This week’s highlights: the acquisition of pisos.com by Immobiliare.it and Sumar’s push to audit ranking algorithms in online real estate platforms. On the AI front, Barto and Sutton received the Turing Award for their foundational work in reinforcement learning. There’s also fresh thinking from Simon Willison on “vibe coding” with AI, and plenty of momentum around tools, models, and governance in the gen AI space.
Real estate and Marketplaces
-
Sumar Quiere Controlar Los Algoritmos Que Usan Idealista Y Fotocasa: Sumar, liderado por Yolanda Díaz, ha propuesto una iniciativa en el Congreso solicitando al Gobierno supervisar los algoritmos usados por portales inmobiliarios como Idealista y Fotocasa. El objetivo es regular estos algoritmos para evitar distorsiones en precios y proteger a consumidores y pequeños propietarios. La propuesta, apoyada por el Ministerio de Consumo y remitida a otros ministerios relevantes, busca que dichas plataformas garanticen transparencia, equidad y protección de datos personales, además de llevar a cabo auditorías periódicas.
-
Vocento Vende Pisos.com Al Grupo Italiano Immobiliare.it Por 22,5 Millones De Euros: Vocento has sold its real estate portal Pisos.com and related entities for 22.5 million in cash to the Italian group Immobiliare.it. This sale aligns with Vocento’s strategic plan, led by CEO Manuel Mirat, to focus on key areas such as journalism, data, and technology. Pisos.com, now the third largest real estate portal in Spain, will bolster Immobiliare.it’s presence in Spain, complementing its existing operations. The acquisition is part of Immobiliare.it’s broader strategy for growth and international expansion.
-
Sustainability PropTech Avani Solutions Raises 4.4M and Prepares UK and Europe Expansion: Avani Solutions, a sustainability-focused PropTech company, has secured 4.4 million in funding from EGX to support its expansion into the UK and European markets. The company’s platform leverages live data from building systems to reduce carbon footprints by minimizing resource waste and boosting operational efficiency. The funding aims to enhance the platform’s algorithms and integrated features, potentially saving the property industry $77 billion by optimizing operations and reducing inefficiencies.
-
Baltic Classifieds Group Acquires Untu: Baltic Classifieds Group (BCG) has strategically acquired UNTU.lt, a leading Lithuanian digital real estate valuation platform. With over five years of operation, UNTU.lt offers precise property valuation services, enhancing data-driven decision-making for users and real estate brokers by simplifying the broker selection process. This acquisition strengthens BCG’s service offerings, broadening its digital transformation in real estate. BCG, a major online classifieds operator in the Baltic region, aims to improve market transparency and expand its services through this integration.
-
AI-Led Data Platforms BuildingMinds and Optiml Announce Partnership: BuildingMinds and Optiml, two AI-led data platforms, have announced a strategic partnership to address industry-wide challenges by integrating ESG data with retrofit and investment planning capabilities at both real estate portfolio and asset levels. The collaboration offers benefits such as simplified data flows, enhanced retrofit analysis, comprehensive regulatory reporting, and optimized CapEx allocations. BuildingMinds focuses on improving asset operational decisions and regulatory compliance, while Optiml empowers asset managers with decision intelligence for decarbonization strategies. Together, they aim to support clients’ Net Zero ambitions and sustainable financial goals.
-
Uniti AI Raises 4M for Its AI Sales Agents for Global CRE Operators: Uniti AI, a leading AI agent platform for commercial real estate operators, has secured $4 million in a seed round led by Prudence, with contributions from several other partners. The company provides LLM-native platforms allowing real estate operators to deploy customizable AI agents tailored to their workflows. Uniti focuses on enhancing leasing and sales processes and plans to expand to broader operational use cases. The funding will help launch AI-powered voice agents and improve integrations with property management systems.
AI
-
2024 ACM A.M. Turing Award Laureates: The ACM has awarded the 2024 A.M. Turing Award to Andrew G. Barto and Richard S. Sutton for their foundational work in reinforcement learning (RL). Beginning in the 1980s, Barto and Sutton developed the mathematical and algorithmic frameworks that form the basis of RL, significantly advancing AI by enabling agents to learn optimal behaviors from reward signals. Their innovations include temporal difference learning and policy-gradient methods, and their textbook, Reinforcement Learning: An Introduction, remains influential. Their work laid the groundwork for major advancements such as AlphaGo’s victory over human Go champions and applications in areas like robotics and AI language models like ChatGPT. RL continues to have far-reaching impacts across various technological and scientific domains.
-
The-State-of-Ai-How-Organizations-Are-Rewiring-to-Capture-Value_final: Organizations are optimizing generative AI (gen AI) to create significant value, highlighting the crucial role of CEO-led AI governance, especially in large corporations. Effective AI integration demands a transformative top-down approach with C-suite engagement. Companies are redesigning workflows, which impacts EBIT, and are also adopting centralized or hybrid AI deployment models. The workforce impact is mixed, with some areas seeing reduced head counts, while others like IT experience growth. Gen AI is primarily utilized in marketing, product development, service operations, and IT, with larger organizations applying it more extensively than smaller counterparts. Though still early in AI adoption, with challenges like AI talent acquisition and risk management, the focus is shifting towards transformation and strategic prioritization rather than incremental deployment.
-
Transformers without Normalization: Jiachen Zhu’s work titled “Transformers without Normalization” introduces Dynamic Tanh (DyT), an element-wise operation, as a replacement for normalization layers in Transformers. Traditionally deemed essential, normalization layers may be unnecessary as DyT can match or exceed their performance. Inspired by tanh-like transformations in layer normalization, DyT simplifies Transformers across various tasks and settings without needing extensive hyperparameter tuning. This challenges the conventional view of normalization’s role in neural networks.
-
SmolDocling-256M-preview: SmolDocling-256M-preview, authored by Hugging Face, is a multimodal Image-Text-to-Text model focused on efficient document conversion while being fully compatible with Docling. It incorporates popular features of Docling, such as seamless support for DoclingDocuments and enhanced efficiency via DocTags, which separate text from document structure, simplifying processing for Image-to-Sequence models. It supports export to formats like HTML, Markdown, and JSON, optimizing performance by reducing token generation overhead.
-
Thoughts on AI: In “Thoughts on AI,” Davo reflects on the term “AI” and its widespread association with technologies like ChatGPT and large language models (LLMs). He expresses a preference for the term “Augmented Intelligence” due to AI’s diverse applications and argues that AI companies often misrepresent AI’s capabilities. Davo, with a background in microbiology and bioinformatics, believes LLMs are useful for tasks like debugging and template generation, but they don’t fully replace human intelligence due to issues like hallucinations. He emphasizes the importance of human oversight in using AI tools and comparisons to outsourcing mundane tasks, like remembering phone numbers, to technology.
-
#227 Software B2B en la era de la IA: In “#227 Software B2B en la era de la IA,” Samuel Gil explores how AI profoundly impacts but does not redefine B2B software. While AI pushes us to enhance and rethink current systems, foundational principles remain unchanged. Companies still need secure, scalable, and customizable software solutions, despite AI’s transformative potential. AI commoditization and open-source models pressure prices, but specialization in domains like finance or health sustains differentiation. Businesses should focus on solving specific client problems rather than offering standalone AI models. Klarna’s internal software strategy highlights AI’s potential to streamline operations, although this is not universally viable due to expertise requirements. AI now crucially influences business strategy, offering significant market expansion and consumer benefits while posing risks to traditional SaaS models.
-
Mistral Small 3.1: Mistral Small 3.1, introduced by mistral.ai, is a cutting-edge model excelling in text and multimodal understanding with an expanded context window of 128k tokens, outperforming competitors like Gemma 3 and GPT-4o Mini. It supports multiple languages and runs efficiently on devices with as little as an RTX 4090 or 32GB RAM, enabling fast and accurate responses ideal for virtual assistants, with capabilities for domain-specific fine-tuning. Available as an open-source model, it encourages community-driven development, offering base and instruct checkpoints for further customization.
-
Introducing Gemini With Personalization: Google has introduced “Gemini with personalization,” an experimental feature that integrates the Gemini 2.0 Flash Thinking model with Google apps, starting with Search. This allows Gemini to deliver uniquely tailored responses based on users’ past interactions, enhancing accuracy and relevance. With user consent, Gemini will soon connect with additional services like Photos and YouTube to provide deeper personalized insights, leveraging a more comprehensive understanding of user activities and preferences.
-
Some AI-Generated Works Are Copyrightable: The article from deeplearning.ai discusses advancements in AI and copyright issues surrounding AI-generated works. Cohere’s Aya Vision, a multilingual vision-language model, understands text and images consistently across languages, offering free use for noncommercial purposes. Google’s AI co-scientist generates research proposals using multi-agent systems, aiding scientific breakthroughs. The U.S. Copyright Office states AI-generated works can be copyrighted if a human provides significant creative input, rejecting the need for new legal frameworks. The office emphasizes flexibility in evaluating human contributions to AI-generated works.
Data Science and Software Engineering
-
Here’s How I Use LLMs to Help Me Write Code: Simon Willison discusses how he effectively utilizes language learning models (LLMs) to assist with coding. He emphasizes that while using LLMs is not intuitive, they can be extremely useful, particularly if seen as a confident, fast-learning partner rather than an infallible expert. Willison suggests setting reasonable expectations, understanding LLMs’ limitations like training data cut-off dates, and managing context effectively. He advocates for using LLMs to augment existing expertise, especially in initial prototyping phases, and advises thorough testing of any code they produce. Acknowledging that LLMs can’t replace human intuition and experience, he also highlights a valuable use-case in helping to answer questions about codebases.
-
Data Science Agent in Colab: The Future of Data Analysis With Gemini: Jane Fine’s article, “Data Science Agent in Colab: The Future of Data Analysis With Gemini,” discusses the innovative Data Science Agent in Google Colab, powered by Gemini. This tool simplifies data analysis by automating setup tasks like importing libraries and writing boilerplate code. Users can describe their analysis goals, and the agent generates a fully functional Colab notebook. This approach accelerates workflows, offers modifiable solutions, facilitates collaboration, and allows users to focus on data insights.