1M-token Sonnet 4.6, Claude Code to Figma, and an evals flywheel that starts with 20 traces.
AI writing goes bland when models sand off high-entropy detail (semantic ablation). Today: Sonnet 4.6 ships 1M context, Composio gives agents 100+ tools, and a pragmatic evals loop that starts from your failures.
Anthropic dropped Claude Sonnet 4.6, now in beta. It brings a 1 million token context window, major coding boosts, and preference rates rivaling older Opus models. It is now default on Free and Pro tiers at the same pricing, with new GA tools.
Read full story→Apple is reportedly accelerating development of several new AI-focused hardware products, including smart glasses, a pendant-style device, and AirPods with integrated cameras. These devices are designed to leverage AI, signaling a shift towards more ambient and personalized AI interactions.
Figma now lets you import functional UIs captured in the browser via Claude Code into editable Figma frames. This allows teams to edit frames, compare variants, and roundtrip back to code via MCP without touching code directly.
Meta is expanding its AI infrastructure by deploying millions of NVIDIA's upcoming Blackwell and Rubin GPUs, alongside NVIDIA CPUs and Spectrum-X Ethernet switches. This multiyear partnership aims to optimize Meta's data centers for both training and inference, integrating NVIDIA's full platform. Networking and CPUs are now first-class in inference clusters.
Companies like Google and Osmo are using ML to profile scents and search for novel molecules. This digital approach to fragrance design could reduce reliance on resource-intensive natural ingredients, and find synthesized molecules capable of evoking specific brain patterns. Builders could use this for dataset creation or representation learning.
Start with 20 to 50 real traces, label failure modes, write critiques, and iteratively build evaluators, including LLM judges. This flywheel guides dataset creation for more targeted improvements.
Do the task with iteration first, then distil the final successful trajectory into a reusable skill prompt. Generating skills before completion can codify initial mistakes, while generating after success captures emergent knowledge.
AI models systematically strip out high-entropy information, leading to generic output. This 'semantic ablation,' partly driven by decoding and alignment pressures, flattens language and structure, creating a 'JPEG of thought' that lacks original data density. Force high-entropy inputs: specific constraints, concrete examples, and keep your own voice fingerprints.
AI coding agents may be creating 'token anxiety,' a phenomenon where constant prompting and babysitting AI code resembles gambling. This, coupled with management pressure for productivity, could push workers into an addictive cycle, blurring human effort and machine output. Expect work-life balance and ethics to be tested as this trend normalizes.
Tunguz says AI inference expenses ballooned from $200/month to over $100k annually. Migrating to an open-source model saved 88%. This shift makes inference costs a major compensation factor. Builders can add token budgets, per-route model selection, caching, and eval-based regression gates.
ComposioHQ's new project equips AI agents and LLMs with over 100 integrations. It uses function calling to let AI models interact with various apps and services, providing a tool catalog, OAuth handling, API execution, and consistent schemas to expand their utility.
This GitHub repo offers a lightweight system for meta-prompting, context engineering, and spec-driven development. It's an opinionated wrapper that packages context engineering, subagents, and state so you run a few commands and verify outputs.