OpenAI just landed $110 billion. Plus: a new sandbox for AI agents, and how to cut Claude's context output by 98%.
OpenAI announced a $110 billion funding round over the weekend, securing backing from Amazon, Nvidia, and SoftBank. This signals continued investor belief in large language models. Builders also gained a new general-purpose sandbox platform for AI applications and an orchestration tool for Claude agents. One team shared a method that cut Claude Code's context window output by 98%.
ACRON, an open-source tool, reduces LLM inference costs by 85-97% using intelligent routing, multi-tier caching, and workflow decomposition. It provides a REST API, an OpenAI-compatible endpoint, and a dashboard for metrics and cache management. Builders can use smart routing based on task, quality, or latency preferences.
This article argues that Congress, not the Pentagon or AI firms like Anthropic, should define military AI usage policy. The author criticizes the Pentagon's threat against Anthropic for restricting AI use in mass surveillance or autonomous weapons. The current ad hoc process lacks democratic input and leaves crucial policy to executives.
Perplexity launched 'Computer,' an AI agent designed as a general digital worker. It executes complex, multi-step digital work autonomously, browsing the web, researching, and connecting with tools like Gmail. The system uses sub-agents to process natural language prompts, automating background tasks.
OpenAI secured a $110 billion funding round, with major investments from Amazon ($50B), Nvidia ($30B), and SoftBank ($30B). This pushes its pre-money valuation to $730 billion and includes a new multiyear cloud partnership with Amazon. OpenAI continues heavy investment in GPUs and compute, signaling ongoing competition.
One team's approach: "Context Mode" reduces AI agent context window consumption in Claude Code by 98%. It processes raw tool outputs in an isolated sandbox, compressing data from hundreds of kilobytes to a few, extending session time from 30 minutes to 3 hours. This saves 99% of context after 45 minutes by capturing only stdout and efficient indexing.
Researchers reverse-engineered Apple's M4 Neural Engine, gaining direct access beyond CoreML. They found CoreML adds significant overhead to Apple's "38 TOPS" claim and revealed the ANE as a graph execution engine, not a GPU/CPU. This direct access bypasses CoreML overhead, revealing the ANE's true performance and enabling custom model training on Apple's fixed-function accelerator.
This analysis introduces "cognitive debt," where AI-assisted development produces code faster than engineers can comprehend it. This mental deficit, distinct from technical debt, causes uncertainty in code reviews and incident response. The author argues that current engineering metrics miss this gap, sacrificing long-term maintainability and hindering senior engineer development.
Daniel Miessler outlines "The Great Transition," a framework for multiple AI-driven changes affecting builders. He argues knowledge moves public, products become APIs, and AI agents replace consumers. Automation changes from assisting to replacing humans, with AI managing "ideal states" to reshape work.
Alibaba released OpenSandbox, an open-source platform for running AI agents in isolated environments. It offers multi-language SDKs, unified APIs, and Docker/Kubernetes for agent evaluation, code execution, and RL training.
ruflo is an open-source orchestration platform for Claude. It deploys multi-agent swarms, coordinates autonomous workflows, and includes RAG integration with native Claude Code and Codex support.