This week on the channel +672K views this week

Asianometry looks past GLM's models to the cost of making them

The Zhipu profile anchors a six-upload week that also ranges from Edison's electrical problems to an antibiotic collected in Indonesia.

Channel:

Asianometry Story of November 4, 2026Trari Newsroom

Why this story? The channel's weekly summary: what went up this week, what is drawing views now, and where the channel stands among similar channels.

The Rise of GLM and China’s Zhipu AI

Things we can learn from the video

Press to jump

0:34Zhang Peng's work at Tsinghua's Knowledge Engineering Group and on AMiner provides the company's academic starting point.
4:28OpenAI's GPT models push Zhipu toward building bilingual language models rather than relying on traditional NLP pipelines.
10:19The team trains GLM-130B on Nvidia A100 servers and open-sources it in 2022.
16:48Competition in consumer chatbots prompts a shift toward enterprise customers.
20:00GLM-4.5 brings an open-weight mixture-of-experts model to the series.
22:04Asianometry questions the sustainability of research spending that exceeds twice the company's revenue.

Building a capable language model is one challenge. Paying to keep building it is another, and Asianometry's Zhipu AI profile ends on that less glamorous question.

In "The Rise of GLM and China's Zhipu AI," the narrator follows the company from academic research into commercial products and enterprise customers. By the conclusion, the emphasis has shifted from technical achievement to financial sustainability: Asianometry says Zhipu spends more than twice its revenue on research and development while making models freely available. That is the tension behind the company history, not just another comparison of chatbot answers.

The profile was the most-watched of six new uploads, reaching about 140,000 views. In the comments, coding users exchange experiences with GLM, while others ask for more histories of AI labs focused on their operations rather than their models. One viewer picks up the same question as the conclusion: can strong research become a business customers depend on?

Elsewhere, "Silicon Valley Walks the Jagged Edge" drew the week's next-largest new audience. Two Edison titles, a computer purchase that pushed TCS into exports, and "The Special Antibiotic Collected in Indonesia" supplied the channel's characteristic detours beyond AI.

The audience grew modestly, adding about 1,000 subscribers. Daily viewing is currently around half its usual level, although that describes today's pace, not the whole week. Nearby education channel BBC Timestamp is having a busier-than-usual day.

Zhipu remains Asianometry's busiest upload right now. Behind it, "Edison's Half Mile Problem" and the slightly older "The Lintel Conquest of UNIX" keep electrical and computing history in the mix.

Published on Trari

Where next?

Also in Explainers, history and educationVideo taking off 416x its usual first dayJon Bogard asks whether feeling good is enough reason to giveHis TEDx talk challenges donors to compare what their generosity achieves, even when the most effective choice feels less rewarding.Channel: TEDx Talks · Story of November 4, 2026

A different discovery todayVideo taking off 70x its usual first dayTalarico, Ossoff and North Carolina's flip run through AP's election-night streamThe Associated Press put its race calls and balance-of-power maps on a single live feed, with cameras inside campaigns in Texas, Michigan and Georgia.Channel: Associated Press · Story of November 4, 2026

What viewers say

Is this Asianometrys AI era? Would be great to have primers on the big labs, foucsing on their history + operations rather than just the models like everyone else
@tanjeeschuan4999See it on YouTube
Nice video.
After programming with GLM for two months, I can say: I like the humor in GLM 5.x, but it doesn’t know it’s being funny.
“Oops, ended up on the wrong branch again. Cleaning up my mess.” :D
@MeffenOneSee it on YouTube
Watching this while Claude burns my last tokens before weekly reset. 😊
@Tential1See it on YouTube

3 more comments

Zhipu’s challenge is turning strong research into a sustainable business. Enterprise customers may offer steadier revenue than consumer AI, but heavy R&D costs can still outpace adoption. The key question is whether GLM can become infrastructure businesses depend on, rather than just another competitive model.
@SiliconNodeReportsSee it on YouTube
The GLM idea of blending BERT's fill-in-the-blank approach with GPT's left-to-right generation was the detail that stuck with me — it helps explain why they thought it trained more stably. Fascinating too that this grew out of AMiner and Tsinghua's 'Paper to Product' culture, rather than starting out as a model company at all.
@ExplainedIn60HQSee it on YouTube
What a coincidence, I was watching your video while I was using GLM 5.3 to train my model.
@mandateonmandalay7632See it on YouTube

Public comments from the video's page on YouTube