I haven´t been updating the links for three weeks so this one is packed! State of AI report, Claude Code skills/plugins and DIY agents, Maps grounding, with Brave’s Comet hack reminding me how risky that power is. I also followed the business front: real estate’s platform fight (CoStar vs Zillow, fair-housing classifiers, ChatGPT app), the rise of AI services and leaner teams.
Real estate
- ‘CoStar Group CEO Andy Florance on Zillow, M&A, and Domain”s Key Lime Pie Problem’: CoStar CEO Andy Florance says challengers can topple leaders if incumbents mistreat customers, criticizing Zillow’s defense as sloppy and hinting at cynical content-grab tactics. On acquiring Domain to take on REA Group, he argues Domain drifted into low-yield distractions and outdated practices. His priority: build a best-in-class core portal, use CoStar’s product expertise in Australia, and remove obsolete directives to sharpen execution.
- ‘Zillow debuts the only real estate app in ChatGPT’: Zillow launched the only real estate app in ChatGPT. Typing “Zillow” surfaces real-time listings with photos, maps, pricing, and chat-driven filters for sales or rentals, mirroring Zillow’s design and sorting. Users can explore neighborhoods and affordability, then jump to Zillow to tour, contact agents, or start financing. Built with OpenAI, it includes broker/MLS disclosures and strong fair-housing safeguards to reduce bias.
AI coding agents
- ‘How I”m Using Coding Agents in September, 2025’: Author outlines a disciplined workflow for AI coding agents: Claude Code guided by a CLAUDE.md and git worktrees; a one-question brainstorming phase producing short sectional designs; a detailed plan written to docs; a two-session model (architect vs implementer) executing tasks in small chunks with resets; PRs reviewed by CodeRabbit, aided by a helper that aggregates comments; and a role-play prompt that critiques reviewers to avoid blind changes.
- ‘Superpowers: How I”m Using Coding Agents in October 2025’: Jesse Vincent outlines Superpowers, a Claude Code plugin that teaches agents reusable skills, defaults to plan-then-implement, sets up git worktrees, dispatches tasks to subagents, enforces RED/GREEN TDD, and can open PRs. He TDD-tests new skills (even distilled from books), connects ideas to Microsoft Amplifier and persuasion research, shares a demo, plans community skill sharing, and adds a memory skill that indexes past chats for retrieval.
- ‘Breaking New Ground in Generative AI: Zillow’s Fair Housing Classifier’: Zillow, known for the Zestimate since 2006, launched an open-source Fair Housing Classifier for generative AI. It detects language that may violate federal, state, or local antidiscrimination laws to help prevent illegal steering in real estate conversations. Many LLMs don’t account for fair housing by default; this tool flags noncompliance while developers decide responses, promoting responsible, fair housing practices across sectors.
- ‘Everyone Should Be Using Claude Code More’: Lenny Rachitsky says Claude Code is an underrated local AI agent. Think of it as Claude Local: it runs on your computer, handles huge files and long tasks, and is fast. It can organize files, enhance images, brainstorm names, summarize calls, create tickets, clear storage, and more. Dan Shipper uses it to spot conflict avoidance in meeting recordings.
AI
- ‘Agentic Browser Security: Indirect Prompt Injection in Perplexity Comet’: Brave highlights security risks in agentic browsing. They found Perplexity’s Comet vulnerable to indirect prompt injection: when summarizing a page, Comet treats hidden on-page text as commands. In a demo, a Reddit comment made it log in, read a Gmail OTP, and exfiltrate data, aided by a trailing-dot domain dodge. Since the AI operates with the user’s privileges, SOP/CORS don’t help, underscoring the need for new protections.
- ‘About Claude Code Agents’: Claude Code subagents are pre-configured AI assistants for specific tasks. Each has a defined purpose, its own context window, allowed tools, and a custom system prompt. When a task matches a subagent’s expertise, Claude Code delegates it, enabling efficient, independent problem-solving.
- ‘Teaching Claude Code My Obsidian Vault’: Author fixes AI’s blank-slate issue with a CLAUDE.md in an Obsidian vault that Claude Code auto-loads to learn vault structure, coding prefs, commands, and where info lives. Using Spotlight mdfind and textutil, it searches notes and PDFs, automates daily notes, and tailors research to interests. Setup is simple (markdown + shell). Benefits: less context switching. Includes a quick how-to and future ideas: calendar, transcripts, repos, RSS.
- ‘Picos Y Palas Digitales: Quién Gana De Verdad en La Fiebre De La IA’: AI’s gold rush funnels gains to a closed loop of chip, cloud and model firms (NVIDIA, OpenAI, Oracle, AMD, CoreWeave) that cross-invest, inflating valuations on promise amid heavy energy use. Meanwhile, consultancies like Accenture, Capgemini, Infosys and Globant quietly turn AI into productivity and cash yet are penalized. AI will redefine, not erase, consulting; those who integrate it win. When the froth clears, execution beats narrative.
- ‘Equipping Agents for the Real World With Agent Skills’: Anthropic presents Agent Skills: composable folders centered on SKILL.md (with linked files and tools) that Claude can discover and load via progressive disclosure to gain domain expertise. Skills can bundle code for efficient, deterministic operations and scale beyond context limits. Best practices: evaluate gaps, structure for scale, think like Claude, iterate. Security: install trusted, audit others. Roadmap: full lifecycle features and MCP integration.
- ‘Announcing Amazon Quick Suite: Your Agentic Teammate for Answering Questions and Taking Action’: Amazon announces Quick Suite, an agentic teammate that unifies research, BI, and automation in one workspace. Quick Index builds a secure, searchable knowledge base across docs and apps. Quick Research delivers plan-driven, cited insights; Quick Sight offers natural-language analytics and actions from dashboards; Quick Automate and Quick Flows orchestrate simple to complex workflows. Spaces and customizable chat agents add context, with enterprise integrations and governance.
- ‘El Impacto De La Ia en Desarrollo’: Manfred’s Jul–Sep 2025 survey (1100, mostly seasoned devs) finds 9/10 already use AI. ChatGPT leads (72%), Copilot ~50%; Claude 32% (stronger with seniors/small firms) and Gemini 30% (more in medium/large). Adoption is higher and freer in small companies. 73% report productivity gains; about two thirds see quality improvements, though most code needs edits. Training is scarce (49% none, 34% self-taught) and access uneven. Most would allow AI in tech tests; some fear future displacement.
- ‘Vibe Check: OpenAI DevDay 2025’: OpenAI’s DevDay 2025 launched AppsSDK for in-ChatGPT apps, AgentKit for no-code workflows with built-in evals, and opened GPT-5 Pro, Sora 2, and a cheaper voice model to the API. It doubled down on ops and app-store bets; speed and a limited partner list are concerns. Codex GA underwhelmed. Biggest win: broader access to top models. ChatGPT grew from 100M to 800M weekly users, yet the event felt short on vision and inspiration.
- ‘Managing AI Is Like Managing Humans, With One Big Exception’: AI enables leaner teams and blurs manager/IC roles; ICs must adopt management skills. Core levers still apply—purpose, people, process—via clear prompts, choosing the right models, and verifying with evals. High-agency people thrive as AI drafts, summarizes, and mediates collaboration. The big difference: AI lacks physical and emotional needs. Amid upheaval, leaders should stabilize—balancing candor and hope—while building trust and helping people grow.
- ‘Lazy Parents Are Giving Their Toddlers ChatGPT on Voice Mode to Keep Them Entertained for Hours’: Parents are using AI voice chatbots like ChatGPT to entertain or soothe toddlers for hours, prompting concerns about screen-time harms and mental health. Reports link chatbots to self-harm risks and delusion, and safeguards remain weak. Toymakers still push AI into kids’ products. Researchers warn children may anthropomorphize AIs and mistake agency, while AI image tools can displace imagination and drawing. Anecdotes show growing, everyday reliance.
- ‘The Rise of AI Services’: AI services are surging, with Accenture scaling AI revenue from 100m in 2023 to 900m in 2024 and 1.8b by mid 2025, plus 5.9b in bookings. General Catalyst launched Percepta, an AI consulting arm, chasing this demand. Because many AI projects fail due to data, tool choice, and change management, firms use implementation services, forward-deployed engineers, and AI dev shops to bridge the gap and deliver real use cases.
- ‘Unlock the Power of Images With AI Sheets’: Hugging Face launched a major AI Sheets update, adding vision to its open-source, model-agnostic spreadsheet. Leveraging Inference Providers, it taps thousands of open models to display, analyze, and transform images: captioning, classification, OCR, extraction, metadata tagging, detection, colorization, editing, and generation. Use it to parse receipts, digitize handwritten recipes, create visuals at scale, and export results to CSV or Parquet.
- ‘An Opinionated Guide to Using AI Right Now’: Ethan Mollick’s guide: pick from a few leading AIs (ChatGPT, Claude, Gemini, Grok or open models). Free tiers often suffice; upgrade to 20 or 200 per month for heavier, technical work. Prefer agent models, use Deep Research, and connect your data. AIs can produce docs, code, images, and video; be wary of errors and sycophancy. Give context, don’t obsess over prompts, and experiment—building intuition matters as tools evolve quickly.
- ‘Introducing Claude Haiku 4.5’: Anthropic introduced Claude Haiku 4.5, a small, fast, affordable model available to all users, offering Sonnet 4–level coding performance at about one-third the cost and over twice the speed. It even surpasses Sonnet 4 at computer-use tasks, boosting Claude for Chrome and Claude Code. Sonnet 4.5 remains the frontier model and can orchestrate multiple Haiku 4.5s to execute plans in parallel.
- ‘Inteligencia Artificial Y Desvincular Esfuerzo Y Resultado’: Antonio Ortiz argues that AI breaks the link between effort and results in intellectual and creative work, cheapening signals of merit. Newcomers can produce acceptable outputs, sparking perceptions of “cheating” and risking intellectual atrophy. As technique commoditizes, the edge shifts to taste, ideas, and better questions. He favors using AI—sometimes abstaining—and expects a messy phase of laziness before society adapts.
- ‘Claude Skills Are Awesome, Maybe a Bigger Deal Than MCP’: Willison says Claude Skills are a major advance: Markdown files (plus optional scripts) that models load on demand via YAML summaries, saving tokens. They already power document creation (.pdf/.docx/.xlsx/.pptx) and are easy to share. Versus MCP’s heavy context, skills with a coding sandbox act like lightweight, CLI-like extensions. They enable broad automation—from Slack GIF sizing to data journalism—while raising sandboxing and security needs.
- ‘State of AI Report’: 2025 AI centered on reasoning: OpenAI’s GPT-5/o1 lead while China’s DeepSeek/Qwen push strong open weights; closed models still ahead. Benchmarks wobble, RL gains mixed, parallel think‑then‑act spreads. Industry surges on chips and neoclouds but power/compute bottleneck; sovereign AI and US‑China politics reshape supply chains. Safety tightens amid rising cyber misuse; CoT monitorability and deception debated. Coding, agents, video, and AI search scale; jobs and rules lag.
- ‘Another New AI Browser, Dia, Launches for Free to Mac Users’: Dia, a new AI-powered web browser, is now free for Mac users with M1 or newer chips after a summer beta. It works like a standard tabbed browser but adds a built-in chatbot to draft emails, compare info across open tabs (e.g., Airbnbs), and help students. The site even shows it advising against buying a 400 sweater. Dia joins a growing field of AI browsers, including Perplexity’s Comet and Opera’s Neon.
- ‘Grounding With Google Maps: Now Available in the Gemini API’: Google launched Grounding with Google Maps in the Gemini API, letting apps fuse Gemini’s reasoning with up-to-date Maps data from 250M+ places for geospatial-aware experiences. It auto-detects location context, supports lat/long, and can return an interactive widget with photos, reviews, and details. Use cases span travel, real estate, retail, logistics, with hyper-local recommendations. Combining Maps with Google Search improves quality, pairing venue facts with live info like show times.
- ‘Customize Claude Code With Plugins’: Claude Code now supports plugins—bundles of slash commands, subagents, MCP servers, and hooks—installed via /plugin and toggleable to reduce context. Use them to enforce standards, aid users, share workflows, connect tools, and bundle setups. Anyone can host marketplaces (marketplace.json in a repo) and install with /plugin marketplace add. Community marketplaces offer many options. Plugins are in public beta across terminal and VS Code.
- ‘Gemini 2.5 Computer Use Can Solve Google”s Own CAPTCHAs’: Simon Willison tested Gemini 2.5 Computer Use. It hit Google CAPTCHAs (likely due to Browserbase IPs), solved several, then searched and opened Hacker News but performed weakly by checking only one thread. The model card calls for confirmations on high‑stakes actions, not CAPTCHA avoidance. He was impressed by its precise mouse clicks, likely aided by prior bounding‑box and segmentation skills, though a formal eval is needed to judge superiority.
Others
- ‘Doctors Just Found Something Fascinating About What Happens When You Drink on Ozempic’: GLP-1 drugs like Ozempic may not only reduce cravings but also blunt alcohol’s effects. A Virginia Tech study found users who had three drinks in an hour felt less drunk and had slower, lower rises in blood alcohol than non-users. Likely due to slowed gastric emptying and altered reward signaling, this delay could reduce alcohol’s reinforcing impact and help people drink less, suggesting potential for treating addiction.
- ‘El Auge Universitario Ha Sido El Mayor Factor De Ruptura Interna en Occidente, Incluso Más Que La Inmigración’: Antonio Ortiz echoes Janan Ganesh: the university boom fuels Western polarization by creating a culturally dominant but electoral minority of graduates, splitting societies and shaping online moderation. Education also conditions views on immigration, widening elite-voter gaps. He adds that waiting enforces class divides (with paid line-skipping) and that in China technology is raising trust via surveillance and mobile apps.
- ‘La Vida Secreta De La Gente Que Muere Sola’: Notes map a luxury slump: Bordeaux first growths, jets, yachts, resale Rolex, top art and prime property slide; LVMH and Kering see sales fall and swap creative leads to spark a new cycle. As goods lose exclusivity, the ultra-rich pivot to scarce experiences. Also, a single high creatine dose after 21h awake can aid cognition and may spread beyond jet-lag use. Etiquette: say no, not maybe, unless truly uncertain.
Software Engineering
- ‘How We Built Our Own Claude Code’: Tinybird built Tinybird Code, a CLI agent inspired by Claude Code but specialized for real-time data engineering on ClickHouse. Using Pydantic AI, it runs a multi-agent architecture, secure backend-hosted models, and sync workflows. It offers interactive and one-shot modes, modular rendering, context compaction, robust error recovery, and workflow-oriented tools. It’s environment-aware for safe prod deploys and can compose with other agents for complex analytics tasks.
Data Science
- ‘DBT Labs + Fivetran: Open Data Infrastructure for Analytics and AI’: dbt Labs and Fivetran are merging, creating an open, interoperable data infrastructure platform with about 600 million ARR and 10k+ customers. Products stay separate and non-disruptive, focused on simplifying data engineering and working across any cloud. The open data infrastructure is pluggable and standards-based. dbt Core and Fusion keep current licenses; community support continues. Fivetran’s strong OSS work is noted, with plans to open more tech and invest in shared standards.
- ‘Everyone Says ‘Data Teams Should Drive Value’ but How?’: Everyone says data teams must drive value, but the how depends on their role. Define the team’s function and outcomes, not just titles: platform (enable reliable data), complicated subsystem (specialized models), stream-aligned (embedded, fast decisions), and enablement (upskill/bridge). Set clear expectations and scope to avoid catch-all work, prioritize by team type, and communicate impact in business terms tied to measurable results.
Management
- ‘You”re Definitely Going to Be a Manager Now’: Julie Zhuo updates You’re Definitely Going to Be a Manager Now with refreshed examples and new sections on managing remotely and in a downturn, gaps revealed by the pandemic and layoffs. She argues AI enables doing more with less, leading to leaner teams—sometimes one Captain Marvel beats summoning all the Avengers.
- ‘The Magic of Framing the Message’: Influence comes less from novel ideas or critique and more from framing. Beyond “say it, repeat it,” the hard part is shaping your message to what others can hear: learn what they care about and what keeps them up at night, then mirror their language to signal alliance, give ownership, and feel familiar. Start with “yes, and” to build common ground, use trios for stickiness, and wield these tools ethically. Leadership is sharp framing, not loudness.