以下文章来源于刺猬公社 ，作者刺猬公社编辑部 刺猬公社 . 本文来自微信公众号： 刺猬公社 ，作者：刺猬公社编辑部，题图来自：AI生成 Gemini3Pro发布后，市场充斥着欢呼声。最引人瞩目的，是其"Vibe Coding"能力，即用户通过自然 语言描述，就能指挥AI牛马，写出自己的代码。 但最近两年，AI圈的"炸裂"时刻实在太多了，论数量可能仅少于电影圈，狼炸多了总显得面目可 疑。比起Crusor、Trae等市面上已有的Vibe Coding产品，Gemini3Pro到底有何新变化？ 刺猬公社决定派出一位文科生作为代表，实测毫无编码基础的普通用户，能用Gemini3Pro做些什 么。 实测1：生成一个牛马时钟 互联网内容行业观察与研究 牛马时钟是本次实测中需求最简单的一个案例。交互层面，用户只需要选择计划退休时间、上班时间 和下班时间，剩下的都是计算和前端显示。没有经过任何调试，最终效果几乎挑不出毛病来。 唯一值得讨论的问题，是页面右下角电量消耗进度条使用了绿色而不是灰色。但由于电量消耗到 100%，用户就能迎来下班，状态确实比上班更快乐，所以你也不好说这到底是一个Bug还是一个 Feature。 在做这 ...

Maor Shlomo is the Founder and CEO of Base44, the AI building platform that Maor built from idea to $80M acquisition by Wix, in just 8 months. Today the company serves millions of users and will hit $50M ARR by the end of the year. Before Base44, Maor was the Co-Founder and CTO of Explorium. ----------------------------------------------- Timestamps: 00:00 Intro 01:13 The Decision Behind Selling Base44 11:56 How Base44 Thinks About Defensibility 16:37 Mapping the Future of the Market in 3–5 Years 23:18 I’m ...

Gemini 3 Capabilities & Features - Showcases Gemini 3's ability to bring any idea to life in AI Studio through vibe coding [1] - Highlights various functionalities of AI Studio's build tab, including one-shot, iteration, deployment, sharing, and "I'm Feeling Lucky" features [1] - Demonstrates features like Nano Banana, Maps Grounding, and Live API within AI Studio [1] - Showcases Gemini 3's multimodal capabilities, including video understanding [1] - Illustrates Gemini 3's multilingual and reasoning capabilities [1] AI Studio Applications & Demos - Presents a hand-tracking rhythm game demo [1] - Features an interactive choreography app created through vibe coding [1] - Demonstrates Google Maps Grounding with a running app demo [1] - Showcases the use of Gemini 3 for web design and aesthetics [1] - Presents a demo of bringing images to apps [1] - Visualizes quantum research papers [1] - Showcases interactive floor plans and 3D home design [1] - Features a live vibe coding session building a fishing game [1] - Introduces Anti-Gravity, a next-generation agentic IDE [1] - Demonstrates Anti-Gravity features like artifacts and browser actuation [1] - Presents a multiplayer collaborative whiteboard demo [1] - Showcases a 3D globe travel app and chess learning application [1] - Features a Veo 3.1 generative video app demo [1]

Core Insights - Google has officially launched its latest foundational model, Gemini 3, which shows significant improvements in deep reasoning, multimodal understanding, and agent programming capabilities [1] - Gemini 3 Pro achieved a new state-of-the-art (SOTA) score of 71.6 on the xbench-ScienceQA leaderboard, surpassing Grok-4 and demonstrating faster response times and lower costs [1][3] Performance Metrics - Gemini 3 Pro scored an average of 71.6 with a BoN of 85, while Grok-4 scored 65.6, indicating a 6-point lead over the second-place model [5] - The average response time for Gemini 3 Pro is 48.62 seconds, significantly faster than Grok-4's 227.24 seconds and GPT-5.1's 149.91 seconds [6] - Cost analysis shows that running the ScienceQA tasks with Gemini 3 Pro costs only $3, compared to $32 for GPT-5.1, making it substantially more economical [6] Technological Advancements - Gemini 3 introduces a cognitive architecture that shifts from reactive to cautious reasoning, utilizing a "Deep Think" mode that allows for multiple reasoning pathways and self-verification [8] - The model employs a sparse MoE architecture, activating only a small subset of its vast parameters during computation, which enhances efficiency while maintaining performance [8] Developer Tools and Features - The introduction of "Vibe Coding" allows Gemini 3 to align code generation with developer intent, functioning as an autonomous agent capable of executing complex tasks within an IDE [9] - Gemini 3 Pro integrates with Google’s Antigravity platform, enabling developers to automate workflows that involve reading web pages, executing commands, and generating code seamlessly [10] Multimodal Capabilities - Gemini 3 adopts a native multimodal architecture, allowing it to process text, code, images, video, and audio using a unified world model, enhancing its perception and interaction capabilities [11] - The model can generate dynamic, interactive user interfaces in real-time based on user intent, marking a shift from static outputs to interactive experiences [12] Hardware Infrastructure - Gemini 3 is trained on Google’s proprietary TPU (Tensor Processing Unit), designed for high-bandwidth and parallel computing, facilitating efficient training and cost management [13]

Financial Data and Key Metrics Changes - Total bookings grew to $515 million in Q3 2025, up 14% year-over-year, driven by robust new user cohorts and better-than-expected results from Base44 [15][22] - Total revenue increased to $505 million, also up 14% year-over-year, exceeding the high end of guidance [15][22] - Non-GAAP gross margin in Q3 was 69%, slightly down from 70% in Q2 [18] - Free cash flow for Q3 was $159 million, or 32% of revenue, an increase from 30% last quarter [21] Business Line Data and Key Metrics Changes - Partners' revenue grew 24% year-over-year to $192 million, driven by traction among professional designers and solid Studio adoption [16] - Transaction revenue was $65 million, up 20% year-over-year, driven by 13% growth in GPV [16] - Base44's share of audience traffic increased to over 10% in October, up from low single digits in June [10][12] Market Data and Key Metrics Changes - Organic traffic improved as more users searched for Wix online, indicating enhanced brand awareness [9] - Base44 serves over 2 million users globally, translating to more than 1,000 new paying subscribers daily [12] Company Strategy and Development Direction - The company aims to democratize software creation through Vibe Coding, enabling users without coding knowledge to build applications [3][5] - Base44 is positioned as a leader among low-code tools, with significant growth potential in the AI-powered app building space [6][10] - The company plans to enhance its marketing efforts to capture stronger-than-anticipated demand for Base44 [23] Management's Comments on Operating Environment and Future Outlook - Management expressed confidence in the durability of the core business and the ability to enter new markets effectively [14] - The company raised its full-year bookings outlook to $2,060-$2,078 million, reflecting strong performance from Base44 [22][23] - Management anticipates improved operating and free cash flow margins over time as AI costs decrease and marketing investments normalize [21][24] Other Important Information - The company issued $1.15 billion in 0% convertible senior notes due 2030, with plans to use the cash for business purposes and share repurchases [22] - The anticipated launch of a new flagship product has been delayed to early 2026 due to technical challenges [7][70] Q&A Session Summary Question: Dynamics of monthly vs. annual subscriptions for Base44 - Management noted that most Base44 users are currently on monthly subscriptions, with expectations that they will transition to annual plans over time as trust in the platform grows [28][30] Question: Cost management and AI compute expenses - Management acknowledged that while AI costs are currently high due to new user consumption, they expect to reduce these costs over time through various strategies [31][35] Question: Vision for Wix in the AI era - Management highlighted efforts to ensure that all Wix websites are indexable by LLMs and to continuously add new e-commerce standards to enhance visibility [39][40] Question: Cohort retention trends for Base44 - Management indicated that retention trends for Base44 are similar to those seen in early Wix cohorts, with expectations for improvement as brand recognition grows [42][43] Question: Pathway to improving Base44 margins - Management discussed the need for time to see improved margins as the user base matures and costs stabilize, with expectations for better profitability in the future [48][53] Question: Impact of hyper-growth on margins - Management confirmed that while there will be short-term margin pressure due to high growth and new customer acquisition costs, they anticipate margin improvement as the business scales [56][66]

Financial Data and Key Metrics Changes - Total bookings grew to $515 million in Q3 2025, up 14% year-over-year, driven by robust new user cohorts and better-than-expected results from Base44 [15][16] - Total revenue increased to $505 million, also up 14% year-over-year, exceeding the high end of guidance [15][16] - Non-GAAP gross margin in Q3 was 69%, down slightly from 70% in Q2 [18] - Free cash flow was $159 million, or 32% of revenue, an increase from 30% last quarter [20] Business Line Data and Key Metrics Changes - Partners' revenue grew 24% year-over-year to $192 million, driven by traction among professional designers and solid Studio adoption [16] - Transaction revenue was $65 million, up 20% year-over-year, driven by 13% growth in GPV [16] - Base44's share of audience traffic increased from low single digits in June to over 10% in October [11][12] Market Data and Key Metrics Changes - GPV grew 13% year-over-year to $3.7 billion, with partners contributing approximately 55% of total GPV [16] - Base44 serves over 2 million users globally, translating to more than 1,000 new paying subscribers daily [12] Company Strategy and Development Direction - The company aims to democratize software creation through Vibe Coding, enabling users without coding knowledge to build applications [3][5] - Base44 is positioned as a leader among low-code tools, with significant growth potential in the AI-powered app building space [6][11] - The company is focusing on building a comprehensive brand and marketing function for Base44 to capture a larger market share [12] Management's Comments on Operating Environment and Future Outlook - Management expressed confidence in the durability of the core business and the ability to enter new markets effectively [14] - The company raised its full-year bookings outlook to $2,060-$2,078 million, reflecting strong performance from Base44 [22][23] - Expectations for free cash flow are approximately $600 million for 2025, or 30% of revenue [24] Other Important Information - The company issued $1.15 billion in 0% convertible senior notes due 2030, with plans to use the cash for business purposes and potential M&A opportunities [22] - The anticipated launch of a new flagship product has been delayed to early 2026 due to technical challenges [7][70] Q&A Session Summary Question: Dynamics of monthly vs. annual subscriptions for Base44 - Management noted that most Base44 users are on monthly subscriptions, with expectations that as trust builds, more users will transition to annual plans [26][27] Question: Churn rates and subscription dynamics - Churn is currently higher than Wix's standard, but management is optimistic about future retention as the product matures [27][28] Question: Pathway to improving Base44 margins - Management indicated that as Base44 matures, margins are expected to improve, with a focus on reducing AI costs and optimizing the customer mix [49][53] Question: Impact of hyper-growth on margins - Management acknowledged that while there will be short-term margin pressure due to high growth, they expect margins to improve as the business scales [56][66] Question: Delays in the new self-creator tool - The delay is attributed to fine-tuning technology and addressing technical challenges, with a confident outlook for a launch in early 2026 [70] Question: Retention trends for Base44 cohorts - Early retention trends for Base44 are similar to those seen in Wix's early days, with expectations for improvement as brand visibility increases [43][44]

Core Insights - Google has officially released Gemini 3.0 Pro, which emphasizes enhanced reasoning and understanding capabilities, allowing users to receive higher quality responses without needing detailed prompts [1] - In authoritative benchmark tests, Gemini 3.0 Pro achieved a leading score of 72.1% in factual accuracy assessments and 23.4% in mathematical tests, indicating its reliability across various disciplines [1] Benchmark Performance - Gemini 3.0 Pro outperformed its predecessor Gemini 2.5 Pro and competitors like Claude Sonnet 4.5 and GPT-5.1 in multiple benchmarks, including: - 91.9% in scientific knowledge (GPQA Diamond) [2] - 95.0% in mathematics (AIME 2025) [2] - 81.0% in multimodal understanding (MMMU-Pro) [2] - 72.1% in parametric knowledge (SimpleQA Verified) [2] User Experience and Practical Applications - The model's capabilities allow users to generate product interfaces solely through natural language prompts, achieving a level comparable to professional UI designers [5][6] - Gemini 3.0 Pro can create interactive games from static images, demonstrating its versatility and ease of use [7][8] - The model significantly reduces the time and effort required for product development, enabling rapid prototyping and deployment [9][10] Future Implications - Gemini 3.0 Pro represents a shift in software production methods, lowering the marginal cost of trial and error in product development [10] - The model is expected to set a new standard in the industry, potentially transforming the capabilities of individual developers and small teams [10]

Core Insights - The term "vibe coding" has been officially recognized by Collins Dictionary as the word of the year for 2025, symbolizing a shift in programming from logic-based coding to a more intuitive, feeling-based approach [1][10][24] - This new programming style emphasizes collaboration with AI, where programmers use natural language to describe their needs, allowing AI to assist in code generation and logic completion [5][15][22] Group 1: Evolution of Vibe Coding - "Vibe coding" originated from a humorous tweet by Andrej Karpathy, highlighting a new way of programming that prioritizes intuition over strict coding rules [5][6] - The term quickly gained popularity across various tech forums and social media, becoming a cultural symbol of how AI is reshaping the programming landscape [8][17] - The definition of "vibe coding" has evolved to represent the act of using natural language prompts to have AI assist in writing computer code, marking a significant cultural shift in the tech community [10][20] Group 2: Impact on Programming Practices - AI tools like GitHub Copilot and Replit Ghostwriter are enabling programmers to focus on high-level ideas rather than syntax, transforming the coding process into a more collaborative and intuitive experience [15][16] - The programming environment is shifting towards AI-assisted development, where the emphasis is on setting the tone and intent rather than controlling every line of code [22][23] - This transformation is seen as a paradigm shift in software development, with AI taking a more central role in the coding process [21][24] Group 3: Language and Communication Changes - The rise of "vibe coding" reflects a broader change in how humans communicate with machines, moving from precise technical language to more emotional and intuitive expressions [25][27] - Research indicates that prolonged interaction with AI influences human language habits, leading to a more AI-like style of communication [25][28] - The blending of human emotional expression with machine logic suggests a future where creativity and computation coexist, redefining the boundaries between human and machine interactions [24][29]

Core Insights - The article discusses the challenges in reproducing algorithms and experimental results from academic papers in the AI field due to the lack of critical implementation details [1][2] - DeepCode, an open-source tool developed by a team from the University of Hong Kong, addresses these challenges by analyzing paper content and automatically generating runnable code [2][4] Group 1: DeepCode Performance - DeepCode has shown outstanding performance in benchmark tests, surpassing human experts and other advanced AI coding tools [3][4] - In the PaperBench benchmark, DeepCode achieved an overall accuracy of 75.9%, exceeding the human expert group’s score of 72.4% [5][6] - DeepCode scored 84.8% in a comparison with commercial code assistants, significantly outperforming Claude Code, which scored 58.7% [17][19] Group 2: Benchmark Testing - The benchmark tests included comparisons against human experts, state-of-the-art commercial code assistants, scientific code assistants, and large model-based agents [3][4] - DeepCode consistently received the highest scores across all four benchmark tests [4] - The PaperBench benchmark involved reproducing 20 ICML2024 conference papers, with 8316 independently scoreable components evaluated [8][9] Group 3: DeepCode's Core Capabilities - DeepCode's primary capabilities include transforming academic papers into production-level code, generating responsive web pages from natural language descriptions, and creating high-performance backend services from functional requirements [24][25][27] - The tool employs a systematic three-phase framework for code generation, which includes architecture blueprint construction, automated code building, and dynamic validation and optimization [29][35] Group 4: Challenges and Future Directions - Current AI programming tools excel in code completion and simple tasks but struggle with complex tasks requiring deep understanding [36][38] - The development of DeepCode indicates that specialized architecture design can lead to better performance in specific domains, although general deep understanding capabilities remain limited [38][39] - The evolution of AI coding tools from simple assistants to comprehensive development partners raises questions about maintaining developer control and ensuring code quality [40][42]

Core Viewpoint - The article discusses the integration of AI in recruitment processes, particularly focusing on Meituan's use of AI interviewers and the emerging trend of "Vibe Coding" in software development [7][28][38]. Group 1: AI Integration in Recruitment - Meituan has implemented AI interviewers in its recruitment process, with a significant portion of its code (approximately 52%) generated by AI [7][8]. - The AI interview experience has been reported positively by candidates, indicating a shift in how technical interviews are conducted [17][18]. - The AI interview questions cover both technical and general abilities, reflecting a comprehensive assessment approach [15][18]. Group 2: Vibe Coding Concept - "Vibe Coding" is introduced as a new software development methodology where developers use natural language prompts to guide AI tools in generating and debugging code [29][30]. - This approach emphasizes speed and iterative development, allowing even non-programmers to create software by describing desired functionalities [30][31]. - The article argues that the future of programming will heavily involve AI, making it essential for developers to adapt to this collaborative model [34][36]. Group 3: Future of Interviews - The article suggests that future interviews should assess candidates' ability to collaborate with AI tools, rather than just their technical knowledge [38][39]. - Different levels of developers will have varying expectations regarding their proficiency in using AI tools, from basic usage for junior developers to leadership in AI development processes for senior developers [36][37]. - The shift in interview focus from traditional knowledge to understanding and collaboration with AI represents a significant paradigm change in recruitment practices [38][46].