engblogs

MIT's Humane UXD course blends anthropology and computer science to teach humane, user-centered chatbot design using linguistic anthropology and real-world projects (Pond, News Nest, M^3) to steer AI toward social guidance rather than addictive engagement.

Modular Structured Mojo Kernels split a GPU matmul into TileIO, TilePipeline, and TileOp, delivering a durable, cross-vendor pattern centered on TileTensor that scales across NVIDIA and AMD without rewriting the kernel structure.

A technical guide to HIMSS26 highlighting how Databricks enables agentic AI at scale on a governed lakehouse for healthcare, powering real-time analytics, predictive insights, and AI-enabled clinical workflows.

An in-depth, developer-focused guide to Databricks Lakebase autoscaling, explaining how elastic Compute Units and scale-to-zero enable responsive, cost-aware databases for feature branches, CI/CD, and modern workloads.

Databricks merges Quotient AI to inject continuous evaluation and learning for AI agents in production, strengthening Genie, Genie Code and Agent Bricks with structured evaluation and reward signals to steadily improve accuracy, reliability and domain expertise.

Genie Code is Databricks’ agentic, autonomous AI for data teams that builds pipelines, debugs failures, ships dashboards, and maintains production systems, tightly integrated with Unity Catalog for governance and data context.

A technical deep dive into Salesforce's enterprise-grade agent platform, detailing Agentforce and Data 360, architecture choices for trust and governance, data and metadata context, cross-vendor interoperability, and observability to scale reliable agent reasoning beyond CRM.

Replacing bulky HTML error pages with RFC 9457-compliant Markdown and JSON payloads to give AI agents actionable guidance and cut error-response tokens by ~98% across 1xxx errors and beyond.

Cloudflare announces general availability of AI Security for Apps, adding free AI endpoint discovery, custom topics detection, and integrations with IBM and Wiz to detect, mitigate, and respond to LLM-driven threats in AI-powered applications.

Integrating Wiz into Google Cloud to deliver a unified, AI-powered security platform that spans multicloud and hybrid environments, protecting AI-enabled workloads with enhanced threat intelligence and proactive defense.

A data-driven analysis of the 2025-26 Duolingo Language Challenge reveals how politicians and academics use gamified language learning to drive cross-sector engagement, collaboration, and cultural understanding.

A two-step VLM-guided planning method that uses SimVLM to describe and simulate a scene and GenVLM to generate PDDL files, translating visual inputs into formal planning for robust, long-horizon tasks.