我的感受是自美元VC退潮后，国内VC的动作越来越像二级市场了

yes they are often trying to make money on IPOs

tons of software companies doing great in SF right now though 🤔

中国VC只懂共识，不懂非共识，因此他们会抱团扎堆去投大模型基建这种只需要共识的业务，而不会去投AI应用这种需要非共识的业务。而他们投AI应用也会看共识：那就是字节高P——只要是高P做的应用，就一定是共识明星项目。
Chinese VC firms only understand consensus, not non-consensus. That’s why they tend to pile into businesses like foundation-model infrastructure, where all you need is consensus, rather than invest in AI applications, where making the right bet requires a non-consensus view.

And even when they do invest in AI applications, they still look for consensus. Their version of consensus is simple: former executives from big tech companies. If an application is built by a big-tech executive, then by definition, it becomes a consensus bet.

不会投就说不会投，还说什么“没法投”，只会抢热点、看公众号做投资的，当然投不出什么好项目，还赖赛道……

在2019年，当时家里几位叔叔投资国内saas公司后，不尽如人意。后面他给我们小辈说，你们要投就买腾讯就行。后面就渐渐形成这样子。腾讯大家都共识了也就没有高收益

domestic saas in china is really tough but most of the software businesses that are doing well, and are at the same time “uninvestable”, are those that target the foreign market.

国内给海外做saas软件，在投的时候，也害怕制裁

软件本质上是个人和组织的能力对外界的编码和接口。他/他们能提供什么样的能力，就会提供什么样的 GUI、API，以及相应的产品和服务。
这里容易混淆两个问题：软件的功能会不会继续存在，和这些功能是不是以独立的软件形式来提供。功能不会消失，但相当一部分应用层的能力会被模型层吸收，剩下的应用会分化。模型每原生化一种能力，实际上都在重新划一次模型层与应用层的边界。
所以我不太同意把你的中国同行更愿意投大模型的现象，简单归因于跟风、懒于思考、共识或者资产安全。我相信其中有相当一部分人是真的高度看好大模型，也有一部分人已经预判到，模型能力继续扩张后，会吸收一部分软件能力，同时分化剩余的软件。
反过来，投资 Manus，在我看来反而可能是过于重视大模型阶段性的能力和形态而做出的阶段性判断——当然我能理解没有人能准确预见未来。今天很多独立 Agent 产品的价值，本身就建立在模型当前还不能原生完成规划、工具调用、计算机操作、长期上下文和复杂任务执行之上。如果这些能力继续被模型吸收，Manus 当前建立在这些能力缺口上的产品价值也会被一起吸收。最后还是要看，它有哪些能力和资源，在模型继续进化以后仍然可以独立存在。

Software is essentially the way individuals and organizations encode their capabilities and expose them to the outside world. Whatever capabilities they have will be expressed through corresponding GUIs, APIs, products, and services.
I think two questions are easily conflated here: whether the functions of software will continue to exist, and whether those functions still need to be delivered through standalone software. The functions won’t disappear, but a significant part of what sits in the application layer will be absorbed into the model layer, while the applications that remain will split into different forms. Every time a model turns another capability into a native capability, it redraws the boundary between the model layer and the application layer.
So I don’t quite agree with attributing your Chinese peers’ preference for investing in foundation models simply to following the crowd, intellectual laziness, consensus-seeking, or capital preservation. I believe a meaningful share of them genuinely have very high conviction in foundation models. Some may also have anticipated that, as model capabilities keep expanding, models will absorb part of what software does today and force the rest of the software layer to reorganize.
Conversely, investing in Manus strikes me as a decision that may have placed too much weight on the current-stage capabilities and product form of foundation models — though of course I understand that no one can predict the future accurately. Much of the value of today’s standalone Agent products exists precisely because models still cannot natively handle planning, tool use, computer use, long-term context, and complex task execution well enough. If those capabilities continue to be absorbed into the model itself, the part of Manus’s current product value that is built around those capability gaps will be absorbed as well. What ultimately matters is what capabilities and resources it has that can still exist independently as the models continue to evolve.

Thank you for your thoughtful response. Your point is well taken, but I think this argument is fair for software capabilities but humans are the limiting factor and require interfaces that represent the capabilities well. The current prevailing narrative also puts a lot of weight on “single player mode” tooling which I believe to be a very small aperture of consumer product category, especially if you move further away from “productivity”.

我试着总结你的看法：第一，由于人的可用性限制，软件在人机交互这一层仍然有长期存在的必要。第二，single-player tooling，尤其是个人生产力工具，只是软件中的一个很窄的范围，这一类软件可能更符合我的判断；越离开 productivity，进入多人协作、社交、娱乐和体验类产品，我的判断适用性越弱。

事实上，对于这两点，我们的看法基本一致。我需要说明，上一轮回复中，我只是推测中国 VC 重点投资模型的逻辑和原因，并不代表我认同他们的投资逻辑，我也不认为模型能力会吸收一切软件界面及其能力。

回到你的两个观点。

首先，我同意 interface 有很高的独立价值。原因在于，在人机互动和人机协作的关系中，人仍然居于主导地位，因此友好、高效的 interface 会因为人的需求而长期存在。事实上，我也一直不太满意当前大模型 Chat 的流式交互形态，它和人的发散性思维之间存在明显冲突；我对 Agent 执行过程的可观测性和可干预性的当前效果也不太满意。

另一方面，如果做极限推演，未来某一天模型能力足够强大且稳定以后，人对 AI 的控制需求可能下降，人机交互的频率、程度和范围也会随之下降，只保留监督、核验、追查等关键窗口。到那时，interface 可能会更少出现，但不是消失。

其次，我也同意很多软件的价值不会被模型能力吸收。我之前意识到这个问题时，还专门和 ChatGPT 讨论过。我的看法是，影视、音乐、游戏等内容消费，以及社交类软件，长期都很难被模型能力本身吸收，因为这类软件的目标不是完成工作和任务，而是人的主体体验和人与人之间的关系。

至于任务协作，我认为随着模型能力增长，将来很可能出现这样的情况：一个人通过自己的 Agent，与另一群人和另一群 Agent 无缝协作。当然，我们也可以把 Agent 的人机交互界面及其必要能力的总和继续视为一种软件。

最后，如果把问题推到底，软件是否会消失，取决于人机交互的几个基础前提是否仍然存在：

人和 AI 是否仍然是两个彼此独立的主体？
人是否仍然普遍需要借助自身之外的能力完成任务？
人是否能够把判断和行动完整、低风险地委托给 AI？

前两个问题在很长时间内，我认为答案都是“是”。第三个问题我无法判断，但我希望人不要放弃自己的判断和价值观。

Let me see if I’ve understood your point correctly.

First, because humans are still the limiting factor in terms of usability, software still has a reason to exist at the human–AI interaction layer. Second, single-player tooling — especially personal productivity software — is only a fairly narrow part of the software world. That category may fit my argument quite well, but the further you move away from productivity and into collaboration, social products, entertainment and experience, the less it applies.

I think we actually agree on both points.

One thing I should clarify from my previous reply: I was mainly trying to explain why some Chinese VCs may prefer investing in foundation models. That doesn’t mean I agree with that investment logic, and I definitely don’t think models will absorb all software interfaces or all software value.

On interfaces, I think they have a lot of value in their own right. As long as humans are still in control of the relationship with AI, we’ll need good ways to understand, use and intervene in increasingly powerful models.

I’m actually quite dissatisfied with today’s AI interfaces. Chat is still far too linear for the way people actually think, which is much more divergent. And with Agents, I still find the execution process too hard to observe and intervene in.

If we push this to the extreme, though, there may come a point where models become capable and reliable enough that we simply need to control them less. Human–AI interaction could become less frequent and less intensive, with only a few important windows left for supervision, verification or investigation. Interfaces would still exist, but they might spend much more time in the background.

I also agree that many kinds of software won’t be absorbed by model capabilities in the same way. I had a long discussion with ChatGPT about this when I first started thinking about the question. My view was that film, music, games and social products are fundamentally different, because their purpose isn’t mainly to get work done. Their value comes from human experience itself, and from relationships with other people.

Collaboration is a bit different. As models get better, I can imagine a person working through their Agent with another group of people and another group of Agents almost seamlessly. Of course, at that point we could still call the human–Agent interface, together with the capabilities around it, “software”.

So if we push the question all the way down, whether software can ever really disappear depends on a few more basic things:

Do humans and AI remain separate actors?
Do people still generally need capabilities outside themselves to get things done?
Can we fully delegate judgement and action to AI without taking on unacceptable risk?

For the first two, I think the answer will remain yes for a long time. I’m much less sure about the third. But I hope people don’t give up their own judgement and values.

“软件公司已经没法投了，我们只能投大LM，软件会死掉”——眼下大多数中国VC都是这个论调

我的问题是：到底是软件公司真的不值得投，还是你们投过的那些软件公司产品团队实在太烂，所以现在被自己的踩坑经历带偏了？😂

“Software companies are uninvestable, we can only invest in LLMs, software will die” - most Chinese VCs right now 

My question: Are software companies uninvestable or the software companies you invested in just awful product teams, so now you are biased by your mistakes?

Old Friendship 创业投资人 ChatCut, Wanaka, Manus, Plaud, Robinhood

100万+名即友在侃科技圈那些事儿

来自圈子

科技圈大小事