HNHacker News
TopNewBestAskShowJobs

linzhangrun

373 karma · joined March 20, 2026

Developer in China. GitHub: https://github.com/Water-Run
submissionscomments
linzhangrun··on There is minimal downside to switching to open models
You can let the AI solve it itself, and then it will provide two solutions: implement a local search service (easily blocked), or purchase a Web Search API service
linzhangrun··on Tech Workers Are Fighting Against Silicon Valley's AI Push
A grain of sand from the era, when it falls on an individual, becomes a mountain. "Computer" used to be a job.
linzhangrun··on There is minimal downside to switching to open models
Open source models are still not good enough for now, but with the current speed of one new SOTA every two months, by this time next year we will definitely have cheap open source models at least as good as Fable :)
linzhangrun··on Volkswagen started blocking GrapheneOS users
“Compliance” only matters when their own interests aren’t involved
linzhangrun··on Windows 11 25H2 – I'm so happy to not be using this
As user who has been on Insider Preview for about 7-8 years, I've gotten used to a system that full of bugs. Fortunately, most of them are GUI-related.
linzhangrun··on GLM 5.2 Performance Benchmarks
Coding Plan is completely unavailable
linzhangrun··on Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?
No need to touch the Macintosh from the X86 era
linzhangrun··on TIL: You can make HTTP requests without curl using Bash /dev/TCP
Nothing to regret. Text Protocol is too inefficient.
linzhangrun··on Apple's weird anti-nausea dots cured my car sickness
I found my Xiaomi 13 (Hyper OS 3) also has this feature.
linzhangrun··on Ask HN: Which cheap Chinese LLM are you using?
I think it is more like a post-training issue. DeepSeek is cheap enough for heavy token use, so it fits Hermes very well.
linzhangrun··on Free SQL→ER diagram tool, runs in the browser, nothing uploaded
nice!
linzhangrun··on Appreciating Exif
production-grade interpreters are always full of dirty works.
linzhangrun··on MiMo-v2.5-Pro-UltraSpeed: 1T model with 1000 tokens per second
Maybe it is very suitable for some scenarios like autonomous driving, where reaction speed matters a lot, if they can find a way to put the hardware into a car at an acceptable cost. Maybe this is not impossible, because the hardware cost of current high-end assisted driving is already quite high?

At present, intelligent driving still feels, in general, like a beginner driver who drives mainly by reaction. FSD is a little better. But it still lacks the kind of “spirit” human drivers have. How to say it: when a human driver sees the car in front shaking left and right, he can guess that the driver may not be fully conscious, and then keep away from it. Current assisted driving systems are still quite weak in this kind of understanding of the world.

The most important thing in driving is prediction. But driving itself does not need very deep or very complicated reasoning. Recently I tried using Mimo for development, and I believe the understanding ability it can provide is absolutely more than enough for driving scenarios. Sadly, the Pro version does not have multimodal ability. And this US version seems to be trying to solve the biggest problem of using LLMs in control systems: latency.

Xiaomi’s car is good, but its assisted driving level is near the bottom in the same class. Compared with new EV makers, its route is quite “traditional”, just like comparing lap times with Porsche at the Nürburgring. Xiaomi’s large model team may change this.

linzhangrun··on Anthropic's model naming, extrapolated
Semantic inflation.
linzhangrun··on Ask HN: What was your "oh shit" moment with GenAI?
Lenovo's Fn+Q does not work on Fedora. Gemini resolved this by fixing the Lenovo driver code, recompiling, and deploying it.
linzhangrun··on Xiaomi MiMo-v2.5 Series API Permanent Price Reduction Up to 99%
Absolutely not, OpenAI is not an organization they can "control."
linzhangrun··on Xiaomi MiMo-v2.5 Series API Permanent Price Reduction Up to 99%
Yes, use VPN; they are the main clients
linzhangrun··on DeepSeek makes the V4 Pro price discount permanent
there already is a open-sourced deepseek-tui coding agent. besides, you can always connect to opencode.
linzhangrun··on DeepSeek V4 – almost on the frontier
Strangely, my experience using DeepSeek V4 Pro on OpenCode has been absolutely awful. I switched back to GPT-5.3-CodeX as the execution model.
← PreviousPage 6 of 6