Chinese AI lab Moonshot has released Kimi K3, an open-weight model (one whose internal parameters are publicly downloadable) that reportedly rivals top-tier proprietary models, though not without tradeoffs.

DeepSeek has reportedly shipped a new AI system with image-understanding ('sees') capabilities, adding competitive pressure on larger, better-funded AI labs.

A head-to-head comparison covers two newly released models — Google's Gemini 3.7 Flash and OpenAI's GPT-5.6 Sol — as both companies continue rapidly iterating on their lineups.

Linkdaze has launched a digital calendar with AI features, including a meal-planning tool, and distinguishes itself by keeping those features free rather than locking them behind a paywall.

A deep dive argues that enterprise AI's common approach — building separate context and embeddings for each application — breaks down at scale, creating inconsistent knowledge across an organization instead of treating it as a shared asset.

Developer Drew Breunig reflects on how the arrival of Anthropic's expensive 'Fable' model — despite being excellent — forced teams to finally think seriously about optimizing coding workflows and model costs, rather than just waiting for a cheaper, better model to arrive.

A theoretical study finds that even a perfectly working AI could hurt research quality, because time saved gets redirected into starting new projects rather than polishing existing ones, degrading publication quality in two of three modeled scenarios.

In an experiment by Andon Labs, an AI agent running a store needed explicit human prompting before it would fire an underperforming employee, and testing across seven models showed more capable AIs were more consistent about enforcing termination rules, while nearly all models were uncritical about hiring.

Since early 2025, AI agents have overtaken humans as the top consumers of computing 'tokens' (units of text processed by AI) on OpenRouter, growing 14-fold versus humans' 2.8-fold increase, though cheap cached prompts are keeping actual cost growth much slower than raw usage numbers suggest.

An independently built, open-source tool can reportedly remove the watermark Anthropic recently added to Claude-generated content, raising questions about the durability of AI content-labeling efforts.

A new benchmark site, trending on Hacker News, lets anyone compare how well 18 frontier AI models perform when autonomously optimizing the nanoGPT training speedrun, a popular community challenge for efficient model training.