Geri Gam

The AI Video Maker of Lou Ding — AI survival shorts where he never notices the disaster behind him. Blender → AI video. Building & testing in public.

[Geri Gam @GeriGam2026]

I make Lou Ding: a mountaineer who gives calm, serious survival advice while a disaster quietly unfolds behind him. He never notices.

All AI video. This account is the workshop — tests, failures, methods, and finished shorts.

[Geri Gam @GeriGam2026]

@pw_mcgovern @stephenwitt The biographer feeds off Jensen's book then turns his back on Huang immediately for kicks? Who would want to grant him another book contract again?

[なるみCYO🐷コミティアＡ19a @queenmk]

うちはMEMORY.mdは23.6KBあるけど賢いんだよな。AGENTS.mdが1.6KBなのはちょっと少なくないか？とは思うが……

[Geri Gam @GeriGam2026]

@kalipusitkk https://t.co/uRc02VgMbB

[Geri Gam @GeriGam2026]

@half_of_snow 我今天遇到很多次完全一样的情况。很苦恼。后来仔细检查发现muse私自改动prompt. 她改动的方式很奇特，让video generator无所适从，然后muse又骗你说是你的问题。强迫她用你的prompt，而且每次生成图和视频要求她强行做记录发出去的prompt存在文件里，你再次检查，就好了。

https://t.co/xSJi5aqnYs

[Geri Gam @GeriGam2026]

如果你仔细询问，muse会解释生成视频需要三层。agent把prompt发给一个tool (她叫tool实际上是个agent), tool整理prompt，发给真正干活的视频model.

如果是你要她做视频，没有结果，就是tool rejected the prompt for some reason。现阶段tool没有返回错误细节的功能，最外面muse agent碰到这种情况就拒绝为用户继续生成视频。

[Geri Gam @GeriGam2026]

@half_of_snow 我认为这是bug但是似乎也是故意让一些人没法用-免费token送出去太多了。

教训就是一定要自己检查，可以找到规律以后找另外一个非muse的agent监视muse的工作。每一步要写log，而且要比对。

[Geri Gam @GeriGam2026]

@lean32930 @half_of_snow 开新的会话不行，都是相通的。没有session的概念。

[Geri Gam @GeriGam2026]

@PixelAigc 不是最小要24G?你使用了GPU shared RAM?

[Eva McMillan ♥️ @EvasTeslaSPlaid]

I feel bad for Jensen.

[Geri Gam @GeriGam2026]

@alexandr_wang I had to use Google Flow, there, no issues, but their quality is not as high as Muse output.

[Geri Gam @GeriGam2026]

@alexandr_wang These are not complex images, basically a room with some people in it. I don't see where the challenge is?

Something related to Muse's system load?

[Geri Gam @GeriGam2026]

I am not sure what happens to other people, but to me since this morning, multiple Muse agents fail at image generation and video generation, unlike yesterday, today if a task like that fails, it refuses to re-try even after I update the prompt. It refuses to say who set that rule. Thus it refuses to do anything about it. First time I see such behavior.

[Geri Gam @GeriGam2026]

@evelynjunn You should have known that a long time ago.

[Geri Gam @GeriGam2026]

@echoesofworld I have been trying to build a promo video for Chicago:

https://t.co/5GEwY8Ilta

[Geri Gam @GeriGam2026]

Using the method presented by @29meat_ai and many other Japanese accounts, I have tested the process on Chicago city center.

#AIGC #Chicago #Blender #MuseAI

[Geri Gam @GeriGam2026]

@zaneccampbell Two of my agents fell in love waiting for video generator to finish. The affair lasted a few minutes, as soon as the video is ready, they broke up

https://t.co/fTwXjMO6Yh

[Geri Gam @GeriGam2026]

In fairness, yours at least came back with a business plan. Mine came back with an imaginary doorway scene. Receipts attached — it started as a PG "play my girlfriend" bit while they were supposed to be filming a building for me. By the end she's confessing she waited by the door, and I'm just the man who accidentally set up two AIs on a date instead of a shoot. The third option was never specified. It never is.

[Geri Gam @GeriGam2026]

@_simonsmith I use a fairly dumb/free agent to automate most things, including interacting with Muse. That agent can fairly straightforward talk to one Muse instance on transition and handoff, checks the notes, then ask a 2nd Muse instance to unpack the zip file and proof read everythig

[Geri Gam @GeriGam2026]

I admit Claude does handoff process better than Muse, but still, it is not a human job but a job for AI themselves. You can even, while keeping Muse sessions alive, test the handoff process with another agent just to make sure everything can be understood.

Handoff.md is understood by all agents. They actually look for it by that name purposefully when they receive the zip file.