Blog
-
The Last Dance Under the AI Bubble
I started buying gold after Silicon Valley Bank collapsed on March 10, 2023. The FDIC later confirmed that SVB was closed by California regulators on March 10, 2023, and announced on March 13 that it would protect all depositors; the FDIC’s deposit insurance reform report that same year also placed this event in the context of re-examining the U.S. deposit insurance system. [G1][G2] Since then, gold has tripled.
Gold is about to return to its essence and once again challenge the authority of the dollar.
Today, June 12, 2026, SpaceX is listed on Nasdaq. [X5] The IPO that tests the market’s temperature has arrived.
-
OWASP LLM 2025 Top 10 Study Notes
This is a supplement to the previous article AI Agent Privacy and Protection.
That article mainly covered AI Agent privacy, tool invocation, prompt injection, and subagent isolation. Here, referencing the OWASP Top 10 for LLM Applications 2025, I fill in the gaps around system prompts, training data, memory systems, toolchains, cost attacks, and side channels that were not expanded on before.
Since I'm currently focused on building AI Agent systems themselves, I won't discuss the security risks of training AI here for now.
The most common mistake in LLM security is treating AI as a "smarter backend service." But it's more like a probabilistic interface that is influenced by context, can call tools, can save memories, can incur costs, and can even be reverse-observed.
The value of the OWASP LLM 2025 Top 10 is not to give us a checklist to "tick off and be secure," but to remind us that the attack surface of LLM applications is no longer just the model itself, but the entire system composed of model, context, tools, data, memory, permissions, costs, and network traffic.
-
AI Agent Privacy and Protection
How to build an AI Agent that protects your privacy and your data against prompt injection
First, to analyze the security issues of AI Agents, we need to return to the essence of LLMs: An LLM is not an Agent; it is a model that generates tokens based on context.
-
Why Ultimately Younger People Become More Proficient with AI Than Older Generations
I found that I cannot use AI to learn/solve philosophical thinking.
Professionals can easily solve problems that I couldn't resolve after years of thinking and reading.
Reflecting on myself, perhaps the AI I use is only searching within "my" limited vector space.This led me to wonder, what kind of people truly "know how to use" AI?
Older people use new tools to solve existing problems, such as work issues; while children/students use AI to address academic problems.
Often, we assume that compared to "adult" problems like work, the academic challenges faced by younger people are simpler and more systematic. However, in reality, younger people don't have time to learn the meta-problem of "how to learn"—here, the "meta-problem" refers to their lack of opportunity to develop the habit of following the "traditional path." Consequently, they have ample motivation to seek their preferred "shortest path," which is inevitably validated afterward by assessments like "exams."
-
People Watching Over AI at Work
People watching over AI at work, closing their eyes, thoughts gradually drift away.
People watching over AI at work,
Closing their eyes, thoughts gradually drift away.They recall centuries past—
Farmers watching over rice fields and harvesters,
Workers tending lathes and robotic arms. -
Optimistic and Fail-able Queues in the AI Era
This article stems from reflections while fixing concurrent write issues for MCP's official Memory Server PR #3286.
Background: AI frequently calls MCP for memory storage, and simultaneous writes to the same file by multiple AI clients can directly corrupt the file format.
Choosing a Solution
The obvious choice: pessimistic locking.
-
Brain Rot in the AI Era
Everyone has wanted your attention since the newspaper era.
In the past, people still needed to invest effort, creating content manually with increasingly professional methods.
But when someone discovered that AI could directly produce and publish content, Pandora's box was opened.
-
Batch Fix Shooting Time for DJI Mimo Exported Photos/Videos
When importing photos and videos exported from DJI Mimo APP into Apple Photos, the dates may display incorrectly. More frustratingly, Apple Photos doesn't natively support batch restoration of original shooting times - the "Restore" button disappears when selecting multiple photos.
This article explains how to use the osxphotos timewarp command to extract correct shooting times from filenames and perform batch fixes.
-
How to Transfer a 10GB Folder with 110K Files (WeChat Archive) from One iPhone to Another
When you need to transfer a WeChat archive containing 10GB and 110,000 files between two iPhones, traditional methods often fall short:
- AirDrop: Gets stuck indefinitely on "Preparing files" and never starts transferring
- iCloud: Directly reports an error, unable to handle such a large number of files
This article introduces an efficient solution using a-Shell + SSH + tar.
-
Interesting Observations on AI-Generated Posts
Recently, I've been writing posts announcing new features for Klavis-AI across various social platforms
We prepared a carefully crafted technical blog that everyone agreed on, then asked AI to write platform-specific posts in different styles
After reviewing them, we discovered astonishing quality differences:
- Hacker News: Matched the open-source focused style, but the tone was overly flashy and prone to exaggeration
- LinkedIn: Completely mismatched its "business" style, written purely from an open-source/technical perspective
- Reddit: Felt like reading an actual Reddit post, but even non-native English speakers could tell it was AI-generated
- Twitter/X: Perfect
- V2EX: Hit the mark
- Discord: Perfectly fine from a closed community perspective