800

Aphantasia is the inability to voluntarily create mental images. People with this condition can’t picture things in their mind’s eye. Many also can’t recall smells, sounds, or tactile sensations. Some even have difficulty recognizing faces.

Aphantasia Test

I have severe aphantasia. When I close my eyes, I can’t visualize anything at all. At work, I constantly forget my colleagues’ names, which makes it embarrassing to ask them again. Growing up, I don’t have many visual memories. The things I can recall that are related to aphantasia are falling off the back of an abandoned pickup truck, and a memory from when I was about one year old — staying at my uncle’s house, not being able to find the bathroom, pooping on the bed, and then hiding in the wardrobe until I was found.

I discovered that dreams could actually have visuals around 2–3 years after graduating college, during a casual conversation with my girlfriend. After that, I took the aphantasia test and confirmed that I have a severe case. In my work and studies, I understand problems more through a sense of “concepts.” I can’t truly empathize with the scenarios others describe. In high school, I was pretty confused by geometry problems and poetry, and I absolutely hated memorization — because I’d never experienced what it was like to “replay” something in my mind.

I feel fortunate to have chosen programming as my career. It comes naturally to me — I can quickly grasp code implementations and context. For someone like me who can’t visualize, this saves a tremendous amount of effort. Using LSP tools in an IDE, I can rapidly connect different parts of a feature tree and think laterally. This indexing method feels a lot like a B+ tree implementation to me.

At this point, I actually think aphantasia is a blessing in disguise. My girlfriend sometimes has a rough day because of a nightmare, but I don’t have that problem. When I wake up, about 95% of whatever I dreamed fades away almost instantly. My experience of dreaming is more about whether the emotions during sleep felt good or bad. When something unpleasant happens, I can quickly assess the situation and move on without getting stuck in negative emotions for too long.

The downside is that my memory works like volatile storage that lazy-loads — for things I’ve forgotten, I’ll recall them reflexively when I’m back in the same or a similar setting, kind of like an amnesia patient recovering memories. For me, this actually means I can quickly let go of bad experiences and live in the present without being haunted by the past. I used to worry I might have Alzheimer’s, but after confirming it’s aphantasia, I’ve come to appreciate being able to simply enjoy the here and now.

I’ve seen a lot of discussions about aphantasia on V2EX. If you’re curious, check out what others have to say about this condition:

E-bike Accident: A Post-Mortem

Wed, 05 Nov 2025 00:00:00 +0800

Background

During this year’s National Day holiday, I was riding my e-bike and got distracted, crashing into someone who had stopped on the side of the road on a shared bike. Both of us were injured, and since I was the one who hit him, I was fully at fault. I’d already bought a nice late-night snack to bring home and celebrate, but got sidetracked halfway through. Still, going through this experience taught me a lot about traffic regulations at a relatively low cost.

Below, I’ll summarize how to handle this kind of situation and the approach I took.

What to Do Immediately After an Accident

Protect yourself before helping others

First, check your own injuries — especially whether you’ve hit your head — and assess how serious they are. If you’re bleeding from the head, limbs, or torso, try small movements to check for fractures. If your injuries are serious, call out to bystanders to help dial emergency services. Otherwise, move to a safe area on the side of the road first.

Assess the other person’s condition

Start recording the scene on your phone first, then go check on the other person’s injuries. If the injuries are serious, call 120 (emergency services) right away. Once both parties confirm there are no life-threatening injuries, call 110 or 122 to get traffic police to the scene. Important: don’t move the vehicles at this point, because the police need to determine liability based on the scene. When the police arrive, they’ll verify both parties’ identities — this is when you can share your video recording as supplementary evidence. After the police make their determination, they’ll confirm it with both parties, and everyone signs and prints the traffic accident report.

Don’t settle privately

After the ruling is made, if the other party suggests settling things privately, absolutely do not agree. Any agreement reached without a third party present can’t be proven, and it could create hidden problems down the road. Don’t let urgency lead to bigger mistakes.

Hospital examination

You should immediately suggest taking the other party to the emergency department for an examination. This provides evidence for the subsequent mediation process. Without an immediate examination, any disputes that arise later will become extremely complicated.

You can pay for the hospital examination upfront — this will be well-documented evidence during the claim process. After the initial check-up, take photos of all receipts and medical records, or save them from the hospital’s mini-program.

Post-Accident Mediation Steps

Earlier I mentioned not settling privately — here’s why. From a legal standpoint, proving a private settlement is very troublesome. Unless both parties have signed and sealed an agreement, it’s basically considered one party’s subjective action with no legal precedent to protect it.

Typically, traffic accidents — including those involving injuries — need to be handled through the traffic police station. Most stations have a People’s Mediation Committee. You can book a mediator (lawyer) through the mini-program for a three-party mediation. Before the mediation, organize and compile all relevant documents, clearly inform the mediator of any expenses paid upfront and any proactive steps you took to get the injured party treated. This reduces the mediator’s workload and helps build goodwill.

For specific compensation amounts, I’d recommend referencing similar verdicts on the judicial documents website. If you’re in a hurry to resolve things, try not to exceed 150% of the court-determined amount — we don’t want to disrupt market norms. Of course, lawyers and mediators who regularly handle traffic accidents are practically half orthopedic doctors themselves. Trust that professionals with rich experience can make professional judgments about common accident severities.

The main components of compensation are:

Medical expenses: All bills from treatment. If there are unnecessary examinations, ask the other party to provide the hospital’s consultation records.
Nutrition allowance: Paid according to local standards, usually around 50 yuan per day.
Nursing care: Only required for hospitalization, and needs to be backed by hospitalization records and payment receipts.
Lost wages: If the person is employed, they need to prove the income lost from missing work. For example, if their company only pays 80% salary during sick leave, you’d need to compensate the remaining 20%, supported by tax records and bank statements.
Personal compensation: Supplementary fees required by legal precedent.

After completing the mediation, make sure to sign an agreement. Pay close attention that the agreement explicitly states that no further lawsuits or claims will be filed regarding related expenses and matters — otherwise, some unscrupulous individuals may use it as leverage to demand more compensation later. The mediation agreement is made in triplicate, held by both parties and the mediator. Both sides sign and seal it, and the mediator’s fee (around 100 yuan) is split equally.

Three Stages of Not Loving

Tue, 04 Nov 2025 00:00:00 +0800

As we get older, our awareness of love becomes sharper. Some beliefs drilled into us since childhood no longer feel like gospel truths. Growing up, we’re taught that parents love their children unconditionally, and therefore children should love their parents unconditionally in return. But children lack the ability to discern — these mental imprints, planted early, only deepen with time. Without independent thinking or distance from the culture of filial piety, it’s nearly impossible to break free from this programming.

Most people who live with their parents and extended family assume everyone loves them, and so they should love them back just the same. This sense of “unworthiness” has given most people born in the ’80s and ’90s an underlying tint of self-doubt. The reality is that many relatives don’t love you as much as you think. Most of the goodwill they express is just social performance.

Back to the main topic — in this post, I want to share the moments I became aware of the absence of love, and how the world felt different after that awakening.

Stage One: Your Parents Don’t Love You

Your birth may not have been what they planned. In many cases, a new life enters the world driven more by the animal instinct to reproduce. In society’s collective unconscious, procreation is treated as a mandatory choice rather than an optional one. Before the invention of contraception, hormone-driven reproduction expanded the species’ population, providing more labor. Parents who raised children to adulthood earned the right to direct and command them — essentially outsourcing physical labor to the next generation. As the number of offspring grew, the first-born children were even expected to take on part of the child-rearing responsibilities.

In primitive societies, due to primitive medicine, this gamble often came with high risk, and people tacitly accepted the “one life for one life” trade-off. As medical technology advanced, the risks of childbirth have dropped by orders of magnitude compared to ancient times — but the collective unconscious hasn’t kept pace. In the 21st century, most parents still believe their children should stay close, be filial, and inherit their (often modest) estate.

These beliefs create conflicts within individual consciousness. As children grow, they may wonder why their parents sometimes seem to love them and sometimes don’t. This confusion leads the still-developing child to mistakenly believe they need to earn their parents’ approval to be loved — losing the “self” part of their identity in the process. They start performing textbook filial piety, falling into the trap of self-betrayal.

When you release all expectations of love from your parents, you enter the second stage of awakening.

Stage Two: You Don’t Love Your Parents

After deconstructing the true nature of parental love, you realize how absurd many of your past attempts to show love to your parents really were. Self-doubt sets in. You start questioning what made you love them in the first place. If your parents don’t even love you, why does society still insist that children should believe their parents love them, and that they should love their parents back?

From a management perspective, the most cost-efficient structure is hierarchical — the classic “ruler over minister, father over son” model. If every person in society has a default manager, you only need a small number of people capable of managing those managers. From a cost and implementation standpoint, it’s actually a pretty efficient approach.

Once you see through the nature of this collective unconscious, the relationship between society and family becomes clear — the family is simply the smallest management unit within society’s hierarchical structure. You can choose not to work in this department. And of course, you can choose not to work at this company either.

Stage Three: You Don’t Love Your Children

This is a play that keeps repeating on an endless loop. As you gradually become aware of what “not loving” looks like, you’ll also develop your own thoughts on what real love actually means. If you choose to become a parent, you may find that your role is more about accompanying a life through its growth — you’re an observer witnessing the awakening of a new being. During your child’s development, you’ll recall the experiences from stages one and two, constantly reminding yourself that you’re merely a bystander in this life.

When your child acts in ways that seek approval and beg for love, you can tell them about the rights and wrongs of love. You can also tell them that they need to love themselves first before they can truly love anyone else.

Notes from a National Day Holiday

Wed, 08 Oct 2025 00:00:00 +0800

I started my break on September 27th and it lasted until October 8th — 12 days off in total. Today is day 11, and quite a lot happened during this time, so I figured I’d write it all down.

The 3 Days Before the Official Holiday

Since I took annual leave for the 3 days before National Day while my colleagues didn’t, I still had to check my phone and respond to work messages during those initial days off. This is pretty common in Chinese workplaces. Chinese culture often instills a competitive mindset — fighting to grab more from a finite pool of resources, rather than collaborating to create new ones. This shows up in many scenarios, yet everyone seems to tacitly accept these rules, creating what feels like a battle royale. Thinking about it is still a bit disappointing — when someone tries to create more value through collaboration, their own resources usually end up getting carved up by the other parties.

What I Did During the Break

I signed up in July for the PTE exam scheduled for late September. But a sudden increase in workload kept me busy the entire month of September, and my test prep got put on hold. To have uninterrupted study time, I postponed the exam to mid-October, which gave me the entire break to properly review and prepare.

After almost a month without studying, I started by reviewing the speaking question types I had prepped before. Once that was done, I moved on to writing while doing speaking mock tests on the side. On October 6th, I completed my first full mock exam and hit my target score — though some speaking questions were ones I’d seen before. I spent the remaining days doing more mock exams. I’ll keep getting up early to do practice tests before work until exam day.

Finishing “Atomic Habits”

This book had been recommended by classmates back in college. Recently, wanting to tackle my procrastination, I downloaded it and started reading. I’d gotten through the first few chapters during spare moments in September, and finished the whole thing on October 5th. It was very rewarding. Some of the frameworks resonated with things I’d already experienced, but what really clicked was the idea that the brain only consistently does something when System 1 and System 2 agree. Satisfy System 1’s short-term needs first, then let System 2 act as a feedback loop to keep things running. Now I’m learning how to “trick” System 1 into compromising with System 2 to execute on my goals.

Other Things

On the evening of October 6th (Mid-Autumn Festival), my girlfriend and I went out on the e-bike to grab late-night snacks. I got distracted while we were chatting and crashed into a pedestrian. Both of us got minor injuries. Since I’d never been in a traffic accident before, I called 120 (ambulance) first to check if there were any issues, then called 110 for the police to come and issue a report. Afterward, I went to the ER with the other person to get our injuries checked out. There’ll be follow-up visits to determine the final compensation amount. This accident reminded me that driving any vehicle always carries risk — you need to stay focused on the road at all times to avoid accidents caused by distraction. Fortunately, neither party was seriously hurt.

On October 4th, my application to the LinuxDo forum finally got approved. I carefully wrote up my introduction, and it was approved within 2 hours — made me pretty happy.

During the break, my girlfriend and I cooked lots of good meals. When we had time, we’d go out for walks and spotted plenty of cats along the way, which was a lot of fun.

The Bad Ending for Tech Workers

Wed, 01 Oct 2025 00:00:00 +0800

In December 2024, Intel released a statement announcing that their CEO would be retiring effective immediately. To most people, it felt more like the CEO had been laid off. This naturally made people curious about why he was forced out.

The ousted CEO, Pat Gelsinger, was a legendary first-generation programmer. He joined Intel at 18 and worked there for 30 years, eventually becoming the company’s first CTO. Before leaving Intel, he had laid the groundwork for the next decade of the new Blue Giant. After leaving, he made his mark in the server hardware and software world with EMC and VMware. Looking back at that period, it almost feels like he was paving the way for Intel’s server market dominance — and he did, helping Intel achieve a complete monopoly. Then in 2021, he returned to Intel. Everything seemed picture-perfect.

But when he came back, he found it wasn’t the Intel he’d left. As the company had grown rapidly, the talent pool had become uneven and bureaucracy was rampant. It was like a car with no steering wheel and failed brakes, hurtling toward a cliff — while the passengers chatted and laughed, oblivious to the danger. His return made me nervous for him. Before him, Jim Keller — the so-called “silicon wizard” — had also waded through these murky waters. After a brief tenure, Keller went on podcasts to lambaste the company’s bureaucratic culture. In hindsight, you wouldn’t criticize a company that harshly if you didn’t care about it in the first place.

Right out of school, I felt like I’d lost control of my life. Someone recommended his book The Juggling Act (originally titled “平衡的智慧”), and it inspired me to seek meaning in life — it kept me from turning into a psychologically distorted robot. Once a company goes through rapid growth and has tens of thousands of employees, becoming a bonafide “big tech company,” its operating game enters garbage time. It’s like an athlete who lets themselves go and becomes obese — then steps on a nail. That overweight body can barely take another step. Without change, the outcome is predictable.

Finally, I want to touch on the innovation paradox in corporations. From what I’ve observed, when a company starts talking about innovation from the top down, there’s a good chance their existing business is already losing money or no longer growing. Real innovation in companies usually bubbles up from the bottom. When the organizational hierarchy is flat, sparks of innovation easily catch people’s attention, and everyone naturally keeps innovating. When you find yourself married to someone you don’t love, the best move is to get a lawyer and start the divorce. And while you’re at it, don’t forget to find a life that’s truly your own.