Manus’ Final Interview Before Acquisition: Surreal 2025

Duration: 211 min · ▶ Watch on YouTube

Guest: 季超 (Peak) · Manus AI 联合创始人兼首席科学家

Switch → 中文

Chapters (46)

  • 00:00 · 嘉宾介绍与早年经历
    • 季超介绍自己的家庭背景，以及高中时期开发iOS浏览器赚取第一桶金的经历。
  • 06:00 · 获得真格基金投资与辍学创业
    • 季超放弃大学Offer，接受徐小平的真格基金投资，正式开启创业之路。
  • 10:50 · 接触NLP与Word2Vec的启发
    • 为了优化浏览器预加载，季超进入NLP领域，并深受Google Word2Vec论文的启发。
  • 12:40 · 第二次创业：语义搜索与知识图谱
    • 受Apple Watch等可穿戴设备启发，季超开始研发基于知识图谱的语义搜索引擎Magi。
  • 16:50 · 研发开放式信息提取技术
    • 为了突破人工标注知识图谱的瓶颈，团队从零开始研发无Schema的开放式信息提取模型。
  • 20:40 · GPT-3的降维打击与公司出售
    • 2019年体验到GPT-3后，季超意识到技术路线的代差，最终选择将公司出售。
  • 26:20 · B2B AI公司的打榜经历
    • 在一家B2B AI公司工作一年半，主要负责内部算法打榜，体验了纯技术驱动的工作模式。
  • 30:30 · 寻找新机会与拒绝大模型公司
    • 离职后与多家头部大模型公司交流，但因不想做纯底层模型或不认同某些理念而未加入。
  • 33:40 · 结识肖弘与加入Manus
    • 被肖弘务实、非“艺术家”的特质打动，决定加入Manus团队，重构浏览器、搜索和模型。
  • 41:30 · AI时代的模型与应用之争
    • 探讨大模型公司与应用公司的边界模糊化，以及Scaling Law（Bitter Lesson）的深刻影响。
  • 50:00 · Team Structure and Decision Making
    • The guest introduces the six co-founders and discusses their evolving decision-making frameworks, including the GPA model.
  • 56:48 · The Failed AI Browser Experiment
    • The team spent months building an AI-native browser but eventually abandoned it after realizing the distribution challenges and observing Arc’s pivot.
  • 1:01:00 · Pivoting and the Value of Monica
    • The team reflects on the sunk cost of the browser project and how Monica provided a rational, cash-flow-positive foundation for the company.
  • 1:06:00 · The Birth of Manus
    • Inspired by coding tools like Cursor, the team realized programming is a general-purpose medium and decided to build a general AI agent for prosumers.
  • 1:11:00 · Delaying the Launch of Manus
    • Despite finishing the product early, the team delayed the launch to align with the release of Claude 3.5 Sonnet V2 for better agentic capabilities.
  • 1:21:00 · Technical Architecture of Manus
    • The guest explains the sandbox environment, the shift from Linux containers to full virtual machines, and the importance of Windows support.
  • 1:30:00 · Model Training and Context Length
    • A deep dive into why current chatbot-aligned models struggle as agents, the illusion of infinite context length, and the need for compaction awareness.
  • 100:00 · Model Capabilities and Differences
    • Discussion on the strengths of different AI models like Anthropic, Google, and OpenAI in coding, multimodal, and reasoning.
  • 100:16 · Product Strategy: General vs. Vertical
    • Why Manus chose to build a general-purpose agent rather than a vertical, specialized tool.
  • 100:41 · Target Audience and Evaluation
    • Defining the ‘Prosumer’ target audience and the shift towards evaluating agents based on real-world labor replacement (RLI).
  • 101:12 · Agent Interaction and Error Handling
    • How users interact with agents, teach them, and how agents handle errors and self-correction.
  • 101:30 · Agent OS and Market Landscape
    • Debating the concept of an ‘Agent OS’ and analyzing the ToB vs. ToC market for AI agents.
  • 102:00 · Competition with OpenAI
    • Comparing Manus to ChatGPT and explaining why general chatbot users differ from agent users.
  • 102:30 · Launch Challenges and Invite Codes
    • The story behind the unexpected surge in traffic during launch and the necessity of invite codes.
  • 103:13 · Pricing and Future Outlook
    • The logic behind Manus’s pricing strategy and the future evolution of agent capabilities.
  • 150:00 · Manus 1.5 Updates & Performance
    • Discussion on the technical evolution of Manus 1.5, highlighting a 3-5x speed improvement for simple tasks.
  • 150:30 · Chatbots vs. General Agents
    • Exploring the relationship between chatbots and agents, and why agents focus on high-value cognitive labor.
  • 152:00 · Business Metrics: Revenue over DAU
    • Why Manus prioritizes revenue and high-value tasks over traditional internet metrics like Daily Active Users (DAU).
  • 153:00 · AI as a Technology Increment
    • The guest argues that AI is a technology enhancement rather than a platform shift, benefiting existing strong players.
  • 156:00 · Network Effects in AI
    • Analyzing why AI products currently lack traditional network effects and how atomic capabilities might create them.
  • 159:30 · Adapting Agents to Human Systems
    • The challenge of making agents work in a digital world designed for humans, using MCP, APIs, and browser automation.
  • 162:00 · Agentic Workflow vs. Pure Agent
    • Debating rule-based workflows versus pure AI-driven agents, aligning with ‘The Bitter Lesson’.
  • 166:20 · Organizational Structure for AI Startups
    • How Manus structures its teams: Sandbox Team, Agent Team, and Research Team.
  • 171:00 · Product Strategy & Feature Restraint
    • The importance of avoiding feature bloat and focusing on synergy, learning from past products like Monica.
  • 175:00 · Online Learning & Mass Personalization
    • Discussing different types of online learning and how mass personalization can be achieved without parameter updates.
  • 178:00 · Advice for Foundation Model Builders
    • Recommendations for model builders: expand context windows and focus on Tool-Integrated Reasoning (TIR).
  • 184:00 · The LLM Competitive Landscape
    • Evaluating the strategies and strengths of OpenAI, Anthropic, Gemini, Google, and xAI.
  • 200:00 · RLVR与Latent Reasoning的探讨
    • 探讨了RLVR在Token Space采样的问题，以及Latent Reasoning在提高推理效率上的优势。
  • 201:35 · 大厂AI战略与人才流动
    • 分析了Meta在AI浪潮中落后的原因以及OpenAI的人才流失与创新潜力。
  • 202:24 · 评价Thinking Machine Labs与开源生态
    • 讨论了Thinking Machine Labs的Tinker API及其对中小规模研究团队的价值。
  • 203:36 · AI领军人物评价
    • 对Mira Murati、Demis Hassabis和Ilya Sutskever等AI行业核心人物的评价。
  • 204:46 · 对国内AI创业者观点的看法
    • 分享了对杨植麟和杨顺渊关于大厂与小公司竞争及问题定义观点的认同。
  • 206:13 · 多模态输入与Agent应用
    • 强调了在Agent场景下，多模态输入（如交错图文输入）比多模态输出更重要。
  • 207:02 · AI泡沫与未来预期
    • 认为AI泡沫是客观存在的但可接受，并分享了对Manus公司最悲观与最乐观的预期。
  • 208:07 · 创业生活与全球化战略
    • 讲述了团队在WeWork的办公日常，以及选择新加坡作为总部以推进全球合规化的原因。
  • 210:07 · 快问快答
    • 分享了个人喜好的食物、地点、冷知识以及影响最深的AI论文。

Specific Numbers (29)

Time Fact Value Context
01:35 App Store出现 2009年 苹果推出App Store的第二年，季超开始开发iOS浏览器。
10:50 开始研究NLP 2011年 为了解决浏览器预加载问题，季超开始接触自然语言处理技术。
11:00 Word2Vec论文发布 2013年 Tomas Mikolov在Google推出了Word2Vec论文，给季超带来了巨大的技术启发。
19:10 从零训练模型 2014年底至2018年 团队花费数年时间从零开始训练用于开放式信息提取的模型。
20:40 接触GPT-3 2019年 拿到了GPT-3的Early Access，意识到自己训练的模型与GPT-3存在巨大差距。
22:35 加入Manus 2024年3月 季超正式加入Manus团队担任联合创始人兼首席科学家。
50:11 Top Chrome extensions user base 50 million Adblock and Grammarly have around 50 million users, which is a ceiling for extensions.
50:19 Chrome DAU 2 billion Chrome’s daily active users, highlighting the massive gap between browser users and extension users.
51:21 Guest’s joining date April 2024 The guest joined the company in April 2024, shortly before they started the browser project.
1:11:13 Monica ARR $12 million Monica’s Annual Recurring Revenue before Manus was launched.
1:14:31 Manus ARR $100 million+ Manus has surpassed $100 million in Annual Recurring Revenue.
1:14:58 Manus subscription price $40/month The default subscription tier for Manus.
1:23:31 Token input/output ratio 100:1 to 1000:1 The ratio of input to output tokens for Manus, compared to 3:1 for standard chatbots.
100:33 Claude Opus 4.5 4.5 Mentioned as an underestimated model for Agentic Coding.
100:63 Current RLI completion rate 2.5% The current success rate of Manus on the Remote Labor Index benchmark.
100:71 Projected RLI completion rate by 2026 20-30% Prediction for how much the agent’s completion rate will improve by 2026.
101:01 Evaluation team size 10+ The number of people dedicated to subjective evaluation of the agent.
101:48 October 2024 2024/10 A time when the team felt no anxiety about competition.
101:59 ChatGPT monthly price $20 Discussing the anchor price set by ChatGPT for AI subscriptions.
150:21 Manus 1.5 speed improvement 3 to 5 times faster Manus 1.5 completes simple tasks significantly faster while allocating more compute to complex tasks.
151:31 A/B testing sample size 5% Manus tested a different model on 5% of users and noticed an immediate drop in satisfaction.
154:21 Pricing sensitivity $40/month vs $200/month Users are willing to pay significantly more if the agent’s effectiveness improves proportionally.
156:11 Search engine market share 90% Discussing how search engines naturally form monopolies compared to chatbots.
164:47 Historical comparison 2002 Comparing the current state of AI chatbots to the early days of Google search.
167:18 Agent Team size 10 to 20 people The number of people working on the core Agent architecture, evaluation, and research at Manus.
171:17 AI era comparison 2018 Comparing the current AI landscape to the post-Transformer era in 2018.
200:53 Pass@1 and Pass@64 1 and 64 Used to compare the energy and cost consumption of reasoning models.
208:10 Daily arrival time 10:30 AM The time the guest and his team typically arrive at the office.
208:25 AC turn-off time 10:00 PM The time the air conditioning turns off at their WeWork office in the mall.

Research Claims & Predictions (17)

  • [15:30] 传统知识图谱无法Scale
    • evidence: 传统知识图谱极度依赖人类专家的手工标注和预设Schema，无法适应无限的实体关系。
  • [41:30] 大模型公司和应用公司的界限会消失
    • evidence: 大模型公司最终会做应用，而成功的应用公司也必须具备做模型的能力，两者会趋于融合。
  • [48:40] Bitter Lesson（苦涩的教训）是AI发展的铁律
    • evidence: AI的历史证明，基于通用方法和大规模算力的投入，最终总是会打败基于人类专家知识的手工规则。
  • [1:08:38] GUI agents for simple tasks lack value.
    • evidence: Human decision-making and clicking is faster than waiting for an AI to process simple UI tasks; AI is only valuable for long-horizon, complex tasks.
  • [1:16:45] Context lengths over 200k are not practically important.
    • evidence: Instead of infinite context, models need ‘compaction awareness’ to understand compressed historical context without losing reasoning ability.
  • [1:34:15] Chatbot alignment hurts agent performance.
    • evidence: Models trained for chatbots suffer from ‘context pressure’ and impatience, causing them to output EOS (End of Sequence) too early instead of completing long-chain reasoning.
  • [100:71] AI agents will be able to complete 20-30% of remote labor tasks by 2026.
    • evidence: Based on the trajectory of model improvements and the new RLI benchmark.
  • [101:62] The vertical agent market will see a ‘hundred flowers blooming’ scenario, especially in ToB.
    • evidence: Driven by the specific needs of enterprises and the maturity of exit mechanisms in the US market.
  • [102:22] Agent token consumption will grow exponentially compared to chatbots.
    • evidence: Because agents require continuous context appending and self-reflection loops without human intervention.
  • [153:05] AI is a technology enhancement, not a platform change.
    • evidence: Current paradigm; traditional strong players with existing distribution will benefit the most.
  • [162:40] Pure AI-driven agents will have a higher ceiling than rule-based Agentic Workflows.
    • evidence: Long-term AI development; aligns with ‘The Bitter Lesson’ that general computation beats human-curated rules.
  • [175:40] Mass personalization can be achieved via in-context learning without parameter updates.
    • evidence: Current/Near future; using prompt injection and context engineering is more efficient than fine-tuning for individual users.
  • [180:00] Tool-Integrated Reasoning (TIR) is more effective than pure reasoning for agents.
    • evidence: Current model training; integrating tool use directly into the reasoning process yields better results than separate steps.
  • [184:00] Scaling law is not dead, but requires new dimensions beyond just token volume.
    • evidence: Future model training; quality of data and new modalities will drive the next leap in performance.
  • [200:35] Latent Reasoning is more efficient than token-space sampling.
    • evidence: It allows for considering multiple possibilities simultaneously in a near-parallel dimension without the need for token-by-token sampling.
  • [207:25] The AI bubble is real but acceptable.
    • evidence: Historically, humans have done crazier things, and the current bubble won’t render the technology useless.
  • [210:58] The next advancement in AI requires user participation.
    • evidence: Stated as a key belief for the future progress of AI technology.

Key Concepts (24)

  • [10:50] NLP (Natural Language Processing)
    • 自然语言处理，季超早期为了解决浏览器预加载问题而进入的技术领域。
  • [11:00] Word2Vec
    • 一种将自然语言文本离散化并转化为稠密向量的技术，是深度学习在NLP应用的重要里程碑。
  • [14:20] Knowledge Graph (知识图谱)
    • 一种结构化知识的表达形式，通过实体和关系（边）来组织信息。
  • [16:50] Open Information Extraction (开放式信息提取)
    • 一种无需提前定义规则或Schema，由AI自动从文本中识别实体和关系的技术。
  • [48:40] Bitter Lesson
    • AI研究中的一个著名观点，认为利用通用计算方法和大规模算力，最终总是胜过人类手工设计的特定领域知识。
  • [50:24] BDFL (Benevolent Dictator For Life)
    • A decision-making model where one person has the final, authoritative say, used by the team for setting goals.
  • [50:35] GPA Framework
    • A decision-making structure dividing choices into Goals (dictatorial), Priorities (mixed), and Alternatives (democratic).
  • [1:08:38] Long-Horizon Tasks
    • Complex tasks requiring multiple steps and long execution times, which are the ideal use case for AI agents.
  • [1:10:00] Prosumer
    • Professional consumers (knowledge workers who aren’t necessarily programmers) who are the target audience for Manus.
  • [1:16:45] Compaction Awareness
    • The ability of an AI model to understand and reason over compressed or summarized historical context, rather than needing infinite raw context.
  • [1:34:15] Context Pressure
    • A phenomenon where models trained for short chatbot interactions feel ‘pressure’ to end their output prematurely when context gets too long.
  • [100:25] Agentic Coding
  • [100:42] Prosumer
  • [100:60] RLI (Remote Labor Index)
  • [101:31] Agent OS
  • [162:20] Agentic Workflow vs. Pure Agent
    • Agentic Workflow relies on human-defined rules and constraints, whereas a Pure Agent uses the model’s intelligence to determine the best path to complete a task.
  • [160:03] MCP (Model Context Protocol)
    • A protocol designed to help models interact with external tools and APIs, though currently facing maturity and ecosystem challenges.
  • [180:28] Tool-Integrated Reasoning (TIR)
    • A training approach where the model’s reasoning process is deeply integrated with its ability to use external tools, rather than treating them as separate steps.
  • [175:44] Mass Personalization
    • Customizing AI behavior for individual users at scale using context engineering and prompt injection, rather than fine-tuning model weights.
  • [162:40] The Bitter Lesson
    • The philosophy that general methods leveraging massive computation ultimately outperform human-curated, domain-specific knowledge.
  • [182:34] Pixel in Pixel out
    • A hypothetical AI architecture that processes raw screen pixels directly instead of relying on text or code interfaces.
  • [200:08] RLVR (Reinforcement Learning with Verifiable Rewards)
    • A method of training models where reasoning is done via token-space sampling, which the guest argues is closer to search than true reasoning.
  • [200:37] Latent Reasoning
    • Reasoning in a continuous or parallel latent space rather than discrete token space, allowing for higher efficiency and evaluating multiple possibilities simultaneously.
  • [206:23] Multimodal Input for Agents
    • The ability of an AI agent to process interleaved text and image inputs (e.g., reading a webpage screenshot) to make decisions, which is currently under-prioritized in training.

People Mentioned (26)

  • 小珺 — 本次访谈的主持人。
  • 季超 (Peak) — 访谈嘉宾，Manus联合创始人兼首席科学家。
  • 徐老师 (徐小平) — 真格基金创始人，鼓励季超放弃学业直接创业。
  • Tomas Mikolov — Google研究员，Word2Vec论文的作者。
  • 肖弘 — Manus创始人兼CEO，以其务实和非“艺术家”的特质打动了季超。
  • 张涛 (Zhang Tao) — CPO of the company, a serial entrepreneur who handles product and external partnerships.
  • 小红 (Red) — CEO of the company, who acts as the ultimate decision-maker for product direction.
  • 潘潘 (Panpan) — CTO of the company.
  • 慧杰 (Huijie) — CMO of the company.
  • CZ (Chen Zhe) — COO of the company, handling operations and finance.
  • Josh Miller — Founder of The Browser Company, whose tweet about discontinuing Arc influenced the team’s pivot.
  • Patrick Collison — Mentioned as an early, high-value user of Manus.
  • Jack Dorsey — Mentioned as an early, high-value user of Manus.
  • Andrej Karpathy — Mentioned as someone whose influence or endorsement cannot simply be bought.
  • Thomas Kurian — Google Cloud CEO, mentioned as presenting Manus at a developer conference.
  • Yann LeCun — Former Chief AI Scientist at Meta; his departure from the role is seen as a potential positive signal for Meta’s pragmatic AI focus.
  • Yuandong Tian — Researcher at Meta, mentioned for his interesting work on Latent Reasoning.
  • Xiaohong (Red) — Team member who handles emotional stability and stress within the startup.
  • Zhang Tao (Tao Ge) — Team member acting as CPO, handling external relations, partnerships, and product insights.
  • Panpan — Technical leader within the team.
  • Mira Murati — Former OpenAI executive, praised for her leadership and appeal based on industry reputation.
  • Demis Hassabis — CEO of DeepMind, cited as the person the guest admires most from a purely technical perspective.
  • Ilya Sutskever — Highly respected AI figure; the industry is eagerly waiting to see his next project.
  • Yang Shunyu — AI founder whose views on big companies copying small companies were strongly agreed with by the guest.
  • Yang Zhilin — Founder of Moonshot AI, quoted for his philosophy on defining and solving problems.
  • Lao Huang (Jensen Huang) — Quoted as saying ‘nothing will be surprising’ regarding the future of AI.

Companies Mentioned (25)

Manus AI · 真格基金 (ZhenFund) · Google · OpenAI · DeepSeek · 千问 (Qwen) · Kimi (月之暗面) · Cursor · Grammarly & Adblock · The Browser Company (Arc) · Google DeepMind · Anthropic · Microsoft · Kimi · xAI · Meta · Slack · Notion · GitHub · Stripe · Cloudflare · Thinking Machine Labs · DeepMind · Manus · Temu / Shein

Notable Quotes (17)

我把我这辈子想试的技术都以合理的方式花投资人的钱给搞定了，所以我觉得在那个项目做完的那一刻，我的人生已经圆满了，我早就无憾了。 — 季超 @ 26:00

别的创始人太艺术家了，小红有一个非常稀缺的特质，他很正常，他身心健全，没有任何不良嗜好，没有极端思想。 — 季超 @ 37:20

Everything added dilutes everything else. (每增加一个东西都会稀释其他所有东西的价值) — 季超 @ 42:25

For every complex problem, there is an answer that is clear, simple, and wrong. — Co-founder of Manus/Monica @ 1:03:58

Monica是一张理智的AI应用的船票。 — Co-founder of Manus/Monica @ 1:06:05

编程其实是一个通用能力，它是解决通用任务的一个媒介。 — Co-founder of Manus/Monica @ 1:09:04

什么是ARR？ARR是MRR乘以12。你不能把你一个月内获得的年付算在当月。 — Co-founder of Manus/Monica @ 1:13:52

如果你做的是一款通用的Agent的话，你其实在做一个类似人的东西。 — Manus Founder @ 100:50

模型是比人更加全能的一个东西，所以你应该充分利用模型的优势，而不要生搬硬套人带来的这条约束。 — Manus Founder @ 101:40

ChatGPT是一个打透了的产品…但实际上有Agent需求的人客观来说是少了一个量级的。 — Manus Founder @ 102:02

AI其实不是一个平台变化，AI是一个技术增量。 — Manus Founder @ 153:05

完成一个任务的所有过程和方式是由智能本身决定的… 这才是更符合The Bitter Lesson的事。 — Manus Founder @ 162:40

如果你真的是在做一个Agent，你同时在做两个产品，一个是给人用的，一个是给Agent用的。 — Manus Founder @ 164:15

学会跟AI共处之后其实没有那么多的恐惧，反而是给你解放了更多人的一面。 — Manus Founder @ 181:15

以前都是小公司抄大厂，现在是大厂抄小公司。 — Guest (quoting Yang Shunyu) @ 204:55

Career Arc & Personal Stories (13)

  • [01:35] 高中时期，季超抓住了苹果App Store刚起步的红利，开发了一款iOS浏览器，赚取了30多万美元的第一桶金。
  • [06:00] 面对读大学还是创业的选择，季超接受了真格基金徐小平的投资，放弃了大学Offer，正式开始创业。
  • [12:40] 为了解决可穿戴设备上的搜索交互问题，季超带领团队研发了基于知识图谱的语义搜索引擎Magi，并死磕底层模型训练。
  • [20:40] 在拿到GPT-3的早期访问权限后，季超发现自己辛苦训练的模型被降维打击，最终决定卖掉公司。
  • [26:20] 经历创业挫折后，季超加入了一家B2B AI公司，度过了一段纯粹为了算法打榜赢显卡的快乐时光。
  • [33:40] 在寻找新机会时，季超拒绝了多家大模型公司，最终被Manus CEO肖弘的务实和产品理念打动，决定加入Manus重构AI时代的浏览器和搜索。
  • [51:00] The team spent from April to September building an AI-native browser. Despite having a working product, they realized the distribution model was flawed and the user value wasn’t there. They bravely decided to scrap the project after seeing Arc’s founder admit similar struggles.
  • [1:06:00] Observing non-engineers in their company using Cursor (a coding tool) to write blogs and analyze data sparked the realization that code is just a medium for solving general problems, leading to the creation of Manus.
  • [102:30] The founder recounts the chaotic launch of Manus, where unexpected global traffic overwhelmed their compute resources, forcing them to implement an invite code system and sleep only 3-4 hours a day while fixing servers.
  • [171:00] The founder reflects on his previous product ‘Monica’, noting that building too many disconnected features led to team bloat and a lack of core synergy, a mistake he avoids with Manus.
  • [175:50] He mentions his past experience working on ‘Maggie’, where he learned valuable lessons about continuous and lifelong learning in AI systems.
  • [208:08] The guest shares his startup routine: arriving at the WeWork office in a mall at 10:30 AM and working late into the night. Because the mall’s AC turns off at 10:00 PM, the team stays to see how long they can endure the heat, highlighting their startup hustle.
  • [209:00] The guest explains the strategic decision to set up Manus’s headquarters in Singapore. It was driven by the need to reduce communication costs across distributed teams in China and to meet strict global compliance standards (SOC2, ISO, GDPR) required for a global market.

Tools & Models Discussed (24)

  • Word2Vec: 将词语转换为稠密向量，用于捕捉词语之间的语义关系。
  • LSTM: 长短期记忆网络，早期用于处理序列数据和自然语言的深度学习模型。
  • GPT-3: OpenAI推出的大型语言模型，展现了强大的上下文理解和生成能力。
  • Magi: 季超早期团队开发的基于知识图谱和开放式信息提取的语义搜索引擎。
  • Monica: 肖弘团队之前开发的一款广受欢迎的AI Chrome浏览器插件。
  • Monica: An AI assistant browser extension that reads context passively and generates cash flow for the company.
  • Manus: A general-purpose AI agent that operates in a cloud sandbox to execute complex, long-horizon tasks for users.
  • Claude 3.5 Sonnet V2: The foundational AI model that provided the necessary agentic capabilities for Manus to function effectively.
  • Cursor: An AI-powered IDE that inspired the Manus team by showing how non-programmers use code generation to solve general tasks.
  • Manus: A general-purpose AI agent designed to act like a remote worker, using a browser and tools to complete complex tasks.
  • Claude Opus 4.5: An Anthropic model praised for its agentic coding capabilities.
  • Gemini: Google’s model, noted for its strong multimodal inputs and video understanding.
  • SWE-bench: A benchmark for evaluating AI models on software engineering tasks.
  • Cursor: An AI-powered code editor that the founder views as a strong, general-purpose coding tool rather than a narrow vertical agent.
  • Manus 1.5: An AI agent designed to execute complex, high-value cognitive tasks autonomously.
  • ChatGPT: OpenAI’s conversational AI, referenced as the dominant chatbot but not necessarily the best general agent.
  • Claude: Anthropic’s LLM, praised for its coding capabilities and alignment with high-value tasks.
  • Gemini: Google’s LLM, noted for its strong multimodal capabilities and access to search data.
  • Llama: Meta’s open-source LLM, mentioned in the context of team turnover.
  • ChatGPT: OpenAI’s flagship product; the guest specifically mentioned the ChatGPT Agent team.

Topics

AI技术演进史 (从NLP到大模型) · 技术创业者的心路历程与反思 · 知识图谱与大语言模型的路线之争 · AI应用层与模型层的商业模式 · 优秀创业团队的合伙人特质 · Startup Decision Making · AI Agents vs Chatbots · Product Market Fit · Model Context Length · Cloud Sandboxing for AI · SaaS Revenue Metrics (ARR) · AI Agents · Model Evaluation · Product Strategy · Market Competition · Startup Scaling · Pricing Models · AI Agents vs Chatbots · Product Strategy and Business Metrics · Agentic Workflows vs Pure Agents · Model Context Protocol (MCP) · Organizational Structure for AI Startups · The LLM Competitive Landscape · Future of AI and Human Employment · RLVR vs Latent Reasoning · AI Industry Dynamics (Meta, OpenAI, DeepMind) · AI Agents and Multimodal Inputs · AI Bubble · Global Expansion and Compliance for AI Startups · Startup Culture

Takeaways

  • 技术发展存在’Bitter Lesson’，算力加通用算法最终会颠覆依赖人工规则的系统。
  • 对于技术型创业者，找到一个务实、懂产品且能互补的CEO（非’艺术家’）至关重要。
  • 大模型时代，纯粹的模型公司和应用公司的边界正在消失，优秀的应用需要深入底层，优秀的模型也需要应用场景。
  • 早期的失败和挫折（如被GPT-3降维打击）能帮助创业者放下执念，以更健康的心态迎接下一个技术浪潮。
  • Don’t build products just because they sound cool; ensure they solve a real problem better than existing tools.
  • For AI agents, long-horizon tasks are much more valuable than simple GUI automation.
  • Programming is no longer just for software engineers; it is a universal medium for AI to solve general knowledge work tasks.
  • Current LLMs are over-optimized for chatbot interactions, making them impatient and prone to failure in long-step agentic workflows.
  • True ARR should be calculated strictly as MRR x 12, not by front-loading annual subscription payments.
  • General-purpose AI agents should be designed to mimic human interaction with computers rather than acting as specialized tools.
  • The evaluation of AI is shifting from static benchmarks to real-world economic value, such as the Remote Labor Index.
  • Scaling an AI agent product is fundamentally different from scaling a chatbot due to the exponential increase in token consumption and compute requirements.
  • The market for AI agents will likely bifurcate into general-purpose agents for ‘prosumers’ and highly specialized vertical agents for enterprise (ToB) use cases.
  • AI startups should focus on revenue and high-value tasks rather than chasing Daily Active Users (DAU).
  • Avoid imposing human constraints on AI (Agentic Workflows); trust the model’s intelligence to find the best path (Pure Agents).
  • Building an AI agent requires creating two separate interfaces: one for the human user and one for the model.
  • The scaling law is not dead, but future breakthroughs will rely on new dimensions like multimodal data and tool-integrated reasoning.
  • Product teams must exercise restraint; adding features without synergy dilutes the core value and bloats the organization.
  • Latent reasoning is viewed as a more efficient and promising path for AI reasoning compared to token-space sampling (RLVR).
  • Despite talent drain, OpenAI still possesses the innovative culture needed to create the next AI paradigm.
  • For AI startups aiming at a global market, establishing a headquarters in places like Singapore is crucial for navigating international compliance (SOC2, GDPR).
  • The AI bubble is a natural phase of technological evolution and should not cause extreme pessimism; the technology will remain valuable.
  • The next major breakthrough in AI development will heavily rely on active user participation and feedback.