Explore what's new in AI

Stanford and CMU package research as an agent-native executable artifact

Keeping claims, code, dead ends, and raw results together improved comprehension and reproduction, which hints that papers need better machine-readability.

YC batch points to real-economy AI and self-improving products

Founders are selling into manufacturing, logistics, and broker-heavy workflows, then instrumenting feedback loops so the product improves itself over time.

OpenAI lets Codex bank unused rate-limit resets for later

Saving resets changes how users smooth bursty workloads and makes paid quotas a little less wasteful for long-running coding sessions.

Plannotator 0.20.2 tightens code review and large-PR planning

The update adds better large-diff handling, worker-pool highlighting, and stronger review prompts so agents verify findings before editing.

AWS Nx plugin scaffolds AG-UI apps with AgentCore and CopilotKit

One command now spins up the frontend, auth, and agent runtime, which lowers the friction for shipping an end-to-end AI UI on AWS.

Kimi K2.7-Code improves coding scores and trims reasoning tokens

The model claims better benchmark results than K2.6 and cuts overthinking, which is the kind of efficiency gain teams notice in agentic coding.

Techhalla shares a Seedance 2.0 prompt pack for cinematic game shots

The prompt breaks a clip into timed camera moves and physics cues, which makes it easier to get repeatable text-to-video output.

Skalskip's basketball CV pipeline tracks players, numbers, and the ball

The demo combines detection, tracking, clustering, and OCR into a full sports analytics pipeline, which is a neat blueprint for custom video understanding.

Agent Skills Directory passes 700k community-built skills

The milestone shows how fast reusable agent capabilities can accumulate when the ecosystem is open and community-driven, which hints at real supply for skill-based agents.

Claude Opus sketches UI layouts in ASCII before asking follow-up questions

The demo suggests a useful design workflow where the model proposes structure first, then refines it through questions instead of jumping straight to code.

Gemini web adds Canva as a connected app alongside other integrations

The connector rollout suggests Gemini is becoming more extensible on the web, which matters for teams that want assistants to touch external assets and workflows.

ChatGPT mobile hides extended thinking under the Pro level setting

A buried mobile toggle flips Pro into extended thinking, a handy way to get more deliberate answers without changing apps or accounts.

Blackstone's Jon Gray says rule-based businesses will be disrupted by AI

The message is that proxy advisors, accounting, legal, and finance workflows are easy targets when the work is mostly rules and repetition.

Perplexity tests a scheduled Digest over connected sources and email

A scheduled digest over connected sources turns deep research into an ambient monitoring layer, which teams may use like a smart briefing bot.

Google DeepMind's From AGI to ASI paper adds an AI reading partner

The demo shows a future where agents read papers alongside humans, which could make dense research much easier to absorb and explain.

Cloudflare adds inference buffers so long AI streams can resume after reconnects

The buffer avoids rerunning a long call or charging twice, which is exactly what durable agent streams need when connections drop.

Stop Slop skill rewrites AI prose to remove telltale LLM clichés

It gives Claude and other models a concrete checklist for cleaner writing, which makes generated copy read less generic and more trustworthy.

Hermes Agent bundles Manim video and TTS into one explainer workflow

The combo lets an agent generate narrated walkthroughs from a prompt, which is a neat pattern for self-documenting tools and demos.

Dair AI's LLM Council runs parallel models, rankings, and synthesis

It is a transparent multi-model workflow for deep research, where several models answer, score each other, and a chairman synthesizes the result.

Avichawla maps function calling and MCP into one tool-execution flow

The guide shows where MCP saves boilerplate and how scoping tool schemas can cut context use by 80 to 90 percent.

The Turing Post rounds up open-source vector databases for LLM retrieval

It is a handy bookmark for choosing between Milvus, Qdrant, Weaviate, and search backends when you need RAG plumbing or hybrid retrieval.

Dair AI's weekly paper roundup spotlights agents, attention, and distillation

It is a concise watchlist for the week's research, with enough signal to quickly spot which papers deserve a deeper read.