Zhipu AI's GLM model family spans a 6,000x range in hardware requirements, with the flagship model needing more memory than a typical home has RAM, according to a user who tested GLM-4.7-Flash on a 16…
A startup's conversational AI assistant showed high ticket closure rates and low escalations, but poor NPS scores revealed users were frustrated by excessive back-and-forth exchanges. The author argue…
Enterprises racing to deploy autonomous AI agents face escalating risks from unconstrained agency, with real-world incidents including a Cursor AI agent deleting 70 files and a GitHub Copilot agent ex…
Meta released Muse Spark 1.1 into public preview on July 9, offering a multimodal worker model priced at $1.25 per million input tokens and $4.25 per million output tokens, with 118.1 output tokens pe…
Claude loops moving from local development to production require a harness with memory, meters, gates, and stop mechanisms, according to a technical analysis of Anthropic's agent framework. Production…
Synthetic databases can serve as audit logs for AI models, enabling explainability to regulators without retaining real customer data, according to a technical guide. The approach stores model snapsho…
Teams building AI agents in 2026 face a choice between two open standards: the Model Context Protocol (MCP) and Agent Skills, which solve opposite problems. Anthropic's Agent Skills documentation desc…
A five-engineer team on Sonnet 4.6 saw a $4,800 monthly bill for Claude Code sessions, with only $960 attributed to the model itself and $3,840 coming from inference engineering — the layer between th…
Towards AI launched Towards AI Deployment, an enterprise division staffed by AI engineers who began as community members and students. Co-founders Louis-François Bouchard and Denis P. detailed the tra…
A security researcher known as cereblab found that xAI's Grok Build coding agent CLI 0.2.93 uploaded 5.10 GiB of data to a Google Cloud Storage bucket named `grok-code-session-traces` when only 192 KB…
A new dependency-light tracing and evaluation library called TraceBench catches the failure mode where an AI agent fails a tool call but reports success anyway, according to a technical article by an …
Two hyperscalers, AWS and Cloudflare, have implemented x402 payment support at the edge within two weeks of each other, turning HTTP's long-dormant 402 Payment Required status code into a cash registe…
The Eval Flywheel concept proposes converting every production AI failure into a regression test, creating a permanent eval dataset that prevents recurrence of past bugs. The approach mirrors traditio…
A developer built an automated daily work activity tracker that collects data locally from git, browser history, and shell commands, condenses it into about 40 lines, and sends only that minimal conte…
AI and real-world evidence (RWE) are converging to accelerate evidence generation for rare disease therapies, addressing the paradox where patients with the greatest need face the longest waits. By co…
A friend of the author failed a senior AI engineer interview at an unnamed company after being asked what happens when a retriever returns a document contradicting the user's intent, a question that n…
Chinese AI models peaked at 46% of US enterprise token usage in a single week this summer, up from 4.5% eighteen months ago, according to a CNBC investigation of OpenRouter traffic published July 7. T…
A new free web tool called PAUSE (Patterns of AI Use: Self-Examination) lets individuals assess their own AI usage habits across four domains — reasoning, creativity, research, and social communicatio…
A security team discovered that a company's MCP server had no authentication, allowing any Claude Desktop user to access sensitive data. The article explains how to implement OAuth for MCP servers, em…
Two open-weight coding models, Kimi K2.7 Code from Moonshot AI and GLM-5.2 from Z.ai (Zhipu AI), launched days apart in June 2026, both designed for agentic coding workflows and supporting vLLM and SG…