Real estate
- ‘Scout24 Q3 2025: Revenues Up 15% as Subscriptions Drive Growth’: Scout24’s Q3 2025 was strong: revenue €165.6m (+15% YoY), EBITDA €97m (+13.8%), margin 62.9% (flat). ImmoScout24 listings +15.7%. Subscription-led growth: Private rev €46.7m (+17.8%, customers 526,740 +14.5%); Professional rev €119m, subs nearly €87m; transaction enablement +14.5% (slower vs Q2). Added to DAX. Guidance nudged up. Expansion: buying Spain’s Fotocasa and Habitaclia (>€150m), closing Q1 2026.
- ‘Zillow CEO Fires Back, Vows to Execute Despite “External Noise” as Q3 Revenue Climbs 16%’: Zillow’s Q3 revenue rose 16% to 676m with a third straight net profit (10m) and EBITDA 165m. Rentals jumped 41% to 173m on 69k listings; mortgages up 36% to 53m; core leads up 7% to 435m vs a 5% market. Traffic grew 7% to 250m uniques. Amid legal clashes with CoStar and allegations over photos and rental collusion, CEO Jeremy Wacksman touted Showcase on 3.2% of new listings and an AI-driven, integrated transaction hub, vowing to execute despite external noise.
AI
- ‘Building Etsy Buyer Profiles With LLMs’: Etsy is testing privacy-first LLMs to convert browsing and purchase signals into structured buyer profiles with explanations. With BigQuery, shorter histories, batching, parallelism, and Airflow, they cut generation for 10M users from 21 to 3 days and lowered cost by 94%. Profiles power query rewriting and personalized refinement pills. Accuracy is measured via CTR, conversion, and engagement. Profiles refresh for drift/seasonality, and cold starts use collaborative filtering.
- ‘Efficient Visual Representation Learning and Evaluation’: Etsy builds efficient visual embeddings for search and recommendations. Starting with EfficientNetB0 and adopting EfficientFormer-l3, they fine-tune multitask models (taxonomy, color, reviews) and evaluate via recall on NN retrieval, plus a text-to-image diffusion–based dataset. Models run on modest GPUs and low-latency mobile. EfficientFormer boosts offline recall and drove lifts in CTR, purchase rate, and clicks across ads and search-by-image.
- ‘Understanding Etsy vast inventory with LLMs’: Etsy uses LLMs to turn unstructured titles, descriptions, and photos into structured product attributes to power search and ML. Prior ML struggled with the long tail; now LLMs, with context engineering, few-shot prompts, taxonomy rules, and human-reviewed silver labels, run in a validated, observable pipeline. Results: attribute coverage rose from 31% to 91%, boosting filter engagement, conversion, and enabling features like color swatches.
- ‘From Image Classification to Multitask Modeling: Building Etsy’s Search by Image Feature’: Etsy built Search by Image by turning photos into embeddings with a fine‑tuned EfficientNet CNN. Category classification gave good taxonomy but weak visual cohesion; triplet loss improved visuals but hurt categories. They adopted multitask classification with heads for category, fine-grain, color and attributes plus a balanced sampler, and added buyer review photos to reduce bias. Retrieval uses an ANN IVF index and real-time GPU inference.
- ‘How I Use Every Claude Code Feature’: Shankar surveys Claude Code, urging focus on outcomes. Make CLAUDE.md the repo’s constitution: concise guardrails, pointers, synced with AGENTS.md. Check /context; avoid auto-compaction; reboot with /clear or document-and-clear. Keep slash commands minimal. Skip custom subagents; let main agent delegate via Tasks. Use resume/continue and logs; enforce commit-time hooks; plan. Prefer Skills and CLI scripting, use MCP as a secure gateway. Leverage SDK; tune settings.json.
- ‘The Tool That Lets You Switch Models Without Losing Your Place’: Droid, Factory’s command-line AI agent, lets you switch between GPT and Claude mid-task while carrying context, aided by subagents. At Every, Danny Aziz and Kieran Klaassen used multi-model panes to plan, build, and refine features in Spiral and Cora; Ben Tossell automates tasks like P&L and YouTube transcripts. Design choices boost SWE-bench results. In Droid Camp, they shared workflows showing gains for both developers and non-coders.
- ‘🐳DeepSeek-Ocr: Run & Fine-Tune’: DeepSeek-OCR is a 3B-parameter vision model for OCR and document understanding, using context optical compression to convert 2D layouts into tokens for efficient long-context processing. Unsloth shows it can be fine-tuned; on a 200K Persian dataset, after 60 steps (bs=8), mean CER dropped from 149.07% to 60.81% on 200 samples—an 88.26% absolute gain, making it 57% more accurate. A free notebook demonstrates the workflow.
- ‘The AI-centric Imperative: Navigating the Next Software Frontier’: McKinsey says software is entering an AI-centric era as gen and agentic AI redefine products, users, and organizations. Advantage shifts from features to data across three agent models (augmentation, agent-centric, expert). Pricing moves from seats to consumption/outcomes; delivery becomes vertical and service-as-software. GTM targets the C-suite with FDEs and partners. Winners will embed AI across PDLC and operations, modernize infrastructure, upskill talent, and execute a focused plan.
- ‘Property Finder Secures US$250 Million From Ares Management to Accelerate Expansion and AI Innovation’: Dubai-based Property Finder secured 250 million debt financing from Ares to accelerate regional expansion, AI and product innovation, marketing and partnerships. The firm reports over 40% revenue CAGR since 2020; UAE revenue rose from 30m (2021) to 117m (2024) and 73m in H1 2025, with EBITDA margin above 60%. It follows a 525 million minority stake by Permira and Blackstone; General Atlantic remains a major shareholder.
- ‘Text2Topic: Multi-Label Text Classification System for Efficient Topic Detection in User Generated Content With Zero-Shot Capabilities’: Text2Topic is a multi-label system for Booking.com UGC, with 239 description-based topics and 1.6M text-topic pairs from 120k texts via smart sampling/partial labeling. A bi-encoder (concat/subtract/multiply) enables fast zero-shot detection, beating MUSE and GPT-3.5; a cross-encoder is more accurate but slow. At scale it supports property typing, fintech message routing, and improved recommendations.
- ‘MCP Colors: Systematically Deal With Prompt Injection Risk’: Simon Willison highlights Tim Kellogg’s MCP Colors, a scheme to cut prompt injection risk in MCP agents by labeling inputs and tools: red for untrusted instructions, blue for critical actions. Store the color in tool metadata, block red-blue mixing at runtime, and batch-label many tools with clear criteria (even via an LLM) for scalable safety.
- ‘Amazon Nova Multimodal Embeddings: State-of-the-Art Embedding Model for Agentic RAG and Semantic Search’: AWS introduced Amazon Nova Multimodal Embeddings on Bedrock, a unified model for agentic RAG and semantic search that embeds text, documents, images, video, and audio for crossmodal retrieval with leading accuracy. It supports 8K context, up to 200 languages, sync/async APIs, segmentation, and MRL-trained outputs at 3072/1024/384/256 dims to balance accuracy, latency, and resource use.
Data Science
- ‘Unas Cuantas Notas Sobre Estadística’: Notes on statistics: model quality hinges on information in training data; adding noise can curb overfitting via ridge-like regularization. Gelman’s case study shows statistical craft colliding with computational pitfalls. Advocates Bayesian inference and cites Harrell’s Bayesian slides. Observes JAX’s rise (NumPyro vs Stan). Reminds that relationships are rarely linear and that binning continuous variables is harmful.
Economics
- ‘¿Para Qué Estudiar, Si Todos Acabaremos en El Paro?’: Juan Luis Jiménez argues to secondary students and educators that studying boosts employment chances and earnings, reduces unemployment and poverty risk, and correlates with living 2–10 years longer, in Spain and across Europe. These benefits persist beyond family background. While no degree guarantees a specific job and outcomes vary by field, education widens choices, builds critical thinking, improves society, and raises the odds of a better life.
- ‘El BCE Congela Los Tipos en El 2% Y El Mercado Augura El Fin Del Boom Hipotecario’: The ECB held rates at 2% for a third time, opting to wait amid controlled inflation and trade uncertainty. Markets expect no near-term cuts and a fade of the mortgage boom: growth slows, banks tighten and avoid price wars, targeting prime clients. Offers are still attractive at low fixed/mixed rates, while Euribor stays around 2.1–2.2 in a plateau, easing variable loans. High home prices hurt affordability. December’s meeting may clarify the outlook.
Software Engineering
- ‘Crafting Engineering Strategy!’: Notes on writing Crafting Engineering Strategy: Larson pursued his 2020s goal of three engineering books, pivoting from Infrastructure Engineering to strategy to create an anchor for the field. He wrote all prose, using LLMs for copy-editing and tooling, plus an LLM-optimized edition with textified images and link refs. He launched a dedicated, referencable site to advance the industry, expects more LLM-ready books, and published with O’Reilly for higher quality.
Others
- ‘Populism Fast and Slow’: Joseph Heath argues populism is less an ideology than a strategy that privileges fast, common-sense intuition over slow, analytic reasoning. It exploits gaps between public intuition and elite evidence (crime, trade, immigration, climate), amplified by social media. Populists’ disinhibited style seems “honest,” fueling illiberalism and conspiracism, and channels resentment at cognitive elites—blunting the left’s appeal.
- ‘Qué Es Un Commonplace Book Y Por Qué Deberías Tener Uno’: Patricia advocates keeping a commonplace book: a notebook to collect quotes, stories, and ideas plus your own reflections, turning information into dialogue. In an age of overload and short videos, it helps you process, remember, and revisit what matters. With roots in antiquity and the Renaissance, it can be an extension of the mind. Tips: schedule time, create a simple ritual, prioritize function over aesthetics, and evolve your organization.
- ‘Por Qué Leer Libros Largos Es Una Buena Inversión’: Juan Carlos Blanco argues long books pay off. PISA data show teens who read 100-plus-page books are about a grade ahead in comprehension, controlling for background. OECD’s Miyako Ikeda adds they also excel at digital skills like separating fact from opinion. In a world of addictive scrolling, long-form reading sharpens attention, resists manipulation, benefits all ages, and offers a career edge. Invest in reading long texts.
- ‘¿Problemas De Atención Y Memoria? Este Neurocientífico Sabe Qué Le Pasa a Tu Cerebro’: Neuroscientist Javier Ortiz Tudela explains that novelty drains cognition as the brain builds new “folders”. Context drives how we process, attend and remember: surprise grabs attention but often yields weak memories, while rich context anchors recall. This informs design, teaching and ADHD support by structuring info for attention vs memory. Brains are highly plastic: amid the digital shift, skills like long reading are recoverable in weeks, reflecting a transitional mismatch.