Explore what's new in AI

Sebastian Raschka shows how to run local coding agents with open-weight models

The guide ties together local models, runtimes, and coding harnesses, plus practical checks for RAM, tok/s, and tool-calling before you bet on local.

Opus 4.8 forgets names in long Claude Code sessions, hinting at KV limits

Long Claude Code sessions can still expose memory limits, even on big models. The screenshots are a useful reminder that context handling fails in mundane ways.

reach_vb says your top five Codex prompts shape how you work

The point is simple, your recurring prompts are becoming operating procedures. Treat them like reusable workflows instead of one-off experiments.

Codex helps reproduce a touch-sensing robot paper in a few hours

The repro shows that touch and proprioception can force compact physical-world representations, and Codex cut the setup time from weeks to hours.

willmcgugan's writing app targets distraction-free drafting with typography controls

The app is aimed at calmer drafting, with layout and typography controls tuned around keeping writers in flow instead of fighting the UI.

Cursor helps build an Electron app that fixes Samsung webcam flicker

The app exposes anti-flicker controls that the monitor and macOS hide, which solves a real hardware annoyance with a tiny custom tool.

Techhalla's Blender MCP workflow turns viewport scenes into stylized video

The workflow shows a repeatable path from 3D viewport animation to stylized video using Blender MCP, Magnific, and Seedance.

Techhalla shares a Seedance 2.0 prompt recipe for stable 15-second shots

The prompt template is a reusable recipe for controlled video generation, with shot timing, camera moves, and character consistency spelled out clearly.

Corey Ganim says AI should filter email and draft replies, not send them

Use AI as a filtering and drafting layer for email, not a blind sender. That keeps control with the human while still saving a lot of time.

OpenTrons OT-2 and Claude Code power a garage Alzheimer's screening workflow

This is a real lab automation story, not a toy demo. Claude Code runs the robot while LLMs are embedded in nearly every screening step.

UC Irvine study says ChatGPT made students faster on math but less likely to retain it

The big takeaway is that AI can shorten task time while reducing retention, so faster completion is not the same as better learning.

The writing-agent repo lays out a Python agent stack from configs to observability

If you're starting a Python agent project, this repo is a strong template because it separates orchestration, memory, MCP, evals, and tracing.

Nous Research adds cron-backed dashboard scripts with sandboxed paths

The cron update keeps scripts sandboxed, snapshots provider drift, and round-trips schedules cleanly, which is the kind of guardrail automation needs.

Pauliusztin's software production line splits agent work into plan, implement, and review

The diagram turns agent work into a reusable production line, which is a helpful model for repeatable planning, coding, and review loops.

Josef Chen shows routing and voting hit a co-failure ceiling across 67 models

The paper argues that model ensembles keep colliding on the same hard queries, so pairwise correlation is not enough to prove routing headroom.

hf-claude lets Claude Code use over 100 open models

This adapter makes Claude Code much more flexible for model routing and comparison, since you can swap in GLM, DeepSeek, Kimi, Qwen, Gemma, or Nemotron.

Bank of Korea finds AI cuts task time 3.8 percent but not output

The data says AI saves time on paper, but organizations often absorb that time into more review and more reports rather than higher throughput.

NUS survey maps World Action Models from prediction to action

The survey is a useful organizer for embodied AI, splitting world models by substrate, action coupling, deployment shape, and the core tradeoffs.

OpenKnowledgeAI turns papers into visual explainers and markdown notes

This notes app makes paper reading more durable by synthesizing explainers, then exporting the result to markdown for later reuse in Obsidian.

Google DeepMind's CUA takes #1 on Odysseys with Gemini 3.5 Flash

Odysseys rewards long, multi-step web workflows, so a #1 finish here is a useful sign that vision-only agents are improving on sustained tasks.

LangChain's Deep Agents course covers planning, context files, subagents, and memory

The course is a practical map of the deep-agent stack, from task planning and file-based context to long-term memory and subagent spawning.

Vercel skills CLI auto-matches a goal to the right Claude Code skill

The find-skills trick lets agents install and reuse the right dev tool instead of hand-writing another prompt flow. That makes skills more portable across assistants.

Athas v0.8.0 ships as a lightweight open-source code editor

Athas is positioning itself as a small, cross-platform editor with built-in chat, which makes it a plausible lightweight environment for agentic coding.