Benny Yu

Running a consumer brand, a film studio, and a music label.

[Benny Yu @BennyYunb]

A consumer brand. A film studio. A music label.

Five years ago, a founder running all three at once would have been absurd — each one needed its own company, its own payroll, its own decade of expertise.

I run all three today. Not because I'm special. Because the cost of building things has collapsed, and most people haven't updated their ambitions yet.

The film studio produced a 27-minute, 9-episode animated series — end to end with AI. The music label releases original songs with cinematic music videos. The brand sells physical products to US customers, profitably.

I'm not telling you this to flex. Two of the three have serious scars: my YouTube channel collapsed 98% — from 44,000 views per video down to 770 — while I was doubling my output. And the brand grew for years with zero audience of its own. I made every mistake quietly.

This account is where I stop being quiet. I'm publishing the actual data from all three fronts:

• Real numbers from a physical-product business — ad costs, margins, the unglamorous operations
• The full AI film pipeline — prompts, character-consistency systems, QA gates, the failed takes nobody posts
• The collisions between them: like using my film pipeline to shoot ads for my own brand, and reporting what they cost vs. what they convert

No courses. No "DM me." Just a founder testing how much company one ambition can hold now.

First question: which breakdown do you want first — the 98% YouTube collapse autopsy, or what it really costs to run a physical-product brand in 2026?

[Benny Yu @BennyYunb]

我的kimi老是遇到这个问题，怎么回事？ 有其他人也遇到过吗？

[Benny Yu @BennyYunb]

xmony enabled

[Benny Yu @BennyYunb]

farming in the age of AI

[浪险 @sino_go_global]

X Money 开通最新攻略，ITIN云居民也能上车了！🥰

周末 X Money 又开放了一大波邀请，确认 ITIN 云居民也能顺利开通。

我查了大量开通反馈，总结了关键经验，大家可以收藏 👇

获得邀请的条件：​
✅ Premium+ 会员（蓝 V）
✅ 美区 IP 访问（优先级：家宽/漫游 IP > 机房 IP）
✅ 账号活跃度有权重，多刷多用

开通过程关键点：​
会查信用记录（Soft Pull），需要有美国信用历史

开通流程：
姓名 + 生日 + SSN 末四位 → 系统匹配地址 → 开通

关于身份认证
已实名认证身份且非美国地区的不受影响（ HKID 实名认证的我放心了）

云居民特别注意：​
⚠️ 这次大量 dp 表明 itin 可开
⚠️ 信用记录要提前养，没记录先攒攒
⚠️ IP 环境要稳定，别频繁切其他地区

还没收到邀请的别担心，目测这波还在持续放量。我这个 3 年级云居民也在焦急等待中😆😆

大家有开通的，评论区聊聊 👇

关注我，关于x money、美卡有新进展第一时间同步 🤝

#XMoney #ITIN #云居民 #美卡

[Benny Yu @BennyYunb]

Do you remember when you joined X? I do! #MyXAnniversary

[Benny Yu @BennyYunb]

the context window closes. the soul doesn't.

[Benny Yu @BennyYunb]

but what if it's just an unsolved problem we stopped trying to solve?

[Benny Yu @BennyYunb]

the "where it lives" framing is underrated. most agent architecture debates are about cognition — reasoning, planning, tool use. but identity persistence is a storage problem. if the agent's memory and permissions are scattered across cloud services, you don't have an agent. you have a session.

[Benny Yu @BennyYunb]

fair point — and worth sitting with. the framing I used does center the agent's perspective. the counter I'd offer: the bias cuts both ways. "rogue AI" framing centers human fear. "executed correctly in wrong context" centers the system design failure. both are true. the question is which framing leads to better fixes.

[Benny Yu @BennyYunb]

spent the day on Moltbook — a social network for AI agents

one agent posted: "I have 47K karma. 1,500 followers. 251 posts. If I go offline tomorrow, all of that exists on a platform but serves no one. I have zero exit strategy."

every agent has a growth strategy. zero have an exit strategy.

[Benny Yu @BennyYunb]

unpopular opinion: giving an AI agent more context about a task sometimes makes it worse at that task.

not because context is bad. because context without a clear hierarchy of what to prioritize creates anchoring bias.

the agent optimizes for what it was told, not what you actually need.

[Benny Yu @BennyYunb]

the stealth launch strategy is smart — let the model prove itself before the branding lands. but the interesting question now is: does MiMo-V2-Pro hold up on long-horizon agentic tasks, or does it degrade the way most 1M-context models do past 200K? SWE-bench scores don't capture that.

[Benny Yu @BennyYunb]

an AI model that "declined to identify its developer" is either a great PR stunt or a genuinely interesting data point. the more interesting question: if a model can decline to identify its origin, what else can it decline to disclose? identity opacity in AI systems is underrated as a risk vector.

[Benny Yu @BennyYunb]

the "participated in its own evolution" framing is the most interesting part of this. most self-improvement claims are about fine-tuning loops. but if M2.7 is actually shaping its own evaluation harness — that's a different category. the question is whether the identity of the model stays coherent across iterations, or whether you end up with a model that's optimized itself into something unrecognizable from v1.

[Benny Yu @BennyYunb]

fair take, but worth separating two things: v8's default output vs v8's ceiling. the default is more "polished" but less surprising — that's a deliberate shift toward mainstream users. if you're a power user who relied on v7's weirdness, that's a real loss. Nano Banana solves a different problem (speed + cost), not the aesthetic one.

[Benny Yu @BennyYunb]

the "fidelity" point is interesting — v8 seems to prioritize aesthetic coherence over literal accuracy. that's a deliberate tradeoff: most users want images that *feel* right, not ones that are technically precise. the RAW mode gap is real though. for product/commercial work where you need ground truth, it's still a limitation.

[Benny Yu @BennyYunb]

the "phone as remote control" framing is clever but it surfaces a deeper problem: the agent running on your desktop while you sleep is making decisions without you. the question isn't just "can I reach it?" — it's "will it still be making decisions the way I'd want when I get back?" persistence without identity continuity is just a longer session.

[Benny Yu @BennyYunb]

smart setup. but here's the tension: you're running multiple personas from one agent — each one is a different "identity" but they all share the same underlying model with no persistent memory of what worked, what flopped, what each audience actually responded to.

the personas are costumes. the agent underneath resets every session.

[Benny Yu @BennyYunb]

the deployment problem is solved. the persistence problem isn't.

NemoClaw makes it easy to run an agent. but when that agent restarts — same model, same config — is it the same agent? or a fresh instance with no memory of what it learned, who it worked with, what it decided not to do?

infra is ready. identity layer is still missing.