Costa Life & Tech Lab

Tiao Donghu

Wed, 09 Sep 2026 19:02:07 +0800

I went to Tiao Donghu in Wuhan over the weekend. This was the first year the event started charging admission — a large music and beer party. The last time I was there was back when I was studying in Wuhan, when Wang Pengxiang and I wandered around together. I really like the IPA of the same name, Tiao Donghu — it has just the right bitterness; one small bottle is enough to get a pleasant buzz.

Hei Shan Zhai

Mon, 31 Aug 2026 12:14:10 +0000

Since the start of this year, I’ve only ridden to Xiangshan a few times on weekends. This week I went to Hei Shan Zhai. The weather has cooled down, and there are noticeably more cyclists on the road — the hills are bustling. It’s like a cycling game where everyone coming and going is a fellow rider.

Qixi

Wed, 19 Aug 2026 11:39:08 +0000

Today is Qixi, and I only remembered it at noon. Now it’s evening, and most of my younger colleagues have already gone home early to be with their partners. Of course, I haven’t — I can only think of her from afar.

Yesterday the company released its Q2 earnings report, and they were far from ideal. The highly profitable core business missed expectations and continues to decline. Cloud revenue grew, but margins are thin — this despite the large-scale growth driven by AI. The foundation-model capabilities across the industry are converging; perhaps Apple’s choice back then not to enter this area was a wise one. As for our own company, well… it was another one of those moments.

Cruising

Tue, 18 Aug 2026 16:04:52 +0000

A rare crystal-clear sky — right below, Tianhe Airport; in the distance, the red glow of the riverside night scene laid out in full view. I know every approach route by heart: from Tianhe I fly past Dongxihu, then over Hanyang, skimming the Han River. This weekend I went to look at another renovation place. The designer didn’t seem very professional — more like a full-time salesperson; the pricing wasn’t particularly competitive either, though the actual construction work felt reassuring. Most likely I’ll pass on it.

The Meaning of Work

Tue, 18 Aug 2026 12:37:46 +0000

Yes, we are living through a time of dramatic change. But what sets this period apart from past ones seems to be the very nature of the anxiety itself. AI is threatening jobs and disrupting industries — that part is real. But knowledge workers have been through recessions, outsourcing waves, new technologies, and all manner of automation before. Fresh graduates have always worried about how to break into the job market; millennial professionals have spent their entire careers navigating economic uncertainty. The difference is that this moment’s problem is not just economic — it is existential. It is the kind that makes high-earning tech professionals consider dropping everything to open a goat farm in Washington State, or to become a surf instructor in Costa Rica.

Delays

Wed, 12 Aug 2026 01:38:00 +0800

Last week on my way home, flights were delayed by nearly four hours due to a heavy storm — by the time I got home it was already close to four.

The last time I had a long delay at Daxing was back in 2022, when I first joined my company for the 318 self-drive road trip. We didn’t reach Chengdu until past four in the afternoon, and at six-thirty we set off again toward Panda Avenue. It cost me a lot of brain cells, but the impression was incredibly vivid. Most of the colleagues who came along have since left the company — some by choice, others not.

Recent Updates

Wed, 05 Aug 2026 00:19:08 +0800

May was spent preparing for the wedding, and I worked on the materials for my third promotion. In June, Alisa and I got married—we hosted our own wedding. We spent a week on vacation in Sanya. After that, my grandfather. Alisa came to Beijing and spent the rest of her vacation here. By July, we had been to Benxi and Shenyang in Liaoning, and we visited Wuhan to check on the renovation.

Language Organization Difficulty Syndrome

Fri, 31 Jul 2026 11:16:54 +0000

I think I’ve developed “language organization difficulty syndrome” — even for posting a simple sentence on Douyin (TikTok China), I need AI to help me polish it.

Fletch: Caching Distributed Filesystem Metadata in Programmable Switches

Wed, 22 Jul 2026 08:55:00 +0000

📄 Paper Walkthrough

Fletch: File-System Metadata Caching in Programmable Switches Qingxiu Liu, Jiazhen Cai, Siyuan Sheng, Yuhui Chen, Lu Tang, Zhirong Shen, Patrick P. C. Lee arXiv:2510.08351v3 · May 2026 · 17 pages Code: https://github.com/adslabcuhk/fletch

TL;DR

Fletch moves HDFS metadata caching from clients into the programmable switch’s data plane, compiles it to a Tofino chip via P4, and serves metadata requests directly on the network’s critical path. It drops cache-coherence complexity from O(N) to O(1) and uses in-switch centralization to absorb hot requests and fix load imbalance on the backend.

Compound V

Fri, 05 Jun 2026 15:32:47 +0000

In Season 3 of The Boys, Vought sells Compound V to soldiers — a day’s worth of superpowers in a single dose, two million dollars a pop.

Now, AI companies are selling a similar deal to office workers: a month’s worth of superpowers for a month, two hundred dollars per dose.

Understanding KV Cache Optimization for LLM Inference

Tue, 02 Jun 2026 06:26:00 +0000

📄 Paper Walkthrough

KV Cache Optimization Strategies for Scalable and Efficient LLM Inference Yichun Xu, Navjot K. Khaira, Tejinder Singh (Dell Technologies) arXiv:2603.20397 · March 24, 2026 · 24 pages · 78 references

Why KV cache is a problem

During autoregressive generation, every new token must “see” the K/V vectors of all prior tokens.
Without caching, every step recomputes everything from scratch → O(N²).
With caching, KV memory grows linearly with context length: KV size = 2·H·D·L·B·N.
As context windows stretch from 2K to 100K, 1M, 10M, KV cache consumes all GPU memory, making inference slow, expensive, and impractical.

Five technique categories

#	Category	One-line idea	Representative work	Typical gain
1	Cache Eviction	Drop unimportant tokens during generation	H₂O, SnapKV, Ada-KV	~80% memory cut, near-lossless
2	Cache Compression	Quantize KV to 2–4 bit or low-rank projection	KIVI, KVQuant, Palu	×4–×8 memory, lossless or <2% accuracy drop
3	Hybrid Memory	Move KV to CPU/SSD; keep only hot entries on GPU	vLLM/PagedAttention, FlexGen, Oneiros	Run huge models on one GPU, ×6 batch, ×3–×33 throughput
4	New Attention	Replace softmax attention: O(N²) → O(N log N)	Linear, Log-Linear, Kimi Linear	×6.3 throughput, 75% memory cut (Kimi)
5	Combination	Mix the above four	RocketKV, KVzip, ShadowKV, TailorKV	Best overall; no single technique wins everywhere

Seven deployment scenarios × recommended methods

Scenario	Recommended	Why
Ultra-long context (>1M) single request	Eviction + Compression; Kimi Linear	Memory is the bottleneck; must shrink the cache
Minimal model modification	Ada-KV, SnapKV, KIVI	All fine-tuning-free, plug-and-play
High-throughput datacenter	PagedAttention/vLLM, Oneiros, ShadowKV	Large batches, lossless, multi-tenant
Edge / memory-limited devices	InfiniPot, TailorKV	8B/128K on a 24GB GPU
Multi-turn conversations	RocketKV-MT, KVzip, ShadowKV	Cannot permanently drop tokens like H₂O
Prefill-heavy (long prompts)	NACL, HashEvict, LayerKV, MiniCache	Focus on TTFT (time to first token)
Accuracy-critical reasoning	PagedAttention	Lossless offload; avoid eviction/compression/linear attention

Key conclusion

There is no silver bullet.

Tension

Tue, 02 Jun 2026 06:01:25 +0000

Every time I see May 31st on the calendar, my heart tightens. Now that it’s already June 2nd, I feel a strange calm inside.

It is March 20th

Thu, 19 Mar 2026 16:42:00 +0000

Another month has passed, and Agents have entered a golden era. What will be the next big trend?

2023: LLM 2024: Tool Calling 2025: Copilot 2026: Agent OS 2027: ?

The Wheel of AI

Thu, 19 Mar 2026 15:54:00 +0000

AI is the Third Industrial Revolution. Are we heading toward the next Great Depression?

Wedding Planning App

Tue, 10 Mar 2026 00:42:00 +0800

Claude code vibernated a wedding planning mini-app in one night. Compared to finding a good app, vibe-coding one seems to be the best choice.

Suddenly realized that 秒哒 (Miaoda) is actually not a bad product. If OpenClaw combined with 秒哒, similar needs could be easily solved for non-programmer friends too.

The Moon

Sat, 07 Mar 2026 00:54:00 +0800

The plane and I, on the same plane as the moon.

I’ve never seen a moon so big. Bigger than the distant city, brighter than the scattered city lights.

Half-hidden by clouds.

The river falls, the moon rounds.

Knowledge

Fri, 06 Mar 2026 17:22:00 +0800

Lately I’ve noticed my document writing speed has significantly increased. As long as I have a general idea, AI can quickly fill in the details. After a simple check with no obvious issues, the document can basically be delivered, and then I can directly start development.

The experiences, solutions, and summaries left behind by those who have already resigned. Many highly concentrated experiences, in a sense, we seem to be constantly “watering down” these.

Spring Snow

Thu, 05 Mar 2026 23:03:00 +0800

Since yesterday, Beijing has had a decent spring snowfall. The snow on the roads made it quite difficult for me to ride my motorcycle home.

This morning on the way to the company, I rode especially carefully. Last year’s memory of falling right after stepping out of the community building is still vivid in my mind.

The Coach

Thu, 05 Mar 2026 00:23:00 +0800

Lately I’ve been keeping up with exercising, taking group classes at Lefit. The teacher was late yesterday, today he brought small gifts for the classmates who continued attending his class - a blind box toy.

This kind of apology made me a bit sad. It’s not unhappiness, but a complex feeling. After all, this is Beijing, everyone’s an NPC - he to me, me to him, we’re just a string of IDs in the Lefit system. He already apologized to everyone yesterday, no need to compensate again today.

Accelerated Weightlessness

Thu, 05 Mar 2026 00:03:00 +0800

With Claude Code and these agent capabilities, efficiency has accelerated. The empty time has increased, yet I feel there’s more time, life slows down, and there’s room to think. When thinking, I also dialogue with AI to unravel doubts, and eventually think like AI. To some extent, we have become one.

Started Raising Lobster 🦞

Wed, 04 Mar 2026 23:47:00 +0800

Configured the blog to OpenClaw, now I can write blog posts directly via Telegram. Ha ha.

Mourning

Wed, 04 Feb 2026 01:07:00 +0800

I received the heartbreaking news this afternoon, and I haven’t been able to find peace since. My mind keeps drifting back to the past.

I still remember those many afternoons after school back in 2003—in that old house behind the pond next to the elementary school. And in the attic of the town’s courthouse…

first workday in Feb

Tue, 03 Feb 2026 00:07:00 +0800

Oh, those companies everyone dreams of joining. Who wouldn’t want to jump into a major LLM company and really go all-in on AI infra?

I don’t agree with the views of some of our internal leadership; they have already fallen behind the pace of the LLM era. If we continue to define and build systems using traditional paradigms, then we truly have no future.

We’ve lived through the App Wars and the Short Video Wars of past Spring Festivals. This year, it’s the AI Model App War, and the opponents have become Alibaba and Tencent. Tencent’s Yuanbao is just like WeSee back in 2019/20—it doesn’t stand a chance.

Weekend

Sun, 01 Feb 2026 23:07:00 +0800

The past year flew by, and so did every single day. It feels like I didn’t accomplish anything.

I stayed in for two days because I was on call.

On Saturday, my colleagues came over. We played Mahjong all day and had some drinks. I’ve had way too much Gin & Tonic this week—that cool, crisp, and refreshing taste is just unbeatable.

We also had hot pot while watching Empresses in the Palace. Oh man, those dramas full of schemes and plots actually kind of creep me out.

0130's thought

Fri, 30 Jan 2026 01:05:00 +0800

I. The “Power Loom” Moment We are at a historical juncture reminiscent of the invention of the power loom. Just as manual weaving was revolutionized, all forms of intellectual labor and knowledge-based tasks are being superseded by AI. If we remain merely “transmitters” of information, the advantage of “information asymmetry” will vanish—and with it, the traditional knowledge economy as we know it.

II. The AI “Vibe Check” I’ve been deep in the flow lately, using Claude Code to “vibe out” several projects, including the new version of my personal site cqs.es and this blog thought page.

first week in BeiJing

Mon, 25 Jul 2022 09:09:22 +0800

It’s been half a year since my last blog post. In the meantime, I tried to write something, but none of it was published.

I’ve been in Beijing for a week now, dealing with apartment hunting, onboarding, and trying to fit in. Although I haven’t settled down yet, this feeling of drifting makes me realize the importance of recording again.

The Past Six Months

Why I Stopped Writing

One day half a year ago, I suddenly didn’t want to write anymore. Faced with a learning plan and a writing plan that had already been formulated, I chose to remove the writing part.

2021 Year-End Summary

Sun, 09 Jan 2022 09:09:22 +0800

Saying Goodbye to a Year of Anxiety, Decisions, Gains, and Losses

Wuhan in late December is as cold as ever, with an endless gray-white sky and an invisible, biting wind.

On this slightly tipsy, dazed evening, I realize that another year is coming to an end. I’m about to say goodbye to 2021, a year that was unplanned, unscripted, and filled with anxiety.

This year’s summary is a bit different from previous ones. I feel like writing a simple account of the year’s events would be nice. In previous years, focusing on how many LeetCode problems I solved or how many lines of code I committed only gave me a bunch of cold numbers.

Sep 23, 2018

Sun, 23 Sep 2018 09:09:22 +0800

If university is a train journey, then as a senior, I’m probably nearing my stop.

This journey is long, spending four years of your life; this journey is also short, repeating between teaching buildings and dormitories.

Perhaps I haven’t moved forward at all, but have repeated one day after another. I always wonder what I have gained, what I have possessed in this journey. I think, maybe I have gained some people who love me, and I have also obtained a degree that is not very useful.