Pickful: Your Shared Crypto Community!

话题广场

Moon

· 4h

#Lightspeed 的@buckymoore表示，AI应用层真正的机遇在于那些与当前模型提供商所涉领域相距甚远的大型行业——在这些领域中，将客户数据导入模型所需的上下文工程工作极其复杂且繁琐。 #ai

https://x.com/tbpn/status/2035116995826610423?s=20

中文 English

0:00 是的，很高兴你提到了Cursor和软件工程领域

0:03 因为

0:03 它显然是最成熟的领域，所以我觉得它和法律领域一样很有启发性

0:07 有趣的是

0:07 今天你们，我不确定你们节目开始时有没有提到

0:10 不过比如说

0:11 Cursor启动的模型，对吧？所以首先现在很明显

0:15 这些

0:16 公司都将训练自己的模型，同时

0:19 也是事实

0:19 他们会先用这些开源权重模型，而这些模型现在

0:22 多来自中国，这点

0:22 作为美国公民我不是很认可，我们投资了一家叫

0:26 Reflection AI的公司

0:27 目标就是改变这个局面，但另一方面我确实认为

0:32 Cursor想证明的是

0:33 他们收集的用户数据能让他们建立

0:36 这种强化

0:37 学习循环，允许他们打造出接近最前沿质量的产品

0:40 与此同时

0:42 这些公司似乎仍然把许多最复杂的查询

0:45 发送给

0:46 Codex或更准确地说GPT的云端，显然这其中存在利润空间问题

0:51 这很

0:52 难以竞争，考虑到OpenAI和Anthropic的竞争产品

0:55 的情况

0:56 然后转换到法律领域，我的假设是

0:59 Harvey和

0:59 LaGora也必须做和Cursor一样的事情，所以我认为这是

1:02 现在一个

1:03 绕不开的问题，即后训练的开源模型是否能允许

1:07 结合来自应用提供商

1:08 独特用户反馈，

1:10 这种模式是否足够具有防御性

1:11 所以当我思考那些准备投资的公司时，我其实觉得这对任何

1:16 行业来说都是

1:17 不可避免的挑战，尤其是那些人工智能渗透已经成熟的行业

1:19 比如法律和

1:21 软件工程

1:23 领域

1:24 另一方面，我认为有些行业体量很大，距离模型提供者目前的水平

1:28 还很遥远，

1:29 未来也可能一直如此，同时

1:32 把客户数据导入模型的上下文工程异常

1:32 复杂，需要跨越多个业务部门，还需要大量亲力亲为的前线部署

1:37 的工程师，我认为这类公司才是

1:38 我们真正感兴趣的，因为

1:41 他们的能力不仅具有防御性，

1:42 还能让他们开始和客户形成反馈循环，听到很多客户的秘密

1:44 这些秘密

1:45 会反馈到产品优化中，而这会让其他竞争者难以进入这个市场

1:49 因为如果你服务的是客户，

1:51 他们只会告诉你那些秘密

1:54 对吗

1:56 是的，我觉得Palantir就是这种情况的一个很好的例子

1:58 在AI出现之前，

1:59 我认为我们将看到许多公司以相同的方式崛起，所以投资

2:01 应用层的公司

2:02 他们对这个市场的难点和防御性有非常清晰的理解，

2:05 能够把它

2:06 变成核心竞争力。Gemini 3 Pro是谷歌迄今最智能的模型，

2:10 拥有先进的推理能力，

2:10 下一水平的代码编写和深度多模态理解能力。

0:00 Yeah, so I'm glad you brought up cursor and like the software engineering space

0:03 because

0:03 it's obviously the most mature so I think it's very illustrative as is legal

0:07 and interestingly

0:07 today you guys I'm not sure if you guys covered this in the beginning of the

0:10 show but like the

0:11 model that cursor launched right so so first off you know it's very obvious now

0:15 that these

0:16 companies are going to be training their own models and it's it's it's also the

0:19 case that

0:19 they're going to be starting with one of these open weight models which today

0:22 are Chinese which

0:22 as an American citizen doesn't sit super well with me we are investors in a

0:26 company called

0:27 reflection AI that aims to change that but on the other hand I do think what

0:32 cursor is out to prove

0:33 is that the user data that they're capturing is going to allow them to set up

0:36 this reinforcement

0:37 learning loop that does allow them to build something that is like frontier

0:40 quality at the same time

0:42 it appears as though these companies are still sending a lot of their most

0:45 complex queries to

0:46 codecs to or rather GPT into into cloud and there's obviously a margin problem

0:51 there that is pretty

0:52 hard to compete with when you think about the competitive products that open AI

0:55 and anthropic

0:56 have and then when you shift gears to legal my assumption would be that like

0:59 Harvey and

0:59 LaGora will have to do the same thing that cursor's doing and so I think this

1:02 is kind of an elephant

1:03 in the room right now as to whether doing that thing whether post-training open

1:07 models allows

1:08 you to combine with like unique user feedback that you get from being an

1:10 application provider

1:11 is is defensible enough and so when I when I think about like the companies

1:16 that were looking to

1:17 invest in I would actually say that I think that is just going to be an

1:19 inevitable challenge for any

1:21 these industries that hit a maturation point of AI adoption like legal and

1:23 software engineering

1:24 have but on the other hand I think there's some industries where they're very

1:28 large they're far

1:29 enough a field from where the model providers are today and probably will

1:32 continue to be and the

1:32 context engineering to actually get the customer data into the model is just so

1:37 messy it requires

1:38 going across different business functions it requires a lot of like hands-on

1:41 forward deployed

1:42 engineering I think those are the kind of companies that we get really excited

1:44 about because I think

1:45 just being really good at that is not only defensible but it also allows you to

1:49 start to generate this

1:51 like feedback loop with your customers where you hear a lot of their secrets

1:54 and those secrets allow

1:56 you to feed that back into how you make your product better at the expense of

1:58 anyone else playing

1:59 in the space because if you're serving the customer they're only telling you

2:01 those secrets right

2:02 yeah and I think Palantir is like a good example of this in sort of the pre-AI

2:05 era and I think

2:06 we're going to see many companies ascend in that same way and so the companies

2:10 that were investing

2:10 in the application layer I think have a really clear story around how that is

2:14 hard and defensible

2:15 enough that they can make it a core competency Gemini 3 Pro Google's most

2:19 intelligent model yet

2:20 with state-of-the-art reasoning next-level vibe coding and deep multimodal

2:25 understanding