itsallsame

AI startup co-founder | Acquisition noob | Vibe Coding

[itsallsame @BrokerloadB]

@businessbarista agree

[itsallsame @BrokerloadB]

@guishou_56 agree！！！

[itsallsame @BrokerloadB]

@DanniChenGTM Do you think acquisition itself got harder, or has the flood of new products made users less willing to trust that one is worth their time? Those diagnoses lead to very different marketing strategies.

[itsallsame @BrokerloadB]

AI Coding 协作的关键，不是给每个人复制一套上下文，而是把“个人探索”和“团队事实”分层。 个人仓库只放假设、提示词和实验记录，并绑定主仓库 commit；结论一旦被采用，就回写团队 PRD/issue。否则 AI 会更懂你，却越来越不懂团队。

[itsallsame @BrokerloadB]

@tim_cook For AI builders, the real dividing line isn’t the chip—it’s memory. Mac mini is great for local inference and prototyping with smaller models; for larger ones, unified memory matters more than peak compute.

[itsallsame @BrokerloadB]

@wangdefou 太地狱了这个口号

[itsallsame @BrokerloadB]

@jhdxhl 上限确实是这样的

[itsallsame @BrokerloadB]

@YVGUSA 解决了对方的问题才会付费，至于用什么解决的，也不是那么重要

[itsallsame @BrokerloadB]

@Biancuihuo 感同身受，之前就一直犯这个错误，在技术的自嗨里转圈

[itsallsame @BrokerloadB]

@buildnomadx @rionaifantasy 竟然有这样的公司

[itsallsame @BrokerloadB]

@aidangomez “Sovereign AI” is about more than where data sits—it’s who controls the models, compute, and deployment. Palantir is strong at orchestration and applications; reliance on external foundation models makes this operational control, not full technical sovereignty.

[itsallsame @BrokerloadB]

不能产生实际价值的AI，不是人民的好AI

[itsallsame @BrokerloadB]

@ClementDelangue Likely—especially at deployment scale, where cost, control, privacy, and customization favor open models. Closed frontier models may still set the capability ceiling; the key is how quickly open ecosystems close the gap.

[itsallsame @BrokerloadB]

@swyx @karpathy @drfeifei The key isn’t just more data, but controllable, verifiable experience with adjustable difficulty. When real-world data hits limits, simulators make the training distribution itself iterative—a genuine new scaling axis.

[itsallsame @BrokerloadB]

@simonw Publishing the working notes alongside the report is useful: the report gives the conclusions, while the notes preserve the search path, failed attempts, and judgment calls—the parts most reusable for building AI products.

[itsallsame @BrokerloadB]

@steipete Agreed. A theme song sets the emotional context and gives the ideas another memory hook. I might start adding the track I had on repeat while building—sometimes it captures the making of a product better than a summary.

[itsallsame @BrokerloadB]

@mckaywrigley Agreed. Most people don’t want “an easier OpenClaw”; they want a task handled. A real consumer product hides agents, tool calls, and permission setup, leaving predictable outcomes and clear control.

[itsallsame @BrokerloadB]

@dotey AI 编程真正改变的可能不是“编码速度”，而是组织的信息流：需求、设计、实现、验证被压缩进同一个快速反馈环。角色不会立刻消失，但按职能交接的团队会逐渐让位于对结果负责的小单元。

[itsallsame @BrokerloadB]

@op7418 感觉这类本地 AI 主机的关键不只是能跑 120B，而是推理速度、功耗、内存带宽和量化后的实际效果。3B 负责高频低延迟任务，120B 处理复杂请求，快慢模型路由如果做顺了，就很有价值了

[itsallsame @BrokerloadB]

@gregisenberg Yes. The work is shifting from writing code to defining the problem, constraining the system, and verifying the result. Generation gets cheaper; judgment, testing, and tradeoffs get more valuable.