New in LangSmith: a Fleet on-call copilot for alert triage, computer use for agents, voice trace debugging, and experiment status tracking. Plus Deep Agents Rubrics, programmatic subagents, a new LangSmith Deployment course, and upcoming events in Chicago, Berlin, DC, and Vegas.
Fleet On-Call Copilot: a prebuilt agent template that triages alerts and drafts updates using code, traces, and runbooks.
Computer Use: agents can now operate an isolated virtual computer for code, files, and authenticated API calls.
This article critiques the 'Ethical AI' movement, spearheaded by Anthropic, arguing that it rests on the unproven assumption that AI development is unstoppable but steerable. In reality, Ethical AI neither renews epistemic habits nor guides AI toward humane ends, but functions as controlled opposition for unethical AI. Through analyzing Jack Clark's world-building narrative, the author exposes the contradictions: claiming helplessness to slow AI while asserting power to control its consequences.
Ethical AI is built on the false premise that progress is inevitable but can be guided.
Anthropic's narrative casts AI development as a sci-fi world-building project, asserting AGI is inevitable but shapeable.
Artificial intelligence is reshaping retail behind the scenes, influencing search rankings, supply chains, and customer engagement. Macy's adopts an 'AI-first' philosophy, embedding intelligence into personalization, operations, and software development. The company's AI shopping assistant, Ask Macy's, exemplifies conversational commerce. AI is seen as an invisible layer augmenting human judgment, not replacing it.
AI's impact in retail is primarily in back-end processes like product search, inventory management, and code deployment.
Macy's 'AI-first' approach integrates intelligence directly into systems to speed decisions and enhance relevance.
Prem Natarajan left Amazon to become Capital One's Chief Scientist, applying deep AI research to solve real-world financial challenges at scale, from fraud detection to agentic customer service.
Capital One treats AI as a scientific discipline, not just technology to deploy.
The bank's cloud-first infrastructure enables large-scale AI research.
DeepReinforce released Ornith-1.0, an open-source coding model family built on Gemma 4 and Qwen 3.5. Instead of a fixed harness, the model learns its own scaffold during reinforcement learning. The 397B flagship reports 82.4 on SWE-Bench Verified, with all weights under the MIT license.
Ornith-1.0 ships in 9B, 31B, 35B-MoE, and 397B-MoE sizes under MIT, built on Gemma 4 and Qwen 3.5.
The model learns its own scaffold during RL, jointly optimizing the harness and the solution.
Meta reverses stance on forcing engineers into AI training task force, now allowing voluntary participation after employee backlash and morale concerns.
Meta initially reassigned 7,000 employees to Applied AI task force
Company now defers to individual choice after backlash
The generative AI economy generated $110 billion in sales over the past 12 months, with an annualized run rate exceeding $175 billion. This bottom-up, deduplicated measure reveals that AI revenues are growing three times faster than mobile or internet waves, and can just cover hyperscaler depreciation expenses. Lower token prices drive 12-18% more usage per 10% price cut, sustaining total spending growth.
AI ecosystem generated $110B revenue (deduplicated) in past 12 months, annualizing to $175B run rate, growing 3x faster than previous tech waves.
The study uses a bottom-up method counting only end-customer spend to avoid double-counting.
This post shows you how to configure training jobs on Amazon SageMaker AI to get the most out of Blackwell’s architecture on AWS. You learn how to select batch sizes and sequence lengths that take advantage of Blackwell’s expanded memory, choose the right precision format for your model size (1B to 64B parameters), and apply activation checkpointing strategically. By the end, you have a practical framework for tuning your training configuration and launching distributed training jobs on P6-B200 instances.
Blackwell's expanded memory supports larger batch sizes, longer sequence lengths, and simplified model sharding.
Activation checkpointing is a prerequisite for stable training with large models (~14B+ parameters).
This article explores the context drift problem in long-running AI systems, distinguishing between LLM hallucination and architectural hallucination. It argues that current approaches like context windows and RAG fail to preserve memory integrity, and proposes structured memory with extraction guardrails as a solution.
AI memory degrades over time due to context compression, not model limitations.
Architectural hallucination arises from self-feedback loops causing context drift.
This post demonstrates how to implement video upscaling using SeedVR2 on SageMaker AI. It covers solution architecture, deployment steps, and performance comparisons highlighting quality improvements and processing efficiency. By the end, you'll have practical knowledge to implement this super resolution solution.
SeedVR2 is an open-source video restoration model by ByteDance, combining diffusion models and GANs for efficient upscaling.
The solution uses a three-tier AWS architecture including security, storage, and SageMaker processing pipeline.
In this post, we show you how to build Chaplin (Customer Health and Planned Lifecycle Intelligence Nexus), an open source solution that uses AI agents exposed through the Model Context Protocol (MCP) to provide self-service health event analytics.
Chaplin is an open-source solution using AI agents via MCP for self-service AWS Health event analytics.
It overcomes the bottleneck of relying on TAMs for health event interpretation.
This post shows how to build a governed, serverless data mesh on AWS that provides the secure, scalable data foundation production agentic AI requires.
Agentic AI requires fine-grained access control at every step from tool discovery to query execution, which traditional RAG governance cannot address.
Amazon S3 Tables with built-in Iceberg support and AWS Lake Formation provide row/column/cell-level security with up to 10x higher transactions per second.
OpenKnowledge is an open-source, AI-native markdown editor and knowledge base that serves as an alternative to Obsidian or Notion. It offers a beautiful rich text editor with markdown under the hood, designed for both humans and AI agents. Features include collaborative editing, git-backed sync, agent-native skills, MCP integration for Claude, Cursor, and Codex, and a local-first, privacy-focused approach. The launch of v2.0 saw significant traction with 1.4k signups in 24 hours and hitting #1 on Product Hunt and front page of Hacker News.
OpenKnowledge is an open-source, AI-native markdown editor and knowledge base.
It features a rich text editor with markdown foundation, designed for human and AI agent collaboration.
Tabularis is an open-source desktop database client designed for both AI agents and human users. It features a built-in MCP server for AI agents to safely inspect schemas and run queries, while retaining powerful human-friendly tools like Monaco SQL editor, notebooks, visual query builder, and more. Supports PostgreSQL, MySQL, SQLite, and extensible via plugins. Local-first architecture ensures data security.
Built-in Model Context Protocol (MCP) server enables AI agents to execute queries within the app
Offers professional tools: Monaco SQL editor, notebooks, visual EXPLAIN, ER diagrams
This tutorial covers three methods to create Google Sheets using Gemini: within the spreadsheet via built-in integration, through the Gemini web app with export, and by generating Google Apps Script for advanced automation. It also includes tips for better results.
Gemini is an AI integration in Google Sheets that lets you create, populate, and analyze spreadsheets using natural language.
Method 1: Use Gemini's side panel inside Google Sheets to generate tables, formulas, and analysis via prompts.
Traditional pre-merge human code review cannot scale with AI-generated code. The shift is toward automated CI/CD gates that enforce policies consistently, with human review reserved for high-risk changes. A four-layer quality gate pipeline and post-merge review create a verifiable, auditable system of controls.
AI-generated code volume overwhelms traditional human review capacity.
This article presents a method to automate software fork maintenance using AI coding agents, framing it as a closed-loop feedback system in control theory. Applied to Cohere's fork of vLLM, it reduces the time to absorb upstream releases from weeks to days. The approach includes automated rebasing, measurement collection, and iterative fixing, with a case study on the Cohere Transcribe model.
AI agents can automate the full fork maintenance cycle: sync, measure, fix, repeat.
Fork maintenance is modeled as a feedback control system, with the agent as the controller.
Ai2 compares its 7B transformer Olmo 3 and hybrid Olmo Hybrid, finding the hybrid excels on content words (nouns, verbs, adjectives) and tokens requiring context, but loses advantage on repeated tokens and closing brackets. Token-level loss filtering reveals architectural differences.
Hybrid models predict meaningful tokens (e.g., content words) better, but not repeated tokens.
Hybrids replace some attention layers with recurrent layers, which have fixed-size memory suited for tracking sequential state.
Researchers introduce generative causal testing, which translates black box models into clear hypotheses and verifies them in the scanner, revealing what specific brain regions respond to in language.
GCT distills brain-prediction models into short verbal explanations.
It validates explanations by generating new stories that causally activate targeted brain regions in fMRI.