This week, I’ve mainly been exploring MCP servers and trying to understand why they’re suddenly generating so much buzz. Additionally, I found Airbnb’s insights particularly enjoyable as they shared the lessons learned transitioning their search ranking system from Gradient Boosted Trees to deep learning. Their paper is highly recommended, offering practical insights and clearly highlighting real-world pain points.
AI
- Applying Deep Learning to Airbnb Search: Airbnb significantly improved its search ranking using machine learning, initially deploying a gradient boosted decision tree model that reached a plateau in performance. To overcome this, they transitioned to using deep learning and neural networks (NNs) for search ranking. This involved iterative refinements and experimenting with new techniques like Lambdarank for rank optimization and leveraging offline metrics like NDCG. They explored multiple models, features, and neural network configurations, ultimately adopting a deep neural network with hidden layers to improve booking predictions. However, challenges included overfitting, the complexities of model architecture, handling large feature sets, and ensuring smooth feature distributions. Despite initial setbacks, deep learning led to a transformation in their optimization approach and online performance, moving the focus away from purely feature engineering.
- Google Y Gemini 2.5 Pro: Aprendiendo a Vivir Sin Clics: The text discusses Google’s launch of Gemini 2.5 Pro, its most advanced AI model, which is available for free with usage limitations and signals a strategic shift from ad-based revenue to subscription models. This transition is urgent for Google as their previously dominant search-based advertising model shows signs of decline. The piece highlights the challenges Google faces as it tries to monetize new AI technologies while maintaining user trust amidst concerns over AI “hallucinations.” As Google attempts to adapt, the success of its subscription model remains uncertain as users may already be paying for other services like ChatGPT Plus, raising questions about whether Google can successfully reinvent itself or risk becoming obsolete in the evolving digital landscape.
- Vibe Code Tools, Not Toys.: In “Vibe Code Tools, Not Toys,” Ramiro Aznar discusses his journey from building a complex anomaly detection system manually at Planet using Python, SQL, and dbt, to rapid development using modern AI tools at Tinybird. Initially, he faced challenges due to the complexity and time required to implement and manage the systems. However, with the rise of AI-assisted IDEs and prompt engineering, he transformed the process, drastically reducing development time to a single day. Although “vibe coding” with AI presents potential issues and is often seen as suitable for non-production projects, Aznar highlights its effectiveness for internal tools, emphasizing a shift from traditional coding to AI-driven development for efficiency.
- The Cybernetic Teammate: Ethan Mollick’s study, “The Cybernetic Teammate,” explores how AI integrates into teamwork, based on a randomized controlled trial at Procter & Gamble. The study found that AI, when used by individuals, can mimic the performance of traditional teams, and AI-assisted teams slightly outperform AI-assisted individuals, especially for top-tier solutions. AI bridges functional knowledge gaps and creates balanced outputs, reducing silos among professionals. Interestingly, AI also improves emotional experiences at work. This implies that AI functions more like a teammate than a mere tool, prompting organizations to rethink team structures and the nature of work itself.
- IA Y Navegadores: The text discusses the evolution of the internet from Mozilla’s perspective, emphasizing the shift from the web as the primary ecosystem to the significant role of data and AI in digital life. It questions how Mozilla can maintain its mission of accessibility as the digital landscape transforms, particularly with the rise of AI and LLMs. Despite the rapid changes, these technologies present an opportunity to reassess the essential role of browsers, applications, and other digital tools in business and daily life.
- Multi Modal Relevance: The “MultiModalRelevance” metric assesses the relevance of a generated answer in relation to both visual and textual contexts using user input, responses, and retrieved contexts. The relevance score ranges from 0 to 1, with higher values indicating better relevance. An answer is considered relevant if it aligns with the provided visual or textual context, and the relevance score is discretely either 0 or 1, based on direct evaluation of these contexts.
- Compact Vision-Language With Open Weights, Faster Learning, Diffusion in Few Steps, LLMs Aid Tutors: This article by The Batch @ DeepLearning.AI discusses the merits and applications of fine-tuning models versus using simpler techniques like prompting and agentic workflows. Fine-tuning, while complex, can enhance model accuracy in critical applications, reduce latency, and mimic certain communication styles. It highlights that many teams might benefit from simpler methods instead. Additionally, Google has released the Gemma 3 models, which improve vision-language tasks and outperform previous versions through knowledge distillation and reinforcement learning, with potential for consumer hardware deployment.
- Audio Models in the API: OpenAI has launched the agent asdk, enabling developers to build custom voice agents, expanding from text to voice interactions. Agents, AI systems acting independently, can now communicate via voice, using either speech-to-speech or chained models. Speech-to-speech models are fast and suited for real-time API, while the chained approach involves converting speech to text, processing it with an LLM like GPT-4, and converting back to speech. Developers favor the modular, reliable, and easy start of the chained approach. OpenAI introduced new models, GPT 40 Transcribe and GPT 4 Mini Transcribe, and improved its Whisper model with advancements for efficiency and streaming audio-to-text APIs, featuring noise cancellation and a semantic voice activity detector for seamless voice experiences.
- 🦸🏻#14: What Is McP, and Why Is Everyone – Suddenly!– Talking About It?: The Model Context Protocol (MCP) introduced by Anthropic is gaining traction as a tool to streamline AI’s integration with external data sources and actions, enhancing AI agents’ effectiveness. Announced in November 2024, MCP facilitates AI connectivity to diverse contexts like files, tools, and databases through standardized interfaces, reducing the complexity of custom integrations. As a model-agnostic, open standard, it supports dynamic interactions and is backed by major AI players, offering a scalable, flexible, and secure way to enrich AI functionalities. Despite challenges like management overhead and evolving standards, MCP’s rapid adoption signals its potential to reshape AI systems and workflows by enabling more comprehensive and integrated AI capabilities across applications.
- Optimizing API Output for Use as Tools in Model Context Protocol: In the context of increasing interest in Agentic AI, Craig Walls discusses optimizing the use of APIs as tools within the Model Context Protocol (MCP). He illustrates the challenges he faced when using the ThemeParks.wiki API as an MCP Server, such as slow responses and high token usage due to large JSON payloads and endpoint mismatches. To address these, Walls optimized server-side handling by restructuring and filtering the data, which reduced unnecessary API calls and improved efficiency without hitting rate limits. The article highlights the necessity of adapting existing APIs for optimal integration with AI models, underscoring that the existing APIs often lack design considerations for such advanced uses.
- Qwen2.5-VL-32B-Instruct: Qwen2.5-VL-32B-Instruct, developed by Hugging Face, is an advanced vision-language model that has undergone enhancements to improve mathematical and problem-solving capabilities using reinforcement learning, offering refined human-aligned responses. Released after five months of feedback from developers using earlier versions, Qwen2.5-VL can visually recognize and analyze a wide range of objects and formats, function as a dynamic visual agent, comprehend lengthy videos, and provide structured outputs for complex data. Key improvements include dynamic video understanding via temporal resolution sampling, aiding in event pinpointing and object localization.
Management and Economics
- Apple Innovation and Execution: The text discusses Apple’s innovation trajectory and market strategy, highlighting criticisms that the company has become stagnant. Since the iPhone’s launch, Apple has introduced the iPad, Watch, and AirPods, but these products and their associated upsells mostly target existing users, limiting growth. While services have expanded financially, they lack distinct innovation. Apple’s attempts to branch into new tech like XR and AI have struggled, as demonstrated by the Vision Pro’s shortcomings and delays with a new Siri feature. These challenges have raised concerns about Apple’s execution capability, though past narratives claiming company missteps have not always been accurate.
- #229 Satya: Satya Nadella became the CEO of Microsoft in 2014, succeeding Bill Gates and Steve Ballmer during a challenging period when Microsoft was losing ground to competitors. Under Nadella’s leadership, Microsoft experienced a significant revival by focusing on empathy, innovation, and adaptability. He championed a “mobile-first, cloud-first” strategy, embraced open-source and cross-platform compatibility, and made strategic shifts such as exiting the mobile phone market. Nadella prioritized cloud computing and Artificial Intelligence (AI), significantly advancing Microsoft Azure and investing in AI through initiatives like OpenAI. Recognizing the potential of quantum computing, he led Microsoft to develop a topological qubit, aiming for breakthroughs in computational capacities. Nadella’s approach also transformed Microsoft’s internal culture from a competitive and siloed organization to one fostering openness and continuous learning. His leadership style emphasized empathy and humility as foundational to Microsoft’s renewed success, positioning the company as an innovator in emerging technologies.
- Colapso Del Product Market Fit: The text discusses the dynamic nature of Product Market Fit (PMF) and how it is perceived as a moving target influenced by evolving consumer expectations and technological advancements. It emphasizes that PMF should not be considered a static milestone but a continuous process requiring expansion and adaptation. With the advent of AI, the threshold for PMF is rapidly increasing, leading to potential collapses as seen in the cases of Chegg and Stack Overflow. Companies face challenges in maintaining their market position as AI disrupts traditional growth models, offering faster, more personalized solutions. Firms must predict these disruptions and strategically adapt, focusing on understanding changing customer expectations and risky product-market fits.
- «¡Ignorante! Facturamos Más, Ganamos Más, Valemos Más»: ¿Ah Sí? Inditex Y Telefónica Te Lo Explican: The article contrasts Inditex and Telefónica, analyzing how each company generates and manages wealth. Inditex, largely owned by Amancio Ortega, excels in innovation and the efficient use of capital, maintaining high profitability without the need for extensive borrowing. It invests its profits efficiently and rewards its shareholders handsomely. In contrast, Telefónica generates higher sales but has significant debt obligations, limiting its financial flexibility. The article emphasizes the importance of the Return on Invested Capital (ROIC) as a metric for financial value and criticizes assumptions that revenue alone determines a company’s worth. Inditex’s ability to reinvest its earnings without resorting to debt gives it a strategic advantage, while Telefónica’s financial structure involves more external shareholders and debt reliance.