‘Introducing Gemma 3 270M: The Compact Model for Hyper-Efficient AI’: Gemma 3 270M is a compact AI model designed for hyper-efficiency, emphasizing fine-tuning to specialize in tasks such as text classification and data extraction. While it’s not suitable for general-purpose large language model tasks due to its size, its efficiency and capability make it ideal for building production systems that are lean, fast, and cost-effective. The model embodies the philosophy of using the right tool for the job, prioritizing quality and specialization over raw power.
‘Piloting Claude for Chrome’: Simon Willison discusses the security risks associated with using Claude for Chrome, particularly focusing on prompt injection attacks where AIs can be tricked into performing harmful actions. Red-teaming experiments have shown vulnerabilities without proper mitigations. Anthropic advises against autonomous mode, instead opting for configurations requiring user intervention, like site-level permissions and action confirmations. Despite the significant demand for browser automation, Willison is skeptical about end users effectively managing these security risks.
‘Too Many Model Context Protocol Servers and LLM Allocations on the Dance Floor’: The blog post by Simon Willison discusses the issue of extensive token costs associated with Model Context Protocol (MCP) servers and LLM allocations, referencing Geoffrey Huntley’s observations. Huntley estimates a usable context window of around 176,000 tokens after accounting for system prompts. The post highlights that overloading prompts with irrelevant information can degrade LLM performance and suggests a more token-efficient method: using existing command-line interface (CLI) tools, like GitHub’s CLI, which many LLMs can utilize effectively with minimal token expenditure. The author shares positive experiences with building custom CLI tools for specific LLMs, emphasizing that these tools can access functionality at negligible token costs.
‘Amazon”s New Alexa AI Sounds Like a Dystopian Nightmare’: Amazon’s new Alexa AI, dubbed “Alexa+,” has sparked controversy with its fusion of data surveillance and generative AI capabilities. Despite boasting enhanced features like improved voice synthesis and smoother app interactions, Alexa+ has suffered backlash for disrupting basic functions and bombarding users with ads, positioning itself more as a marketing tool than a personal assistant. Limited information is available on its AI technology, suspected to be proprietary to Amazon, but privacy concerns heightened after reports of indefinite voice data storage surfaced. Critics view the rollout as a profit-driven move rather than enhancing user experience.
‘Claude Sonnet 4 Now Supports 1M Tokens of Context’: Claude Sonnet 4, by Anthropic, now supports 1 million tokens of context, increasing from the previous 200,000 limit and aligning with similar offerings from Gemini and OpenAI. This enhancement enables more extensive input and output capabilities. Additionally, Anthropic introduces variable pricing based on context length, with lower rates for prompts up to 200K tokens and higher fees for those exceeding this limit, a strategy similar to Gemini’s existing pricing model.
‘Google Gemini URL Context’: Simon Willison’s weblog discusses a new feature in the Google Gemini API called “url_context,” which allows models to request the contents of URLs to enhance responses to prompts. Accompanying this release is version 0.25 of the llm-gemini tool, which supports this feature through a new option. This update allows users to enable URL content requests, potentially enriching the interaction capabilities of the models using the Gemini API.
‘Ted Lechterman: “Democratizar La IA No Es Hacerla Accesible, Es Compartir El Poder”’: Ted Lechterman, a prominent philosopher and leader of UNESCO’s first Chair in AI Ethics and Governance, argues that democratizing AI involves sharing power, not just making it accessible. He highlights the need to address the opaqueness of AI’s “black boxes” and the challenges posed by private corporations controlling powerful AI technologies. Lechterman emphasizes that true democratization means allowing citizens to influence AI’s design and governance. He also notes the importance of proactive ethical policies beyond legal compliance to ensure AI supports democracy and the public good.
‘Enseñando Inteligencia Artificial a Diseñadores UX’: Nicolás Bronzina’s article discusses teaching artificial intelligence (AI) to UX designers, emphasizing the shift from deterministic design software to probabilistic AI models. The article highlights how AI changes the approach to digital design by introducing statistical elements, necessitating new methodologies to manage imperfection and uncertainty. Bronzina explains using AI as a creative collaborator and presents a detailed educational framework with lectures by Bronzina, Fabien Girardin, and Rohit Gupta at a Madrid university, stressing AI’s role in expanding creative possibilities beyond traditional methods.
‘GPT-5 Es Mejor De Lo Que Crees’: The article “GPT-5 Es Mejor De Lo Que Crees” by Miguel A. Román discusses the release of GPT-5, highlighting both its modest advancement compared to expectations and significant improvement over GPT-4. While initial reactions to GPT-5 were lukewarm, its incremental improvements have raised AI capabilities to remarkable levels. The landscape has shifted from developing AI models to building value-added products atop these technologies. Providers will compete based more on service and cost than raw model intelligence. Furthermore, the focus is moving toward creating AI-driven applications in the emergent Software 3.0 era, emphasizing productivity and innovation. This new phase integrates AI into everyday business practices, demanding both engineering and creative efforts to harness its potential fully, challenging developers to leverage AI for unprecedented productivity.
‘Best Practices for Pair Programming With AI Assistants’: The article from Graphite.dev outlines best practices for pair programming with AI assistants, illustrating how this introduces unique challenges and benefits compared to traditional human pair programming. AI-assisted programming offers constant availability and broad technical knowledge but lacks domain-specific insights and requires text-based communication. Effective collaboration involves clear role definition, providing contextual project information, iterative refinement of code, critical review for errors and vulnerabilities, and using detailed prompting strategies to improve code quality. Developers can leverage AI to learn new concepts, frameworks, and languages, enhancing their skills through AI-driven insights.
‘AGENTS.md’: AGENTS.md is a concept designed to provide detailed context for coding agents, separate from the human-focused content in README.md files. While README.md files are concise and aimed at human contributors with project intros and guidelines, AGENTS.md offers more in-depth instructions like build steps, testing, code style, and security considerations specifically for coding agents. Projects can include AGENTS.md files at their root or in subdirectories for monorepos, allowing agents to access the most relevant instructions automatically.
‘Quoting Bruce Schneier’: In a blog post titled “Quoting Bruce Schneier,” Simon Willison discusses the vulnerability of AI systems to prompt injection attacks. He highlights Bruce Schneier’s view that there are currently no AI systems secure against such attacks, especially in adversarial environments with untrusted data. This issue is seen as existential, yet it appears that most technology developers are ignoring the problem.
‘Quoting Sam Altman’: In a blog post titled “Quoting Sam Altman,” Simon Willison summarizes a remark from Sam Altman during a dinner with a small group of reporters in San Francisco. Altman discusses the profitability of their company’s inference operations, stating that they would be significantly profitable if training costs were excluded. The post highlights Altman’s comments as noteworthy from the event, focusing on the financial aspects of their business model.
Real estate
‘Boost conversions by +140% with AI-powered property search’: SearchSmartly provides AI-powered property search solutions that significantly boost conversion rates by 140 percent. Their customizable APIs and whitelabel products offer personalized property searches, aligning with seekers’ specific preferences. The intelligent recommendation engine uses diverse datasets to provide hyper-personalized results, including factors like local schools and transportation. It uncovers “Hidden Gems”—overlooked properties in nearby areas, enhancing seeker satisfaction and advertiser visibility.
‘Coraly Raises 2M Pre-Seed to Transform Real Estate Lead Generation With AI’: Coraly.ai, formerly Coralytics, has raised a $2 million pre-seed funding round led by the Salica Oryx Fund, with participation from EQ2 Ventures and strategic angel investors. The Dubai-based proptech company aims to enhance its AI-driven platform for real estate lead generation. The funds will accelerate product development, including expanding Coraly.ai’s engineering team and market footprint in the UAE, Saudi Arabia, France, and the United States. Coraly.ai, rebranded for a global vision, focuses on solving the industry’s fragmented tool challenges by streamlining real estate processes and facilitating growth and innovation for agents worldwide. Strategic partnerships and the backing of investors highlight its potential to transform real estate technology and expand into key growth markets.
‘Kasa Raises 40M to Invest in Technology and AI’: Kasa, a property management company specializing in boutique hotels and apartment-style lodgings, has raised 40 million to enhance its technology and AI capabilities for its hospitality operations system. With backing from Silver Lake Waterman and partnerships with Starwood Capital, Berkshire, Greystar, and Prudential, Kasa plans to improve performance and profitability for its partners. The company operates over 85 properties across 45 U.S. markets, achieving more than 100 million in annual booking revenue. Kasa aims to provide independent hoteliers with access to advanced technology and operational systems typically reserved for larger hotel chains.
‘Unite Students Acquires Empiric Student Property PLC in £723M Deal’: Unite Students has acquired Empiric Student Property PLC for 723 million pounds following negotiations from May, marking a major consolidation in the UK student accommodation sector. Empiric shareholders will receive new Unite shares and cash, and own 10% of the new entity. Empiric’s assets, located in top UK university cities, were praised for their strategic value. Despite a year-on-year revenue growth of 3.3%, Empiric’s pretax profit declined. The deal is projected to enhance long-term growth for stakeholders.
‘Domain Shareholders Back 3 Billion CoStar Group Takeover’: Domain Holdings Australia’s shareholders have overwhelmingly approved a 3billiontakeoverbyUSpropertydatagiantCoStarGroup,with99.984.43 per share, a 50% premium over the prior month’s average. The deal, pending court approval and regulatory conditions, enhances CoStar’s presence in the Asia-Pacific region. Completion is expected by August, reshaping Australia’s proptech landscape.
Economics
‘No Es El EBITDA, ¡es El ROIC! El Caso De La Creación De Valor en El Sector De La Consultoría Tecnológica’: In the realm of technological consultancy, evaluating value creation transcends traditional metrics like EBITDA, focusing instead on ROIC (return on invested capital) as a crucial indicator. Effective financial value creation hinges on how firms utilize and multiply cash. For instance, two companies may yield different returns on the same investment, indicating varied efficiency and value generation. Renowned experts Tim Koller and Aswath Damodaran assert that ROIC is essential for accurately measuring both current and anticipated financial value in companies.
‘Beneath the AI Bubble, the Economy Looks Bleak’: Despite the booming appearance of the US tech sector and record highs in stock markets, such as the S&P 500 and the Nasdaq Composite, the underlying economy remains bleak. The AI-driven gains create a misleading veneer of prosperity, while widespread economic stagnation looms beneath. A handful of tech giants—Nvidia, Amazon, Google, Tesla, Microsoft, Apple, and Meta—dominate this apparent growth, masking profit declines in other sectors like energy and consumer goods. Stripping away AI’s impact reveals economic stagnation, with non-tech companies underperforming. AI has effectively bailed out the stock market and obscured deeper economic issues.
Management
‘Ramblings’: “Ramblings” by Steph Ango suggests creating personal “ramblings” channels for remote teams of 2-10 people to foster social cohesion without cluttering main group channels. These are likened to personal journals within the team chat app, allowing members to share thoughts, ideas, and updates casually. Typically, team members post short updates on various topics a few times a week. The channels, muted by default, encourage deep work while maintaining connection. This approach at Obsidian replaces water cooler talk, promoting creativity and connection.