Commentary on external content.
This article demonstrates how Jev, a typed decision-making model from TypeSafe, can evaluate agent responses by turning natural language judgment into structured classification with confidence scores. Rather than parsing generated prose, Jev returns constrained answers with probability and confidence metadata that lets developers identify uncertain evaluations and set appropriate automation thresholds based on risk tolerance.
Meta's engineering team describes an AI agent system that captures organizational expertise as structured, auditable knowledge separate from reasoning procedures. The system combines domain-specific knowledge architecture with a self-improvement loop that implements expert feedback without model retraining, creating persistent institutional memory that scales specialist knowledge across organizations.
Cohere Labs' analysis of modern LLM training pipelines reveals that cultural diversity is systematically lost during post-training, even when models support multiple languages. The authors argue that building truly global AI requires moving beyond multilinguality to explicit cultural representation in training data itself, as inference-time alignment techniques assume cultural knowledge is already present in the model.
As AI systems automate routine customer service tasks, job titles remain unchanged but the actual work transforms dramatically. Support workers now handle exceptions and quality control of AI outputs; QA analysts evaluate both human and AI-generated interactions; knowledge managers become responsible for training data that directly shapes AI responses. This shift concentrates difficult cognitive and emotional labor among humans while creating an entirely new operational layer requiring constant monitoring, calibration, and governance, yet most organizations remain structurally unprepared for this transformation.
Maryam Miradi presents a seven-step architecture for building closed-loop evaluations of multimodal AI agents at scale, specifically addressing the core challenge that AI can improve image quality while altering truth. The framework emphasizes faithful reproduction of content (matching ingredients, portions, and identity) through human ground truth anchors, multimodal routing, cross-modal verification, QA gates with retry logic, production drift monitoring, and iterative tuning before deployment.
This empirical study examines whether people can reliably distinguish between stories written by humans versus generated by AI systems. The research investigates the cognitive and perceptual barriers that affect our ability to detect machine-generated text in narrative contexts, with implications for trust, transparency, and how we evaluate authenticity in written communication.
Bureau for Listening Manifest 3 presents listening as a multifaceted artistic, social, and philosophical practice that extends beyond hearing to become a mode of being and relating to the world. The manifesto frames listening as a radical form of attunement, resistance, and connection that can transform perception, community, and understanding across everyday, artistic, and institutional contexts. It positions listening proposals as generative invitations to reorient how people engage with sound, silence, resonance, and each other.
Storybook is a frontend workshop for building, testing, and documenting UI components in isolation. It enables teams to develop components without entanglement in business logic or application infrastructure, and serves as a single source of truth for UI patterns that can be shared across teams and consumed by AI agents for automated code generation.
DAIR.AI's weekly survey synthesizes four key AI research directions: self-improving agents that evolve their own scaffolds and weights, metacognition frameworks that unify confidence and self-regulation in LLMs, the structural conditions required for meaningful routing in multi-model systems, and empirical critiques of harness evolution showing that simple search baselines often outperform supposedly optimized prompts. The curated selection positions these papers as navigational tools for builders implementing agent systems in production.
Solo journaling TTRPGs are a growing sub-genre of tabletop RPGs that enable single players to tell stories through writing and random elements, without requiring a group. These games use mechanics like dice rolls, prompts, and resource management to guide narrative development, with examples like Thousand Year Old Vampire demonstrating how constraints like memory limitations create compelling storytelling dynamics.
The article argues that the most effective workflow for creating skill.md files (structured instruction files for AI agents) combines AI-generated drafts with human refinement. It contrasts purely automated generation, which lacks nuance and product-specific judgment, with fully manual creation, which is time-consuming. The recommended hybrid approach leverages AI for structure and speed while relying on human expertise to add the contextual knowledge that makes AI agents actually effective.
Interactional AI is a research programme that applies Conversation Analysis to improve conversational AI models and user interfaces. The core argument is that interactive interfaces are not naturally intuitive unless they are grounded in the foundations of human social interaction, requiring design teams to ground conversational UIs in conversation analytic research methods and social science approaches.
This article examines the friction that arises when users shift from exploratory conversation with AI tools to precise task delegation. Through direct observation of seven people working with AI, the author identifies three interaction styles (collaborative, commanding, over-explainers) and reveals how conversational fluency in AI interfaces creates false expectations of shared understanding, leading users to accept suggestions that drift from their original intent.
This piece examines the iterative workflow patterns users develop when interacting with AI tools to achieve desired outcomes. It questions whether current AI-assisted workflows are effective or if they devolve into endless cycles of back-and-forth refinement, and points to Figma's exploration of AI-enabled product development as a case study in optimizing idea-to-product processes.
Eve Fairbanks argues that AI-generated writing betrays itself through a distinctive, "frictionless" quality that emerges from the absence of human struggle and revision. Rather than being merely a neutral tool like spell-check, AI writing bypasses the cognitive work through which humans develop meaningful ideas. The efficiency that makes AI appealing to writers creates a crisis of trust for readers who sense something crucial is missing.
Taguette is an open-source web-based tool for qualitative data analysis that allows researchers to upload documents, create hierarchical tag systems, and annotate text with tags and notes for later recall and organization. The tool addresses the practical need for structured, systematic approaches to analyzing and organizing qualitative research data.
The human brain is inherently lazy and tends to take cognitive shortcuts. This cognitive tendency affects how we interact with systems and interfaces, suggesting that interface design should account for human cognitive limitations and our tendency to conserve mental effort rather than fighting against these natural inclinations.
This LinkedIn post highlights UC Berkeley Law School's AI policy that restricts student use of generative AI for core legal reasoning tasks like brainstorming, drafting, revision, and exams. The post curates and endorses this policy as a model for responsible AI governance in higher education, emphasizing that cognitive skill development remains essential for quality legal practice and education.
This webpage documents Doug Engelbart's legendary 1968 demonstration at the Fall Joint Computer Conference, where he and his team showcased revolutionary interactive computing technologies including the mouse, video conferencing, and collaborative editing. The page serves as a portal to archived footage, retrospectives, and historical materials commemorating what became known as "The Mother of All Demos," emphasizing both the technical innovations presented and the ambitious vision of augmenting human intellect that motivated the work.
Faculty at three universities are using Hypothesis social annotation alongside generative AI tools to promote critical thinking and ethical technology use rather than banning or ignoring AI. By structuring assignments that require students to evaluate, critique, and analyze AI-generated content, these educators help students engage deeply with course material while developing essential skills for thoughtful AI use.
Ben Shneiderman presents eight foundational principles for interface design that have shaped HCI practice for three decades. The rules address consistency, universal usability, informative feedback, closure, error prevention, reversibility, user control, and cognitive load. These principles, derived from experience and continuously refined, provide a practical framework applicable across desktop, web, and mobile environments.
Adi Leviim argues that the chat box interface became the default for AI products not through thoughtful design but through shipping convenience: it was the fastest UI to build around LLMs' text-generation capabilities in 2022. The essay contends that this default choice sacrifices decades of established interface design principles, maximizing both the gulf of execution and gulf of evaluation, forcing users into iterative refinement loops when they should have direct manipulation and structured information surfaces.
Amelia Wattenberger argues that chat interfaces are poorly designed for most use cases because they lack affordances, force users to articulate context that should be built into the interface, and isolate responses without enabling iterative refinement. She advocates for UI patterns that expose parameters and constraints upfront rather than burying them in natural language prompts.
This article presents a comprehensive architectural framework for agentic AI systems in 2026, arguing that most AI failures stem from architectural problems rather than model quality. The author defines four canonical design patterns (Reflection, Tool Use, Planning, and Multi-Agent) and emphasizes that agentic AI represents a paradigm shift from monolithic systems to distributed, observable, and bounded agent architectures.
This article argues that 2026 marks a fundamental shift from generative AI (passive, read-only text generation) to agentic AI (active, autonomous systems that plan and execute work). The author contends that AI engineers must move beyond prompt engineering and embrace systems engineering, treating AI as operational infrastructure that performs end-to-end tasks rather than conversational oracles that simply generate responses.
This article argues that AI value comes from redesigning workflows rather than inserting AI tools into existing organizational structures. The author contends that traditional org charts show hierarchy but obscure how work actually flows, and proposes work charts as a management tool that makes workflows, decisions, handoffs, and accountability visible for effective AI transformation.
This article argues that Retrieval Augmented Generation (RAG) has become a strategic imperative for enterprises in 2026, addressing critical challenges like LLM hallucinations, outdated outputs, and high retraining costs. RAG bridges the gap between large language models and organizational knowledge by retrieving verified, real-time data at the moment of generation, ensuring outputs are accurate, compliant, and trustworthy without requiring constant model retraining.
A DeepLearning.AI short course taught by Atai Barkai (CopilotKit co-founder) covering three patterns for letting AI agents render dynamic interfaces instead of text-only replies: controlled components, the declarative A2UI framework, and open-ended UI via MCP. The capstone builds a LangChain plus React plus CopilotKit agent that produces charts and interactive widgets.
This is a documentation hub for GitHub Copilot adoption and best practices at NAV (the Norwegian Labour and Welfare Administration). The site provides news updates, tools, guidelines, and usage statistics for AI-driven development, focusing on custom agents, skills, and organizational implementation of GitHub Copilot across teams.
Anthony Masure traces UCD from Xerox Star through Don Norman, arguing that design cannot be centered on anything. A world built on invisible technology is a world without experience.
Tommi's personal site, framed as 'a perpetual work in progress' โ a digital garden built around identity rather than credentials, with explicit anti-surveillance and indie-web commitments. Links community workshops, decentralised internet practice, and a Leibniz-style acceptance of incompleteness as core to what the site is.
Argues that perception isn't a window onto reality but the substrate of reality itself. Sheffield draws on phenomenology, neuroscience, and animal cognition to introduce 'somatic deixis' (a two-step process of designation and adjudication) and 'execution states' as alternative frames for what minds are. Opening line: minds do not create experience, experience creates minds.
Fieldwork for Future Ecologies (Onomatopee 225, 2022) argues art practice and art-based research can radically expand the concept of "fieldwork" beyond its conventional scientific framing, positioning creative and speculative methodologies as essential tools for environmental and climate inquiry.
Treats prompts as first-class delivery artefacts: version-controlled, reviewed, and refined alongside code. Introduces the REASONS Canvas โ a seven-part template covering requirements, entities, approach, structure, operations, norms, and safeguards โ and a six-step workflow that fixes the prompt before fixing the code when reality diverges.
Vincent Rump introduces the Tri-System Theory (Shaw & Nave, 2026), which extends Kahneman's dual-process model by proposing a third cognitive system โ artificial cognition (AI) โ that actively participates in human reasoning and decision-making. The post argues that humans increasingly defer to AI outputs with minimal critical reflection, a phenomenon the authors call "cognitive surrender," which improves decision quality when AI is correct but degrades it when AI is wrong โ while trust remains consistently high either way. This raises urgent questions about human autonomy, expertise, and responsibility as cognition becomes a hybrid, human-AI process.
Microsoft researchers introduce DELEGATE-52, a benchmark of long document-editing workflows across 52 professional domains. Even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT-5.4) corrupt about 25% of document content by the end of these workflows, and agentic tool use does not improve performance. The paper argues current LLMs are unreliable delegates: they introduce sparse but severe errors that compound silently over long interactions.
Frank A. Kalman, a formally trained journalist, argues that using AI as a collaborative drafting partner is not a betrayal of craft but an honest extension of a writer's editorial process. He frames AI as a responsive thinking partner whose "wrongness" is productive, pushing the writer to clarify what they actually mean. The piece is a transparent, process-level account of how an experienced writer uses voice memos, AI chat, and iterative editing to reach a finished essay, positioned against the performative outrage of writers who conflate the suffering of drafting with the value of writing itself.
Tactical Tech is a Berlin-based international non-profit that develops creative interventions, toolkits, and training programs to help individuals, communities, and educators critically understand the socio-political and environmental impacts of digital technologies. Its flagship announcement, "Supercharged Human?", targets teens and educators with ready-to-use AI literacy resources designed to foster critical engagement rather than passive adoption. The organisation frames its work around building civil society capacity to resist misinformation, digital extractivism, and uncritical AI use.
Argues that the mental model you bring to AI determines how effectively you work with it. Catalogues thirteen metaphors (bionic mind, smart intern, word calculator, genie, sparring partner, and others), each surfacing different affordances and risks of LLMs. The framing is operational: pick the metaphor that fits the task, switch deliberately, avoid being trapped by one.
China's proposed AI law acknowledges AI-related human vulnerabilities and establishes contextual technical measures to prevent AI harms | Edition #264
A peer-reviewed conference paper from DRS 2022 that examines how metaphors shape designers' understanding of machine learning and AI systems, exploring both where common metaphors mislead and what qualities make metaphors genuinely useful for design thinking.
Nora Bateson argues that applying static, context-free data to living systems is dangerous. Her concept of warm data contextualises information through relationships and interconnection, as an alternative to reductionist measurement. Spans ecology, healthcare, food systems, and education.
Marek Tuszynski (Tactical Tech) argues that the synthetic intimacy people develop with generative AI tools (an 'ELIZA effect') makes critical conversation about AI almost impossible: dependency forecloses critique. The piece curates 30 resources mapping AI's labor exploitation, environmental costs, military use, and concentration of power, framing AI critique as a moral and political project rather than a technical one.
A short Instagram reel by Lucio Arese visualises the song of a hemp bunting (Linaria cannabina) as a scatter plot, where pitch, amplitude, and timbre map directly to spatial position and density. The post argues that birdsong is not creativity but a tightly structured signal carrying territorial, genetic, and hormonal information: a 'biological blockchain' rather than a song.
AI has collapsed centuries of human knowledge into a single point of failure. Discover Algorithmic Epistemic Injustice and why statistical probability isn't enough.
6 plugins that turn Claude Code into a full dev team in 5 minutes. Each one covers a different role; planning, design, code review, security, memory, and team coordination.
Ben Wigler highlights a paper on emotion circuits in LLMs: discovery and control.
AI tools can accelerate research, but using them to avoid the hard work of learning undermines the development of independent thinking.
Apple's hybrid batch-incremental knowledge graph platform. Key inspiration for building smarter connections in this garden.
A process post by Lucio Arese: he is recording many interactions between robins and using the material to build visual models of 'shared manifolds' between multiple individuals, alongside reading the literature on territorial interactions. The work continues his project of treating birdsong as data; the new dimension is multi-individual rather than single-bird.
A comprehensive overview of the digital garden concept and its evolution.
Weblinks