Nvidia's Alpamayo pushes autonomy, Mac control arrives via Cloudflare, and a Wikipedia chunker does 120ms. Oh, and your reserved GPU capacity just got 15% pricier.
AWS updated EC2 Capacity Blocks for ML pricing (about +15%), a real tax on reserved GPU capacity. Nvidia used CES to formalize Rubin (rack-scale) and Alpamayo (autonomy), plus two builder tools: Mac control via Cloudflare + a chunker that does English Wikipedia in ~120ms.
US foreign policy signals interest in Venezuela and Greenland to shape long-lead supply of critical minerals like gallium and lithium, essential for AI chips and data centers. This aims to reduce reliance on China, which dominates rare earth refining, through mechanisms like sanctions, investment, offtake agreements, and export controls, with impacts expected over years rather than immediately.
Read full story→At CES 2026, Nvidia signaled a push into "physical AI" with its Cosmos world model and Alpamayo for autonomous driving, aiming for robotaxis by 2027. They unveiled Rubin, their next-gen AI platform for data centers, and are promoting "open" AI models across various domains, skipping new consumer GPUs.
OpenAI's 2026 strategy outlines primitives for agents: proactive scheduling, permissions, tool orchestration, memory, and enterprise connectors. This includes transforming ChatGPT into a personal assistant and building an agent operating system, alongside enhancing enterprise automation and evolving Codex into an automated teammate that integrates with existing developer tools.
Google Cloud's VP of Engineering for Databases discusses integrating generative AI with operational databases, addressing the disconnect between LLMs and secure enterprise data. Expect built-in vector indexes, in-database embedding generation, policy-aware retrieval, and first-class agent tool surfaces (MCP-style). He highlights initiatives like the MCP Toolbox, natural language to SQL, and the evolution towards 'AI-native databases'.
`memchunk` is a new text chunking library that processes English Wikipedia in ~120ms on a 7950X in release mode, outperforming other popular tools by orders of magnitude. It achieves this speed through low-level byte search optimizations like SIMD (AVX2/SSE2) and efficient backward delimiter searching, though tokenizer and delimiter choices affect results. Python and WASM bindings are available.
A new "agentic approach" proposes to solve flakiness and complexity in end-to-end web application testing. This method uses AI agents to observe UI state, attempt tasks, and produce a structured trace (steps, assertions, DOM snapshots) that can be replayed or converted into deterministic tests, improving testing efficiency and reliability.
An analysis critiques "Capital in the 22nd Century," arguing its predictions of extreme AI-driven wealth inequality rest on flawed assumptions: humans retain control, AIs do not own capital, institutions remain legible, and production functions resemble today. The author questions applying current economic models to a future with superintelligent AI, suggesting human survival and control are the real issues.
AWS raised prices for its EC2 Capacity Blocks for ML by about 15%. If sustained, this signals that constrained GPU capacity can reprice upward, potentially affecting other cloud resources.
Despite widespread predictions, "none of that ended up happening": 2025 didn't see AI agents significantly integrate into the workforce as anticipated. The article reports agents largely failed to meet expectations, highlighting their current limitations and expert skepticism from figures like Gary Marcus and Andrej Karpathy.
AI hasn't eliminated code review; it's shifted the focus to proving functionality. The article introduces a "PR Contract" framework: PRs must include proof (tests, steps, logs) and an explicit AI contribution note. This allows solo developers to use automation for rapid shipping, while teams still rely on human judgment for critical aspects like security and context, given AI's current weaknesses.
SYSTEM, an open-source project, lets you control your Mac remotely using natural language. It pairs a Claude-powered AI Agent on Cloudflare Workers with a local Bridge to execute commands via AppleScript, shell, and Raycast, secured by Cloudflare Access gating, least-privilege command allowlists, and local-only bridge execution boundaries.
`musistudio/claude-code-router` is a TypeScript proxy/router that integrates Claude's general APIs (including coding capabilities) into your custom infrastructure. It runs as a service, allowing developers to configure routing rules, authentication, logging, and retries, while abstracting model updates.