AI is bundling into trusted vertical stacks where workflow hooks and proprietary data—not another chat veneer—decide wins; I’m biased toward marketplaces and data moats, and real estate is the lab: Zillow/Rightmove’s AI modes look like rails, while Walmart’s 3x‑worse ChatGPT checkout shows what happens when the UI outruns the data. On the build side, work is going agentic and model‑centric: I’m the foreman, models are the crew, and the compounding comes from infra and roles (Git for agents, editing AI writing, database shake‑ups, the rise of the data‑infra PM) that let systems ship while we supervise.
AI
- ‘AI”s Bundling Moment’: AI is rebundling software after the SaaS era of point solutions. Rapid model change makes buyers seek trusted, multi‑year platforms, not stitched stacks. Firms are verticalizing: Harvey extends to professional services, Glean to Work AI, ElevenLabs to voice agents; OpenAI and Anthropic build industry teams. By lowering cognitive load and learning workflows, platforms scale faster. In AI, breadth and trust win over narrow specialization.
- ‘Editing AI Writing’: Every’s “Editing AI Writing” shows how AI unlocks value when tackling real pain: using Atlas agents streamlined hiring tasks. Codified editorial taste via a 400-rule style guide in Claude lifts draft quality and shifts editors to higher-level judgment. With late-2024/2025 model advances (e.g., Claude Code, Cowork), small teams can scale like big ones—if they adopt new workflows and learn from peers.
- ‘I Have Seen the Compounding Teams’: Sam Schillace profiles “compounding teams” whose output grows fast by building proactive, model-centric frameworks instead of just using Copilot-like helpers. They let AIs make and use tools, commit them to git, and work via programmer infra (filesystems, git, Markdown, Kubernetes). With many parallel runs, human attention is the bottleneck. Success hinges on modular, test-driven coordination; mixing with hand coding hurts. This is spreading to wider knowledge work: a platform shift.
- ‘Coding After Coders: The End of Computer Programming as We Know It’: AI coding agents are shifting programming from handcrafting to directing: developers describe intent, bots write and test code, boosting productivity from modest gains at big firms to orders of magnitude at startups. Many coders welcome the creativity, others fear skill loss, quality bloat, ethics, and job erosion, especially for juniors. Work becomes judging, system design, and conversation. Cheaper coding may expand demand and spread software to more people.
- ‘How StrongDM”s AI Team Build Serious Software Without Even Looking at the Code’: Simon Willison profiles StrongDM’s software factory: agents generate and verify code with no human writing or review. Spurred by 2024–25 model gains, they tackle validation via scenario testing: end-to-end, holdout user stories judged probabilistically with a satisfaction metric. A Digital Twin Universe stress-tests systems. Techniques include Gene Transfusion, Semports, and Pyramid Summaries. Core question: how to prove AI-built software works.
- ‘Thoughts on OpenAI Acquiring Astral and Uv/Ruff/Ty’: Simon Willison reacts to OpenAI acquiring Astral (uv, ruff, ty). Astral joins the Codex team, pledging to keep building in the open; OpenAI pitches the deal as accelerating Codex with Astral’s tooling and talent. He highlights uv’s central role and huge adoption, with ruff/ty aiding AI coding. OpenAI’s OSS track record is limited, but Astral’s permissive licenses keep a forkable escape hatch; overall he’s cautiously optimistic.
Real estate
- ‘Zillow Debuts AI Mode, Bringing Guided Intelligence to Every Step of the Housing Journey’: Zillow launched AI mode, a conversational tool that helps buyers and renters through the entire housing journey: discovering homes, comparing options and tradeoffs, estimating repair costs, gauging affordability, providing neighborhood insights, interpreting price cuts, scheduling tours, and connecting to agents. Built on verified data with transparent, fair-housing guardrails and privacy, it follows regulations and supports—not replaces—licensed professionals.
- ‘How Zillow’s New AI Mode Works Throughout the Real Estate Journey’: Zillow’s new AI mode uses a coordinated multi‑agent system to answer housing questions in natural language, invoking skills for search, financing and valuation. It interprets intent, routes to tools, and keeps context across visits. With computer vision and multimodal models, it reads home photos for layout, light and condition to compare options. Integrated with Zillow, it turns chat into actions—similar homes, affordability, tours, agents—backed by housing‑specific safeguards and compliance.
- ‘Rightmove Launches Next Phase of AI-powered Property Search’: Rightmove, the UK’s largest property platform, launched a beta AI conversational search built with Google’s Gemini, accessible from its homepage. Users can describe needs in plain language—beyond location, beds, and price—to get tailored shortlists (e.g., commutable areas, solar panels, near trains). It complements core search, starts with a limited rollout, and will evolve. It builds on AI Keywords (smart prompts scanning listings) and Style with AI (visualizing home styles).
- ‘The New Industry Leader in AI Home Search? In Conversation With the Team Behind Homes AI’: CoStar’s Homes AI is a speech-to-speech, multilingual search on Homes.com that blends AI with the site UI to filter listings, swap images, and handle open-ended queries via proprietary data and a Microsoft-secure stack. Promising but early, it has lag, errors, and high costs; upgrades (memory, richer signals) are planned. Marketing and agent workflows are pivotal, with rollouts to other CoStar portals ahead, aiming for faster, fun, agent-aligned home searches.
- ‘Why Data, Not AI, Will Decide the Future of Property Search’: AI hype in property search outpaces reality: only 23 of about 900 portals offer AI search, many testing or rolling back. Benchmarks show the bottleneck isn’t AI or interface but data. Systems excel on objective, structured fields (beds, price) and fail on subjective or contextual needs (safety, pollution). LLMs parse intent yet surface unreliable listings. Without richer data sets like image, geospatial, and environmental context, accuracy lags; data depth will decide the future.
- ‘Can AI Help Me Find a Home?’: AI is reshaping homebuying by improving search, pricing and upfront information, potentially halving UK transaction times. Innovations include 3D digital twins, photorealistic virtual tours, de-furnishing tools, and natural language search with smart suggestions; portals also identify serious buyers. Yet agents remain essential; those who leverage AI to analyze behavior, advise faster, and personalize service will keep a competitive edge.
- ‘VerbaFlo Raises $7 Million to Build Al Communications Platform for Global Real Estate’: VerbaFlo.AI raised 7 million in seed funding led by Pi Labs, with Haatch, Navigate Ventures, Old College Capital, the University of Edinburgh’s venture arm, global family offices, and follow-on anchor investors. Funds will drive US and global expansion, product development, and hiring. Its vertical AI platform for residential real estate integrates with existing systems to automate and optimize resident lifecycle conversations across 40+ use cases.
- ‘Inquilinos Cada Vez Más Ricos: La Paradoja Del Esfuerzo en Alquiler’: Spain, especially Catalonia, shows falling rent-to-income among market renters, but due to selection: tighter supply pushes low-income households out to family/informal housing. Median renter income rose (notably in Catalonia) and the share of bottom-quintile renters fell while higher-income shares grew. Evidence on rent controls shows reduced supply, misallocation and spillover price rises. Using rent effort as success confuses selection with real access.
Data Science
- ‘AI Is Redrawing the Database Market’: AI is redrawing data platforms: apps become agentic, analytics conversational, and observability AI-driven, requiring high concurrency, real-time, full-fidelity data. Legacy warehouses lag. The stack is converging on unified transactional+analytical systems, notably Postgres with ClickHouse. ClickHouse positions as a unified AI platform—columnar speed, open formats, ClickStack for observability, LibreChat for agents, and LLM observability—urging teams to replatform before switching costs rise.
- ‘Data Infra PM: The Most Important Role in Data Product’: As AI agents permeate products, their data-hungry, context-heavy behavior strains existing data stacks. A Data Infra PM is crucial to set strategy: rapidly realign foundations to agent needs, prioritize access over full democratization, control warehouse and token costs, and build a robust context layer. They must master agent patterns, challenge legacy designs, link spend to value, and combine business savvy with deep data/infra skills amid ambiguity.
- ‘How We Give Every User SQL Access to a Shared ClickHouse Cluster’: Matt Aitken presents TRQL, a SQL-style DSL that gives every Trigger.dev user safe access to a shared ClickHouse cluster. Its grammar forbids dangerous ops; the compiler injects tenant and time filters, hides internal schemas, and adds features like virtual columns. Using ANTLR and an AST, TRQL validates and parameterizes queries, compiles to ClickHouse SQL, runs read-only on replicas, and returns JSON.
Online Marketplaces
- ‘Walmart: ChatGPT Checkout Converted 3x Worse Than Website’: Walmart says purchases completed inside ChatGPT via OpenAI Instant Checkout converted at one-third the rate of click-outs to its site, calling the experience unsatisfying. OpenAI is phasing Instant Checkout out; Walmart, which had offered about 200k products since November, will embed its Sparky bot in ChatGPT to sync carts and finish checkout on Walmart, with a similar integration coming to Google Gemini.
Philosophy
- ‘El Santo grIAl’: AI is a new monolith, not a rupture: the old cycle returns—fear, resistance, journey, transformation. Like the Grail knights, the path is the darkest, trackless one; the aim is change, not goals. Modernity productivized knowledge, making us machine-like; AI merely mirrors and excels at that. What remains uniquely human is purposeless, initiatory knowing: a risky, solitary quest out of the cave, not a shortcut or group route.
- ‘La Ensalada Del McDonalds, Las Preferencias Declaradas Y Las Preferencias Reveladas.’: Antonio Ortiz contrasts declared vs revealed preferences on digital platforms: algorithms favor behavior over stated tastes to maximize engagement, creating guilty pleasures and reshaping interests. Spotify tests Taste Profile to let users adjust inferred tastes. He likens feeds to ultra-processed food as supernormal stimuli. On food prices, he challenges blaming supermarkets, noting lower household food burden and limited role of retail margins in Spain.
Technology
- ‘How ROOST Is Advancing Online Safety’: Discord co-founded ROOST to make open, shared, auditable safety tools that counter fast-evolving, AI-driven threats. It open-sourced Osprey, a production rules engine for real-time event analysis and enforcement, plus Coop for reviews. The same Osprey runs at Discord and is adopted by platforms like Bluesky, with vendors offering managed services. This open ecosystem accelerates innovation, raises baseline protections, and strengthens safety for hundreds of millions.
- ‘A New Era of Personalization: Shape Your Taste Profile on Spotify’: Spotify is launching a beta Taste Profile that lets you see and shape how it understands your taste across music, podcasts, and audiobooks. Rolling out to Premium users in New Zealand, it enables asking for more or less of certain vibes and using habits (workouts, commutes) to refine homepage recommendations.
Software Engineering
- ‘La Bonilista — Qué Carallo Son Los Forward Deployed Engineers…’: Forward Deployed Engineers, popularized by Palantir, embed with clients to model data, integrate systems, and iterate solutions, solving problems rather than tickets. On-site work often stems from sensitive data. Palantir reframed this classic services role with elite branding (Deltas). As AI cheapens coding, advantage shifts to tailoring and implementation; FDEs spearhead product and sales. Tech firms and consultancies now expand FDE teams, blurring product and services.
- ‘Using Git With Coding Agents’: Git pairs well with coding agents: keep work in version control, let agents init repos, commit, and sync with GitHub, load context via git log, and choose merge strategies. Agents can resolve conflicts, recover code via reflog or stash, run tests, and drive git bisect to find regressions. Treat history as an editable narrative; agents can rewrite and curate it to aid future development.