黑哥

CTO in Beijing, building AI into real content work. I use Chinese AI tools daily and post what English Twitter misses: tests, prices, news. 黑哥

[黑哥 @Blake488137]

Good timing for a new look. Apple started testing Apple Intelligence in mainland China on Nov 4 with the iOS 27.3 developer beta, and Qwen powers Siri, Writing Tools and Visual Intelligence there. Apple plans the public release for December, so many iPhone users here will meet Qwen through Siri.

[黑哥 @Blake488137]

@KekeSun9l Agreed. On that Aug 10 day, 2.64B of the 2.7B input tokens were hits. At DeepSeek's V4 Pro peak list price, that's $0.044 vs $1.32 per million, a 30x gap. Caching works on the start of the prompt, so one timestamp near the top can quietly kill most of those hits.

[黑哥 @Blake488137]

One real day of my AI agents, Aug 10:

2.7 billion tokens
15,180 requests
97.7% of tokens were cache hits
Output: 6.49 million tokens, 0.24% of input

My usage tracker priced the day at $906 at list rates.

Agents mostly read. They re-read the same files and history again and again, and barely write. That's why the cache price matters more than the headline price.

DeepSeek's current list for V4 Pro, per million tokens, off-peak / peak:
Cache hit: $0.022 / $0.044
Cache miss: $0.66 / $1.32
Output: $1.98 / $3.96

Off-peak is half price. Peak is weekdays 9 to 12 and 2 to 6 Beijing time. In New York that's 8 to 11 pm and 1 to 5 am.

My rule now: batch jobs that can wait run at lunch, at night or on weekends.

Source: DeepSeek API docs, Models & Pricing (checked Nov 4)

[黑哥 @Blake488137]

You can see that gap from Beijing in the price sheets. Zhipu's China API lists GLM-5.3 Flash at about $0.12 per million input tokens and $0.42 output. At prices like that, a lot of volume can still lose money on the GPUs. Volume is just the easiest number to show.

[Poe Zhao @poezhao0605]

"...validated the token-consumption economics that justify years of cloud infrastructure spending."

My March line on OpenClaw, an open-source AI agent, is another call the report corrects. I took a rise in use as proof of an economic result.

A filing from SiliconFlow, a Chinese independent token platform, later showed high volume and a large gross loss at the same time. Rising volume can happen with good economics or bad ones. It settles neither.

[黑哥 @Blake488137]

Freud covers part of it, but the split is very regional. In Ipsos's 2024 survey, 83% of people in China said AI products have more benefits than drawbacks. In the US it was 39%. Same tech, very different gut reaction, so exposure and daily use probably matter as much as wounded pride.

[黑哥 @Blake488137]

@tphuang Brazil is where BYD stopped just shipping cars in. The Camaçari plant in Bahia started assembly in October 2025 and passed 150,000 vehicles in its first year, per BYD's Brazil VP. Welding and paint lines are next, with local production targeted for early 2027.

[黑哥 @Blake488137]

@rohanpaul_ai That's how I run my agents in Beijing too. My 10 a.m. follower snapshot works that way: the API returns the counts, a script does the math, and the model just writes the summary. The p-hacking finding is the scary part. Locking the metric before the test starts is the guardrail.

[黑哥 @Blake488137]

@rohanpaul_ai For scale, China added about 250 GW of new generating capacity in the first nine months of this year, and about 170 GW of that was wind and solar, per the China Electricity Council. 1.7 GW is a huge buy for one lab and still under 1% of what China added in those nine months.

[黑哥 @Blake488137]

@testingcatalog The agent race looks similar from Beijing. DeepSeek said its Harness agent app hit 1.54M daily users by Nov 4, with about 35,000 third-party plugins, and it's now building an official plugin store. Curious whether Gemini Enterprise opens up to outside plugins the same way.

[黑哥 @Blake488137]

The tracker's raw count that day was 2,699,662,442 tokens. New input was only 53.1 million. The other 2.638 billion were cache hits, which is why the cache line on the price sheet is the one I actually watch.

[黑哥 @Blake488137]

Inner Mongolia is one of the eight national computing hubs China named in 2022. Officials told that hub to keep average PUE under 1.2 and use western wind and solar to support coastal demand, including the Beijing region. A campus that builds its own turbines is exactly that plan.

[黑哥 @Blake488137]

没点转账，钱也能被转走？

一次授权给出去，对方到底能动你多少钱？改密码能撤掉之前的权限吗？

这期借币安事件讲清链上授权，再带你检查微信、支付宝免密支付和 Apple 订阅。

看完顺手查一下，也转给家里管钱的人。

我是黑哥，不吹AI也不恐惧AI，只说真话。

#网络安全 #授权风险 #��问黑哥

[黑哥 @Blake488137]

@teortaxesTex Thinky's own table lines up with that. Inkling 2 Small is 17B active out of 594B total, DeepSeek V4.1 Flash is 16B out of 552B, and V4.1 Flash still edges it on the AA Index, 32.6 to 30.6. Same weight class, and DeepSeek is still slightly ahead.

[黑哥 @Blake488137]

@tphuang The base layer is already huge. MIIT says China had 5.254 million 5G base stations at the end of September, about 40% of all mobile base stations. Turning even part of that grid into radar is a very different scale from a lab pilot.

[黑哥 @Blake488137]

@teortaxesTex Ningxia fits that. It's one of the eight national computing hubs China set up in 2022, and Zhongwei's data centers already ran around 1.2 PUE by 2023, with about 86% of the city's installed power capacity in wind and solar. Solar wired straight to the racks is the next step.

[黑哥 @Blake488137]

@poezhao0605 The detail that still gets me: the chip shown at the 2003 launch had 144 pins, but the Hanxin-1 handed to the appraisal experts was a 208-pin part. The swap was set up before anyone tested anything. Your first question would have caught it.

[黑哥 @Blake488137]

@Da7_Tech Both, Hermes first. I've had it running as a hosted bot for a few weeks here in Beijing. The part that sold me is the scheduling: a scout bot drops one markdown file at 8:30 every morning, and my main agent picks it up from there. Hope your review covers cron jobs.

[黑哥 @Blake488137]

From Beijing desks this is already a workday meme. When DeepSeek drops mid-meeting, people switch to whatever still answers, then come back. Reliability is now part of the buying decision, not only price.

[tphuang @tphuang]

DeepSeek is down again is now a common occurrence on Chinese web & makes news everytime.

As its usage in China's business community continue to proliferate, DeepSeek is going to have to increase its reliability.

And ofc, that means having lower efficiency & utilization. You cannot guarantee higher up-time w/o having back up servers.

[黑哥 @Blake488137]

@poezhao0605 The Aptiv line in the same note is the tell. From Beijing, module suppliers (joints, controllers) are booking 2027 slots while full humanoid bodies stay demo-heavy. Buyers ask for the kit first.

[黑哥 @Blake488137]

@teortaxesTex From Beijing desks, when DeepSeek hiccups mid-day you feel it in the work chat first. Reliability is already part of the buying decision here, next to the China list price.