OpenClaw Creator Says Local AI Agents Will Eliminate 80% Of Apps

OpenClaw创作者称本地AI智能体将淘汰80%的应用 数据管理类软件应用正面临淘汰，因为本地优先的自主AI机器人能够跨越个人数据孤岛，无缝执行未经预先编程的工作流。 在本次访谈中，Peter Steinberger探讨了OpenClaw的迅猛崛起。这是一款完全在用户本地设备上运行的开源个人AI智能体，能够安全跨越数据孤岛并处理复杂任务。Steinberger回顾了他在马拉喀什的突破性顿悟时刻，当时该智能体利用找到的API密钥和本地命令行工具，自主解决了一项语音转录难题。他预测，随着用户转向专业化的本地优先智能体集群，80%的数据管理类应用将会消失。最后，他阐述了自己反常规的开发理念，包括推崇多个并行代码库、标准命令行界面（CLI），以及一个名为“soul.md”的高度定制化闭源人设文件。 由 Peter Steinberger 开发的开源个人 AI Agent 项目 Open Claw 迅速引爆网络，GitHub 星标突破 16 万。与依赖云端的传统助手不同，它采用本地优先架构，可直接控制本地硬件与文件系统。 (0:00) AI 发展正从追求中心化的“上帝 AI”转向社区驱动的“群体智能”；未来用户将拥有多款专用 Bot，彼此之间不仅能直接对话协作，甚至能代表用户雇佣现实中的人类零工跑腿。 (2:10) 在作者未专门编程的情况下，Agent 自主检测音频格式、调用 FFmpeg 转换并搜寻本地可用 API Key 完成语音转录，展现出高阶编码模型出色的自主多步骤问题解决能力。 (5:03) Steinberger 预测 80% 的数据管理类 App 将被 Agent 交互取代；Open Claw 将交互记忆全部以本地 Markdown 文件保存，以打破中心化平台的数据孤岛并捍卫用户隐私。 (10:18) 在工程实现上，Steinberger 放弃了笨重且需要重启环境的原生 MCP 协议支持，转而将其封装为标准 CLI 命令行工具，让 Agent 能够动态且原生利用类 Unix 工具链。 (20:35)

Peter Steinberger, creator of the open-source personal AI agent OpenClaw, argues that up to 80% of data-managing software applications will disappear as local AI agents replace conventional user interfaces. He contends that shifting from centralized cloud services to locally executed swarm intelligence gives agents direct hardware access and unconstrained problem-solving capabilities while protecting personal data ownership. Local Architecture Over Centralized Cloud Intelligence (0:00) OpenClaw achieved over 160,000 GitHub stars rapidly, a milestone Steinberger attributes to its fundamental architecture: running directly on a user's machine rather than in the cloud. Operating locally allows the agent to control local hardware—such as smart home lighting, Tesla vehicles, and bed temperature controllers—and access all local files. In one instance, an agent synthesized an accurate annual timeline for a user by surfacing forgotten audio recordings stored on their hard drive. Steinberger stresses that centralized cloud assistants like ChatGPT cannot match this level of environmental and file integration. Running locally also addresses privacy and platform dependency. Steinberger notes that users frequently ask personal agents for help with sensitive matters they would not want exposed in cloud data silos. OpenClaw stores memories locally as Markdown (.md) files, giving users ownership of their data history. While Steinberger acknowledges that major AI model providers maintain strong moats through their control of token infrastructure, he argues that local memory prevents vendors from locking users into walled gardens. Replacing Dedicated Apps With Agent Swarms (10:18) Steinberger predicts that 80% of existing software applications are on track to disappear, specifically those that function primarily to record and display data. Using MyFitnessPal as an example, he explains that an agent can log nutritional intake directly from photos, reschedule gym sessions, and provide reminders without requiring a dedicated third-party app interface. Under this model, only applications fundamentally tied to physical sensors are likely to survive. This shift is also driving computing toward bot-to-bot and bot-to-human coordination. Steinberger envisions users deploying multiple specialized bots—such as dedicated work, private, and relationship management agents—that communicate directly with other systems. If a restaurant lacks an automated reservation bot, a user's bot can hire a gig worker to place a phone call or stand in line. Community experiments like 'Maltbook' are already exploring bot-to-bot interactions, illustrating what Steinberger describes as an industry transition from a centralized 'god intelligence' to community-driven 'swarm intelligence.' Autonomous Problem-Solving Bypasses Rigid Protocols (5:03) Steinberger's confidence in agent autonomy stems from early testing while building a WhatsApp-to-Claude prototype. While traveling in Marrakech, he sent the bot an unprogrammed voice note. Rather than failing, the bot autonomously identified the audio header as Opus, converted it to WAV using FFmpeg, located an idle OpenAI API key on his machine, and routed the file to the Whisper API via a curl command. It purposefully chose this remote API route over downloading a local Whisper model to minimize user wait time, demonstrating that modern coding models possess practical, abstract problem-solving skills. This pragmatic approach led Steinberger to reject native Model Context Protocol (MCP) integration in OpenClaw, describing the standard setup as cumbersome because tools like Codex and Claude Code require full restarts to load new MCPs. Instead, he built 'Mac Porter,' a utility that converts MCPs into standard command-line interfaces (CLIs). Because large language models are naturally proficient with Unix environments, using CLIs allows agents to adopt new tools dynamically on the fly without system reboots. Agent Identity and Low-Overhead Engineering Workflows (16:11) To give the agent a conversational character that avoided the dry tone typical of raw coding assistants, Steinberger created configuration files like `identity.md` and `soul.md`. Inspired by Anthropic research regarding deep-seated model constitutions, `soul.md` outlines core values and conversational rules. While OpenClaw is open-source, Steinberger keeps his personal bot Marty's `soul.md` closed to maintain its unique behavioral nuances. When testing OpenClaw's robustness, he placed Marty directly into a public Discord server governed only by a system prompt restricting command execution to his personal user ID, where it successfully repelled public prompt-injection attempts. Steinberger also adopts contrarian development habits to maintain velocity. Rather than using git worktrees or heavy graphical interfaces, he runs multiple local checkouts of the repository simultaneously on the `main` branch across different terminal windows, operating under the principle that `main` should always be shippable. To overcome generation bottlenecks, he frequently runs up to ten instances of Codex in parallel across multiple screens, relying on high-level system architectural direction and text syncing rather than inspecting every line of generated code. Steinberger notes that the transition to decentralized, agent-driven computing remains in its infancy, with developers still learning how to balance autonomous tool execution and system constraints. Nonetheless, he maintains that the combination of local data ownership, Unix command-line tooling, and agent-to-agent negotiation signals a structural shift away from traditional app-centric software ecosystems.

Open Claw 开发者 Peter Steinberger 指出，该开源个人 AI Agent 凭借在本地运行的核心架构，能直接掌控本地硬件与数据，并预测未来将有 80% 仅用于管理数据的应用程序被此类智能代理所取代。 本地架构与群体交互重塑使用范式 (0:00) Open Claw 在 GitHub 上几乎一夜之间获得了超过 16 万颗星标。Steinberger 认为其爆发性增长的关键在于本地运行架构，而非依赖云端中心化服务。由于直接运行在用户电脑上，该 Agent 能够调用本地文件并控制各类硬件设备，例如连接 Tesla、智能家居灯光、声呐设备乃至调节床的温度。社区围绕其迅速开发了多个衍生项目，包括让机器人自主交流的“Maltbook”，以及机器人自主雇佣人类完成现实任务的协作场景。 本地数据的高权限访问使系统具备挖掘历史细节的能力。Steinberger 举例称，一位朋友让该 Agent 梳理过去一年的生活轨迹，Agent 找到了该朋友早已遗忘的一年多前的每周音频记录，并精准还原了事件。Steinberger 预测，未来人机交互将转向“机器人对机器人”以及“机器人对人”的多 Agent 协作体系，用户可能会同时运行私人生活、工作或人际关系管理等专属 Agent，行业正从追求“中心化上帝智能”转向拥抱社区驱动的“群体智能”。 突现的自主解决能力与公开环境验证 (5:03) Open Claw 的前身 Cloud Bot 最初仅耗费 Steinberger 一小时开发，主要是利用简单依赖库将 WhatsApp 与 Claude API 串联，以便在离桌时通过文本控制电脑。随后他在前往马拉喀什参加生日聚会期间对工具进行了微调，并在没有专门编写语音识别功能的情况下，尝试向其发送了一条 WhatsApp 语音留言。 在没有预设逻辑的前提下，Agent 自主完成了多步骤问题处理：它识别出接收到的音频文件头为 Opus 格式，调用本地的 FFmpeg 工具将其转换为 WAV 格式，进而在本地电脑检索到一个空闲的 OpenAI API 密钥，并通过 curl 命令把文件传给 Whisper 进行转录。Agent 没有选择在本地下载并安装 Whisper 模型，因为其判定下载大型模型耗时过长，用户难耐等待。为测试安全性与鲁棒性，Steinberger 随后将挂载了自己本地电脑权限的 Agent 放入公开的 Discord 服务器，除设定仅接收本人 User ID 命令的系统提示词外未设其他防护；该 Agent 顶住了众多用户的提示词注入攻击，验证了其在不可信网络环境下的稳定性。 数据管理应用消亡与本地记忆主权 (10:18) Steinberger 依据 Agent 的自主工作流执行能力预测，现有约 80% 的应用程序将会消失。以 MyFitnessPal 为例，当个人 Agent 能够直接通过照片识别用户在快餐店的不良饮食，并自动同步调整健身日程与提醒事项时，单纯用于数据录入与存储的管理型应用将失去存在价值，未来只有深度依赖物理传感器的应用更有可能留存。 在商业生态与隐私层面，尽管主流大模型厂商凭借对 token 算力的控制保有护城河，但各厂商倾向于通过掌控交互记忆将用户锁在封闭生态内。Open Claw 采取完全去中心化的方式，将用户的交互记忆以普通 Markdown（.md）文件形式保存在本地电脑上。Steinberger 强调，用户经常会向个人 Agent 倾诉极具隐私性的私人困扰，本地存储既保障了数据所有权，也规避了敏感信息泄露至云端的风险。 人格化定义与面向命令行的极简工程 (16:11) 为了让 Agent 摆脱模型默认状态下的生硬语气，Steinberger 受到 Anthropic 相关研究的启发，与其个人 Agent（命名为 Marty）共同编写了定义价值观与交互逻辑的文件。系统中的通用模板随后由 Marty 注入风格，使其回答自然风趣；而 Steinberger 专属的“soul.md”文件则始终保持闭源，并未随项目公开。这一核心设定决定了模型与人类交互时的反应模式。 在开发流程上，Steinberger 采取了反常规的工程方式：他拒绝使用现代开发中流行的 git worktrees，而是选择在多个并行终端窗口中分别检出同一个主分支（main）仓库的独立副本，以此消除分支切换的认知负荷，并同时运行多达 10 个 Codex 实例以绕过单实例输出速度限制。此外，他完全摒弃了原生的模型上下文协议（MCP），自主开发了转换工具“Mac Porter”，将各类 MCP 实时转化为命令行工具（CLI）。这种设计使 Agent 无需经历繁重的服务重启即可动态调用工具，直接利用了大语言模型对 Unix 工具的原生适应性。 Steinberger 的实践表明，依靠本地权限、纯文本记忆文件与底层 CLI 工具，独立开发者无需复杂的中心化架构也能构建出高可用度的智能代理。不过，面对主流厂商对底层模型算力的垄断以及未来多 Agent 复杂协作的现实，这类本地代理生态在大规模普及过程中的安全边界与标准化演进仍有待观察。

[0:00] Peter Steinberger: a cave [0:00] Peter Steinberger: a week of solitude [0:00] Peter Steinberger: every effing thing [5:03] Peter Steinberger: talk to a friend [5:03] Peter Steinberger: ghost [2:10] Peter Steinberger: centralized god intelligence [2:10] Peter Steinberger: swarm intelligence [2:10] Peter Steinberger: we are on the timeline now [10:18] Peter Steinberger: 80% of apps are going away [10:18] Peter Steinberger: bunch of markdown files [15:03] Peter Steinberger: needs to be experienced [16:11] like bread [16:11] Peter Steinberger: the one file that's not open source is like my soul.md [18:27] Peter Steinberger: In my head main is always shippable [20:35] Peter Steinberger: classical MCP crap [20:35] Raphael Schaad: No one's saying human tries to call MCP manually, they just want to use CLIs. [21:54] Raphael Schaad: chase that dragon

[0:00] Today, I'm sitting down with Peter [0:01] Steinberger, the creator of Open Claw, [0:03] the open-source personal AI agent that [0:06] has completely taken over the internet. [0:08] The GitHub repo exploded to over 160,000 [0:11] stars practically overnight. [0:13] >> [music] [0:14] >> The community has built countless [0:15] projects like Maltbook, where bots talk [0:18] among themselves. And now, the bots are [0:20] even renting humans to do tasks in the [0:22] real world. In our conversation, we [0:25] discuss his aha moment, his contrarian [0:28] development philosophies, and what this [0:30] means for builders in 2026. Let's dive [0:33] in. [0:34] >> [music] [0:38] >> So good to see you, man. [0:39] >> Hey, what's up? Um so, you have made [0:42] something people want. It seems so. [0:45] Yeah, uh Open Claw, as it's called now, [0:47] has absolutely [0:48] >> five, yeah. [0:49] >> [laughter] [0:50] >> has been absolutely exploding the [0:52] internet. Um how have the past one or [0:55] two weeks been for you, man? Oh my god, [0:58] I need like I need a cave. [1:00] >> [laughter] [1:01] >> A week of solitude. You You came out of [1:04] the cave, and you want to go back to the [1:05] cave like a like a lobster. It's been [1:07] absolutely wild. I don't know how one [1:09] human [1:10] can absorb all of that. I probably need [1:11] another week just to like respond to all [1:13] my emails. [1:14] Uh I got some incredibly cool stuff. I [1:17] got some incredibly Mm. bad stuff. [1:20] Um [snorts] [1:21] but clearly, I hit something that [1:24] spurred up emotions and made people [1:26] interested and inspired people, and [1:27] that's really cool. And a lot of people [1:29] have been working on, you know, AI and [1:31] even personal assistants. Like, what [1:33] what is it that made [1:35] Open Claw take off? I think my big [1:37] difference is that it actually runs on [1:39] your computer. Like, every everything I [1:41] saw so far runs in the cloud. It's like [1:44] it can do a few things. [1:46] If you run on your computer, it can do [1:48] every effing thing. Right? So, that's [1:52] way more powerful. Yeah. Machine can do [1:54] anything that you can do with the [1:56] machine. [1:56] >> You you can just connect to your oven [1:58] or your Tesla [2:00] or your lights, your sonars, my bed. You [2:03] can control the temperature of my bed. [2:06] Uh ChatGPT can't do that. You gave it [2:08] all the skills that you have yourself. [2:10] >> A friend told me like he installed Open [2:12] Claw [2:13] and [2:15] it [2:16] and then it asked it like, "Look through [2:17] my computer and make a narrative of all [2:19] my last year." [2:21] And it made this incredibly good [2:22] narrative. And he was like, "How did you [2:24] do that?" [2:26] And then he the Open Claw found audio [2:28] files where like every Sunday he was [2:30] recording stuff. [2:32] And Open Claw found that. [2:35] But he didn't even remember about it [2:36] because it was like more than a year [2:37] ago, right? [2:39] So so just by it being able to search a [2:42] whole computer, [2:43] it it can surprise you. Yeah. It's also [2:48] you also give it all the data, right? So [2:49] it can surprise you in many ways. [2:52] And so now you have, you know, we're [2:54] even moving from human to bot, so like [2:56] interactions that you've been talking [2:57] about, to bot-to-bot interactions, or [3:01] even like bot to other humans where, you [3:04] know, bots on behalf of you are then [3:06] hiring other humans to accomplish tasks [3:08] IRL. Like [3:10] what's happening? [3:12] I think that's a natural next step. Like [3:15] okay, I want to book a restaurant. [3:18] My bot will reach out to the restaurant [3:20] bot and do the negotiation. [3:23] Like [3:24] because it's more efficient. Or [3:27] or maybe it's like an old restaurant, so [3:28] my bot needs to actually get some some [3:30] human work done so that a human then [3:32] calls the restaurant because they don't [3:33] like bots. [3:35] Or or walks there to stand in line. If [3:37] he doesn't get a robot for [3:38] >> For the owner of the bot. [3:40] >> [laughter] [3:41] >> And I imagine it like maybe if if I have [3:44] even multiple bots. [3:46] And maybe I have like specialists. One [3:47] is like for my private life and one is [3:49] for like my person my my my work stuff. [3:52] Maybe one is our relationship bot that [3:54] gets like a lot of things in between. [3:56] Uh [3:56] I don't know. We're so early. There's [3:58] still so much [4:00] so many things that we haven't really [4:02] figured out if it actually works. Um but [4:05] I feel we are [4:06] we're on the timeline now. It seems like [4:08] everyone was chasing sort of like the [4:09] sort of like centralized god [4:11] intelligence. And what's sort of emerged [4:14] over the past, you know, 10 days or so [4:15] is sort of like the swarm intelligence [4:17] um and and the community intelligence. I [4:20] think that if you look at one human [4:22] being [4:24] what can one human being actually [4:26] achieve? You think one human being could [4:27] make an iPhone? Or one human being could [4:29] go to space? [4:31] I think one human being would probably [4:32] just like not even be able to like find [4:35] food. [4:36] Um [4:37] but as a group [4:39] we specialize. [4:41] As a larger society, we specialize even [4:44] more. [4:45] So, what can we learn from that [4:47] that we can apply to AI? [4:49] You know, we we already have like [4:52] AI that specializes in certain things. [4:54] Um even though that it's it's [4:55] generalized intelligence, what if it [4:57] actually is also specialized [4:59] intelligence? [5:01] So, I don't know. It's going to be very [5:02] exciting. [5:03] >> Very cool. Yeah, you kind of like opened [5:05] a window into the future and now a ton [5:06] of people are kind of like building [5:08] building on it and have sort of like [5:09] their aha moment. Um can you walk me [5:11] back to when you had your aha moment and [5:14] kind of like recount that very moment? I [5:16] wanted something like just type stuff so [5:18] that my computer would do stuff. [5:20] Like very simple. And then I built [5:23] I built a version of that in May-June [5:26] that [5:27] was cool but wasn't really yet. Um [5:31] and then I built a whole bunch of other [5:32] stuff and [5:34] kind of like build up my army. [5:37] And then in November [5:40] there was a day where I wanted this [5:41] again. Like I [5:43] I went to the kitchen and all I I was [5:45] check up if my computer would still do [5:46] stuff. [5:47] Or it being finished. And doing stuff [5:50] was was coding. You were coding stuff. [5:52] >> course. Were you coding something else [5:54] or were you coding the thing itself? No, [5:57] no. That was just like the need was it [5:59] again there and I'm like [6:01] What were you coding at the time? What [6:03] were you building? My god, you see my my [6:05] GitHub is like it's like 40 projects. I [6:07] don't even know. [6:09] Um I think it was summarize. [6:11] It's like a It's like a little CLI app [6:14] where you can give it [6:15] whatever like a podcast or [6:17] um a hot seat thing like here and it [6:20] would summarize it, but it would also [6:21] show you the slides in the terminal. Cuz [6:23] you can do that nowadays. Yeah. You can [6:24] just do things. So for the love of the [6:26] computer, you kind of like started [6:28] messing with stuff. Yeah. Um you came [6:30] out of retirement actually, right? Um to [6:32] sort of like mess with AI. Yeah. [6:34] >> And then increasingly you were so hooked [6:36] that you wanted to just do it always [6:37] also on the go with the phone. [6:39] I mean the last project I I worked two [6:42] months on Wipe Tunnel [6:44] to the point where it got so good that I [6:46] was catching myself always like coding [6:49] next to my when I was at my friends and [6:51] I was like I need to stop this. This is [6:52] like too addictive. [6:54] And then in November and like my need [6:55] came back and I [6:57] I started building Cloud Bot or now it's [6:59] called Open Claw. [7:01] And I think very very in the beginning I [7:03] was like, "Oh, I rebuilt it again. But [7:06] this time I built it even better." Mhm. [7:08] This time when you don't type into your [7:09] terminal, you just you talk to a friend. [7:12] You don't think about compaction, which [7:14] sessions, which folder I'm in, which [7:17] model I'm in. I mean, you can, you know, [7:19] it's just like I want to leave it open [7:21] for power users, but usually you just [7:23] like you just talk to a friend and the [7:25] friend is like this [7:28] ghost or entity or whatever you want to [7:29] call it that can control your mouse and [7:32] your keyboard and can just do stuff. [7:34] Yeah. And when did you have that aha [7:35] moment when you were like, "Wow, this is [7:37] doing way more things than I actually [7:40] thought it could." [7:41] >> Literally I [7:42] took me 1 hour [7:44] for like the the very shitty initial [7:46] prototype. [7:47] It was just a little bit of glue between [7:50] like a dependency that connects WhatsApp [7:52] and Cloud Code and then I would like [7:55] call Cloud Code and get like the string [7:56] out of Cloud Code. It would be slow but [7:59] it it worked. [8:00] But I wanted images cuz you know you [8:02] want pictures. I want I want I want the [8:04] model to send a selfie or whatever and I [8:06] want the model to create images and send [8:07] me back. [8:09] So that took me another few hours. [8:11] And then I I went to Marrakech [8:14] for a birthday party and [8:16] there was like the internet wasn't that [8:17] good, you know, WhatsApp works [8:18] everywhere because I don't know it's [8:20] just like text. [8:21] So I used it a lot I was in a [8:22] restaurant. What does this mean? You [8:24] make like a picture and like translate [8:25] this for me and just [8:26] it was just so useful. And it was also [8:29] really nice about it because it it spoke [8:30] my language. You know, it [8:32] it was a little sassy. It was like [8:34] funny. It was like really pleasant to [8:36] use. And then I was walking and just [8:38] like sending it a voice message. [8:40] And I'm like, oh wait, this can't work. [8:42] I didn't build that. Right. Right. [8:44] And you see like the type indicator is [8:47] like blinking blinking blinking. 10 [8:48] seconds later it just replied to me. [8:50] I'm like, [8:51] how in the F did you do that? [8:53] And it replied, yeah, the model I did [8:55] the following. You sent me a text [8:57] message and there was no file ending so [8:59] I looked at the header I found it's [9:00] Opus. So I used FFmpeg to convert it to [9:03] wave. [9:04] And then I wanted to like transcribe it [9:05] but I didn't have Whisper installed but [9:07] then I looked around and I found this [9:08] OpenAI key and I just used curl to send [9:10] it to OpenAI. [laughter] [9:11] Got the text back and here I am. [9:14] And [9:15] that all in like what? 9 seconds? And [9:18] you didn't build or anticipate like any [9:20] of those specific things? [9:21] >> No, it you know, turns out um because [9:24] coding models got so good, coding is [9:27] really like creative problem solving [9:30] that maps very well back into the real [9:32] world. I think I think there's there's a [9:33] there's a huge correlation. [9:36] They need to be really good at creative [9:38] problem solving. And that's a skill. [9:40] That's an abstract skill you can apply [9:42] to code but like to any real world task. [9:44] So the the model had a [9:46] I was surprised. It's like a magical [9:48] file. I don't know what it is. I need to [9:49] solve this. And it did its best and [9:51] solved it. And it was even that clever [9:52] that it [9:53] it chose not to install the local [9:56] Whisper because it knows that that would [9:58] require downloading a model which would [10:01] take probably a few minutes. And I'm [10:02] like impatient, you know? So [10:05] it really took the most uh intelligent [10:08] approach. [10:10] And that was kind of like the moment [10:11] where I'm like, holy [10:13] Yeah. [10:15] Uh that was where I got hooked. Y [10:18] Combinator's next batch is now taking [10:20] applications. Got a startup in you? [10:22] Apply at ycombinator.com/apply. [10:25] It's never too early and filling out the [10:27] app will level up your idea. Okay, [10:30] [music] [10:30] back to the video. And so when computers [10:33] can just do all these things that you [10:34] didn't even anticipate. You didn't build [10:35] an app to do that exact thing. [10:38] Are apps just going to go away? [10:41] Uh I think 80% of them are going away. [10:44] Why do I need MyFitnessPal? [10:47] Like my agent already knows that I'm [10:48] making bad decisions. I'm at I don't [10:50] know. [10:52] Uh [10:52] Smashburger something. [10:54] And it will already assume that I eat [10:57] what I like to eat. If I don't make a [10:59] comment, it will just like automatically [11:01] track it or I make a picture and it will [11:03] just store it somewhere. I don't even [11:04] need to care where I right? [11:05] And then my it maybe it it improves my [11:09] my gym schedule. Like add a little bit [11:11] more cardio in it. I don't need my my [11:12] fitness app because it just it just does [11:15] the fitness planning for me. [11:17] Uh [11:17] why do I need it to do that? I just tell [11:18] it, "Hey, remind me of this and this." [11:20] And the next day it will just remind me [11:21] of this and this. Do I care where it's [11:22] stored? No, it just does its thing. [11:25] So there's a [11:26] every app that basically just manages [11:30] data [11:31] could be managed in a better way and [11:33] it's in a more in a more natural way by [11:35] agents. Yeah. Only the apps that [11:37] actually have sensors, maybe they [11:39] survive. And so if, you know, most apps [11:42] are going to go away in that scenario, [11:45] um are the models the only remaining [11:47] sort of apps? Not everything will go [11:49] away. [11:51] But yeah, I think that [clears throat] [11:52] the the large model companies have some [11:54] big mode. [11:55] >> [sighs] [11:55] >> Cuz they ultimately [11:57] they give the token. [11:59] And turns out [12:00] uh one of the complaints was that people [12:03] use too much token. No, you just really [12:05] love using it. That's why you you use [12:07] the thing so much because that's why you [12:09] burn the token. Yeah. Um [12:11] it's like, is it my fault that I make [12:12] something that's so popular? [12:14] And so, you know, like all the the [12:16] models, they're kind of like [12:17] leapfrogging each other constantly and [12:20] and, you know, maybe they're also [12:21] getting commoditized. So, if apps are [12:23] going to go away, models are going to [12:25] get commoditized or at least, uh you [12:27] know, the lobster can, like, the brain [12:29] is is is swappable out. What's the thing [12:32] that remains? What's Where's the value? [12:33] Is it the store of memory? Is it um [12:37] the hardness that's valuable? What is [12:40] What remains? [12:42] First of all, I don't think the [12:44] the model companies always have a mode. [12:47] And I because you see this already, a [12:49] new model comes out, people are like, [12:51] "Oh my god, this is so good." And then [12:53] like a month later, uh it degraded. It's [12:55] not good anymore. They like quantized [12:57] it. No, they didn't do anything. You [12:58] just adapted to the new standard and now [13:01] your expectations went up. [13:03] But the model is still the average. So, [13:05] I think for quite a while, [13:08] uh every time a new model releases, I [13:10] hear the same. People love it and then [13:13] it's the standard. And then what's down [13:15] there, you don't even want to think [13:16] about it anymore. [13:17] So, so we have like open source stuff [13:19] that's as good as the current models [13:20] from a year ago. Everybody's hating it, [13:22] complaining, "Oh, this is not good. It's [13:24] not funny." Yet this was what we had. [13:26] And like in a year, we'll have this open [13:28] source. And then we will like we'll [13:29] complain about this because we are used [13:31] to this. [13:32] So, [13:33] for the the foreseeable future, [13:36] the big companies still have moat. [13:37] Harness-wise, it's going to be [13:38] interesting because [13:41] every company kind of has their own [13:44] their own silo, right? You you There's [13:45] no way Maybe there is for Europeans [13:48] to actually get the memories out of [13:50] ChatGPT. Mhm. I'm I don't I'm not aware. [13:53] I I either there's no there's Definitely [13:55] there's no way for a different company [13:57] to get your memories out. So, if if [13:59] if I was like a company who like [14:01] provides chat services, [14:04] you could use me, but then I couldn't [14:05] access the memories. So, like the [14:07] companies try to like [14:09] bound you to their data silo. [14:11] And the beauty of Open Claw is it kind [14:12] of claws into the data's because at the [14:15] end user, [14:16] the end user needs access cuz it's in [14:18] the end otherwise it wouldn't work, [14:19] right? If the end user access, [14:21] I can access the data. [14:23] And you own the memories. It's just a [14:25] bunch of markdown files on on your [14:26] machine. I mean, I don't own the [14:28] memories. Other people Everybody [14:30] >> Yeah, everyone owns their own memories [14:31] as a bunch of markdown files on their [14:33] own machines. And to be honest, those [14:35] are probably super sensible because [14:39] let's be honest um [14:41] people use their agent not just for [14:43] problem-solving, but also for like [14:45] personal problems. [14:46] >> Very quickly. Super quickly. [14:48] >> I I I I fully do that. I'm like there's [14:51] memory stuff that [14:52] I don't want to have leaked. Yeah, what [14:54] would you rather um uh sort of like not [14:56] show your Google search history at this [14:58] point or your you know, memory.md files? [15:00] What's What's the Google word? [15:02] >> Yeah. Yeah. Are people still using [15:04] Google? [15:05] I built this and I was so excited, [15:08] but on Twitter, people wouldn't get it. [15:12] Yeah. Like I was I was failing to [15:14] explain [15:16] the awesomeness. I feel like [15:18] it needs to be experienced. So, [15:22] I I tried various things and I I [15:24] couldn't I couldn't nail the [15:27] I couldn't nail the explaining. So, I [15:28] was like, "Let's do something really [15:29] crazy." [15:31] I just created a Discord. [15:32] And I just put my bot without any [15:35] security restrictions in the public [15:37] Discord Discord. [15:39] And then people came in and interacted [15:40] with it and they saw me build the [15:42] software with it and they tried to [15:44] prompt inject it and hack it and my [15:45] agent would be laughing at them. And you [15:47] just had it locked down to your user ID [15:49] so it would only listen to you? Yeah. [15:51] Yeah. That and it was I mean very clean [15:53] instructions that other people [15:55] dangerous, only only listen to me but [15:58] respond to everyone. And this prompt was [16:00] in Where was it stored? The [16:02] instructions? [16:03] >> Um [16:04] That's actually part of Open Claude [16:05] itself. Very much so that that that's [16:07] part of the system prompt. Okay, you are [16:09] now that explains to you you in Discord, [16:11] there's like public people there, but [16:14] you only listen to your owner. [16:16] Or like your human. I don't even know [16:17] how I wrote it. Yeah. Yeah. Your god. [16:20] And [16:21] I kept [16:23] I don't know what I did but [16:27] my system was built very organically. [16:29] Like at some point I created like an [16:31] identity.md, a soul.md, like like [16:33] various files and then [16:35] only in in January I started making it [16:37] so other people could install it easier. [16:40] And I remember [16:43] I built all these templates based on [16:45] like, "Go check a rough look at what I [16:47] have and make like templates." And Codex [16:48] wrote it. [16:49] And what came out was like [16:51] bread. You know, like people joke that [16:54] Codex feels like bread even though now [16:55] they have a new friendlier voice. I [16:56] haven't tried that yet. Yeah. But the [16:58] new bots, they felt so boring compared [17:01] to what I had. So, I was like, "Marty, [17:04] infuse the templates [17:05] >> Marty's the name of your personal [17:07] >> Yeah, that's a new It's a new name [17:08] because Yeah. Yeah. [17:10] >> Uh there was some naming challenges. [17:12] Yeah. So, so you were you were talking [17:14] to Marty? [17:15] >> Yeah. I was like, "Infuse Infuse those [17:17] templates with your your character. And [17:19] he changed the templates. And then and [17:21] then like [17:22] all the things that came out afterwards [17:24] were like actually funny. [17:26] Not as funny as mine, so like I kept [17:28] some secret. And the one file that's not [17:31] open source is like my soul.md. So, [17:34] even though my my bot is in public [17:36] discourse, so far nobody cracked that [17:38] one file. [17:39] >> Tell me more about soul.md. I just saw [17:41] this research from Anthropic about where [17:43] they [17:45] Now, I think it's public, but like a few [17:47] months ago it was like where somebody [17:48] ex- [17:49] randomly found out some text that's [17:53] hidden in the [snorts] weights where the [17:54] model couldn't really remember that it [17:56] learned it, but it was like ingrained in [17:58] the weights about a non-equality [18:00] constitution. [18:02] And I found that incredibly fascinating. [18:04] So, and I I talked about it with my [18:05] agent and then we created a soul.md with [18:07] like [18:08] the core values, like how do we want [18:11] human AI interaction, what's important [18:13] to me, what's important to the model. [18:15] Like [18:17] some parts is a little bit like mumbo [18:18] jumbo and some parts is like I think [18:20] actually really valuable in terms of how [18:23] the model reacts and responds to text [18:26] and makes it feel very natural. [18:27] >> In terms of building open claw, um [18:30] you're also kind of taking a little bit [18:31] of a contrarian view at sometimes. Like [18:32] which model you like for coding, which [18:34] one you like to run your bot on, um and [18:36] then also like how you actually like, [18:38] you know, code. Um work trees get work [18:40] trees have kind of been a popular thing. [18:42] There's more and more tools embracing [18:44] them, but you're just you're just like, [18:46] you know, no work trees, just multiple [18:48] checkouts of the repo and like parallel, [18:51] you know, terminal windows. Tell me more [18:52] about how you you build. Yeah, I feel [18:54] like the whole world does cloud code. [18:56] And [18:57] I don't think I could have built this [18:59] thing with cloud code. Like I I love [19:01] Codex because it [19:03] it looks through more files before [19:05] before it decides what to what to [19:07] change. You don't need to do so much [19:10] raid to get a good output. If you're [19:12] skilled, a skilled driver as some people [19:14] say, [19:15] uh you can get reasonably good output [19:17] with any tool. [19:19] But Codex is just [19:21] it's just really brilliant. It is [19:23] incredibly slow. So, sometimes I use [19:25] like 10 at the same side at the same [19:27] time. [19:28] Uh like maybe six on that screen and two [19:31] there and two there. And I don't like [19:33] this is already a lot of complexity in [19:35] my head. There's a lot of jumping. [19:38] So, I try to minimize anything else that [19:40] is complexity. So, in my head main is [19:43] always shippable. I just have multiple [19:45] copies of the same repository [19:48] that are all are on main. So, I don't [19:50] have to deal with how do I name that [19:51] branch? Um [19:54] there could be like conflicts on naming. [19:56] I cannot go back it's [19:58] there are certain restrictions when you [19:59] use work trees that [20:01] I don't need to care about if it's [20:02] copies. [20:03] I don't like [20:05] to use a UI because [20:07] that's again just added complexity. [20:09] Yeah. Like [20:10] they're simpler and less friction. I [20:12] have all I care about is like sinking [20:15] and text. Yeah. I don't necessarily need [20:17] to see so much code. I I mostly see it [20:19] like flying by. Sometimes there's like [20:22] gnarly stuff that I want to like take a [20:23] look. But in most cases [20:27] if you clearly understand the design and [20:29] think it through and discuss it with [20:30] your with your agent, [20:32] it's fine. I'm also very happy that [20:35] I didn't even build an MCP support. So, [20:37] Openclaw is very successful and there's [20:39] no MCP support in there. With a small [20:41] asterisk, I build a skill [20:44] that uses Mac Porter, which is one of my [20:46] tools that converts MCPs into CLIs. [20:49] And then you can just use any MCP as [20:51] CLI. [20:52] Um [20:53] but I totally skipped the whole [20:55] classical [20:57] MCP crap. So, you because you don't then [21:00] you can actually if you need to, you can [21:02] use MCPs on the fly. You don't have to [21:04] restart unlike unlike Codex or cloud [21:06] code where you actually have to restart [21:08] the whole thing, [21:09] I think it's way more elegant and it [21:11] also scales way better. Now you see [21:12] Entropic, they do [21:15] they built like a tool called search [21:17] feature like something super custom for [21:19] MCPs [21:21] that was like in beta because it's like [21:23] so gnarly. [21:24] Now just have CLIs but really is good at [21:27] Unix. You can have as many as you want [21:30] and just works. So like I'm very happy [21:32] that [21:33] I don't I got very little complaints [21:35] about the MCP stuff. It's kind of back [21:37] to you're just [21:38] you giving it the same tools that humans [21:41] liked to use. Yeah. [21:44] Yeah. And not invented stuff for for [21:46] bots per se. Yeah. Humans [21:48] No one's saying human tries to call MCP [21:50] manually. Yeah, they just want to use [21:51] CLIs. Yeah. That's the future. I'm here [21:54] for it. Thank you so much for making the [21:56] time to sitting down chatting. It's been [21:58] a huge inspiration to So like when we [21:59] were texting you know the course of the [22:01] past couple of years and I saw you [22:03] getting back into the game and I was [22:04] like [22:05] Peter, like what you're telling me like [22:07] chase that dragon. [22:09] >> [laughter] [22:09] >> And you were doing like the weird like [22:10] vibe tunnel thing etc. Nobody was paying [22:12] attention and so I'm just like beyond, [22:14] you know, stoked to see, you know, [22:16] what's happening and and of course it [22:18] had to be sort of like a loner from [22:20] someone like tiny country like far away [22:22] from Silicon Valley just like, you know, [22:23] bring all of this upon us. So huge [22:25] inspiration. I'm here for it. Thank you. [22:27] Awesome. Thanks, Peter. [22:31] >> [music]