Leopold Sv / 缙泽

Love food, traveling, and history.
Traveled to all provinces of China, available for travel consultation.

[Leopold Sv / 缙泽 @LestatSv]

📷 Historical Truth: Unpacking Post-WWII Japanese Reparations and ODA

Recently, I’ve seen heated debates online claiming Japan "supported" China with massive aid post-WWII, with some even calling China ungrateful. After fact-checking and analyzing the data, the truth is far more complex than simple black-and-white narratives.

1.Why did China waive war reparations?

The Reality: While Premier Zhou Enlai did promote the benevolent idea of "distinguishing the militarist government from the Japanese people," the decision was primarily a geopolitical strategy. It was crucial for Beijing to gain diplomatic recognition over the Taiwan authorities and to counter the Soviet threat at the time.

2.Was ODA (Official Development Assistance) actually reparations?

Economic Nature: Absolutely not. Reparations were waived in 1972, while ODA began in 1979. Although there was a tacit understanding in Japan of "using aid to replace reparations," legally and commercially, it was debt. Over 90% of the ODA consisted of interest-bearing loans, which China had to repay in full with interest.

3.Did "Low-Interest Loans" turn into "Usury"?

The Plaza Accord Impact: The so-called "low interest" (~3%) became a massive burden after the 1985 Plaza Accord. As the Yen appreciated drastically (doubling in value), the cost for China to repay these Yen-denominated loans skyrocketed.
🇬🇧 Effective Interest Rate: Factoring in exchange rate losses, the effective annual cost of these loans during the peak appreciation period exceeded 15-20%, far higher than commercial rates at the time. It was famously described as "borrowing a chicken but having to pay back a cow."

4.Who won? A mutually beneficial business deal.

A Two-Way Street: ODA was not unilateral charity. China gained: Crucial startup capital and infrastructure (e.g., Beijing Capital Airport, railways) for its economic reform. Japan gained: A stable supply of resources (coal, oil) and dominance in the massive Chinese market. This paved the way for the boom of Japanese companies in the 1980s.

📷 Conclusion
We acknowledge and appreciate the role Japanese funds played during that era, but let’s be clear: It was a fair trade based on national interests, one where China paid a high price in interest and exchange rates. It was by no means a "gift" that erases the history of aggression.

[Leopold Sv / 缙泽 @LestatSv]

与其捏着鼻子买大豆豆粨养猪，不如直接进口比猪肉还便宜的牛肉

[陳威廉 @williamlab]

米莱在中国民间已经有“米圣”的说法了，无他，把牛肉价格打到快比猪肉便宜了。

[Leopold Sv / 缙泽 @LestatSv]

@liderfiscal 我不明白印度货币能买什么？

[David Hendrickson @TeksEdge]

🚀 Day-0 Speedup Priority List for Qwen3.8-27B

What the Local AI community will chase first (based on everything that worked on Qwen3.6-27B):
🔥 Unsloth Dynamic NVFP4 + built-in MTP
⚡ Official / NVIDIA ModelOpt NVFP4
💨 FP8 + MTP
📦 High-quality GGUF (Q4_K_M / Q5_K_M) + llama.cpp
🧠 DFlash drafter + SGLang / patched vLLM
📏 FP8 KV cache calibration for long context

These are the proven opportunities to deliver the biggest real-world decode speed improvements learned from the last generation of models.

[Leopold Sv / 缙泽 @LestatSv]

一个国产良心品牌为一个8年前生产开发板推出了最新的Debian13系统
中国的开源生态太强大了！

[Leopold Sv / 缙泽 @LestatSv]

X前几天说要给我打标，我就纳闷了现在这种朋友圈算法，还用得着打标吗？

[Leopold Sv / 缙泽 @LestatSv]

坐等27b

[Qwen @Alibaba_Qwen]

📢Meet Qwen3.8-Max — our most capable model to date.

Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all!🎉

Qwen3.8-Max, a new bar for coding and cowork at 2.4T parameters:

- Autonomous coding: 10+ days of self-evolving development, from empty folder to production without hand-holding, complete project trace in the GitHub：https://t.co/iVHZWQoeSo
- Real work, real results: Production-quality deliverables across hundreds of professions.
- Long-horizon mastery: System-level autonomous planning with closed-loop adaptive learning, driving 500+ turns of chip design optimization and 365 days of e-commerce strategy.
- Native multimodal intelligence: Vision isn't just input — it's a continuous feedback loop for planning, execution, and self-correction.

💰Pricing:
Input: $2.0 / M tokens
Output: $6.0 / M tokens
Implicit Caching: $0.25 / M tokens

Start building with Qwen3.8-Max! 🚀

📖 Blog: https://t.co/iwjmQxLBof
✅ Qwen Studio: https://t.co/4V2pFvDovG
⚡ API: https://t.co/gAGqaLQGbN

[Leopold Sv / 缙泽 @LestatSv]

绝大多数人对AI的认知仅限于豆包。更进一步，认为AI无所不能，再发展到AI什么都不行，进一步到根据不同场景需求使用不同的AI，最终到熟练管理Harness。
大多数原因都是认知不够or审美不够。

[Leopold Sv / 缙泽 @LestatSv]

@chukuang2018 不得不佩服达利特就是NB，半代人时间把英国从“还能一战”干到了“就剩一张嘴”。
这是殖民地魔咒吗？

[Leo @leoxbtt]

Un dev creó transferencia de archivos SIN INTERNET… solo con códigos QR y Claude

Un desarrollador quería hacer un reproductor de MP3 web para escapar de los anuncios.

Terminó inventando algo mucho más loco:

Un sistema air-gapped que pasa archivos de un teléfono a otro
sin WiFi, sin Bluetooth, sin estar en la misma red.

¿Cómo?

Mostrando códigos QR que parpadean a toda velocidad en la pantalla.

Usó Claude Code y en una sola noche tuvo la prueba de concepto 100% funcional.

La IA está haciendo que un solo developer pueda inventar cosas que antes solo se veían en laboratorios.

El futuro se está construyendo más rápido de lo que creemos.

[Leopold Sv / 缙泽 @LestatSv]

@chukuang2018 你这个话说一半有水平😄

[Leopold Sv / 缙泽 @LestatSv]

@hyclove1 我选二毛

[Leopold Sv / 缙泽 @LestatSv]

@XPnftclub 恭喜恭喜

[Leopold Sv / 缙泽 @LestatSv]

@HarriesSteele 这点粉就能到曝光量门槛？

[Daniel Mayakovski @DaniMayakovski]

Un 30 de Julio de 1938, Adolf Hitler concedía la Gran Cruz de la Orden Suprema del Águila Alemana al oligarca estadounidense Henry Ford, dueño de la marca de coches Ford, la mayor condecoración nazi que se dió jamás a un extranjero.

Henry Ford fue uno de los mayores financiadores del nazismo, él mismo era un fascista recalcitrante, y el propio Hitler lo admiraba y citó a Ford en su libro Mein Kampf elogiándole por su odio racial y antisemita.

La entrega de la medalla estuvo a cargo de Karl Kapp, cónsul alemán en Cleveland, y Fritz Heller, representante consular alemán en Detroit.

[Kyle Hessling @KyleHessling1]

Hello everyone!

I am pleased to finally release an incredibly long experiment that seems to have yielded solid results! As teased last week, using a geometry guided weight merge, I was able to restore the thinking stability of Qwopus v2, and merge it with the curriculum training delta of Qwopus coder!

While Qwopus coder without thinking is still probably my favorite model, I like to use it as an executor to a larger orchestration model, it is incredibly fast and will do pretty much anything you throw at it with thorough guidance.

But what if you want the whole package, the creative problem solving of Qwopus v2 with the tight execution of Qwopus coder? No planning model sitting above it? This is the ticket!

As you can see in the long winded model card, I ran this thing for weeks doing various agentic loops in Claude code, and it would carry everything to the end and always yielded a working result. Which I thought was impressive as Qwopus coder with thinking enabled would’ve gotten lost, and doing this in the coder without thinking would’ve probably worked but would’ve missed a lot of the feature nuance and creativity without being directly told to implement.

This gives you a solid comprehensive stack, and it runs at 100 tps on a single 5090 with MTP!

An interesting happy accident, is that somehow the resulting merged model is more temp stable than either base, it doesn’t seem to loop even ran down at temp 0.3, but in my testing it still seems to be more creative and performant at temp .85-1 as usual. However, if your harness forces a lower temp, it won’t be the end of the world with this model!

Anyways, this is a research proof of the custom geometry guided weight merge, and I’m elated to see how much control i can impose with it; it opens up a lot for future iterations as we can effectively modify, by delta, different portions of the weights without breaking another portion!

I really recommend giving the card a read because I had a blast putting it together and the theory is fun!

Please let me know how it does for you, it’s still definitely a research preview, but I’ve been hitting it hard for weeks and have not been able to find any glaring issues, it seems to be rock stable for agentic work in a fine tune!

As always I made a very different kind of game with it, called “orbital” it’s very hard to play and unconventional, but it’s quite aesthetically pleasing. The creativity of the concept is kind of a flex of the thinking capability, the gameplay is very original, with continued refining it could be awesome and unique, which is sometimes hard to get from a smaller local model without explicitly asking for it!

Blessed as always to be here with you all! Please hit me with all of your wins and critiques in the comments below! I will try to respond to everyone!

Keep shipping legends! Also, Jackrong has some really awesome stuff coming out in the next few days for the 9B and under folks, will be launching them here, so stay tuned!

https://t.co/xz7TzhRxIU

[Unsloth AI @UnslothAI]

Kimi K3 can now be run locally! ✨

The 1-bit model retains ~78.9% accuracy after we shrunk it from 1.56TB to 594GB (-62% size).

Run on a Mac Studio + 128GB RAM device.

Kimi K3 is the strongest open model to date.

Guide: https://t.co/1mVwOMLpDW
GGUF: https://t.co/bt1c1ADdCZ

[Forbidden HQ 🐇 @WearForbidden]

BREAKING: Using experimental quantum lip-reading AI, classified DARPA audio reconstruction tech, and a leftover NSA microwave mic we found in a dumpster, we exclusively recovered the FULL private conversation between Netanyahu, Pete Hegseth, and Marco Rubio at Lindsey Graham’s funeral.

Here’s exactly what was said:

[阿毛 @LmhdbAi]

长鑫科技是吴存荣在合肥当市委书记的时候投出来的，而且亏了10年合肥没有减持。

这老哥2008年的时候，拿市一半的财政收入豪赌京东方，也成功了

今年3月份因受贿1.27亿，被判无期徒刑

帮合肥赚了一万多亿，不知道能不能减2年刑😂

[Leopold Sv / 缙泽 @LestatSv]

@hyclove1 +1

[Leopold Sv / 缙泽 @LestatSv]

@chukuang2018 墙上挂的啥AI?