Explore what's new in AI

Higgsfield chains Claude, Nano Banana Pro, and Seedance into a cartoon episode

Claude writes the story, Nano Banana Pro draws the boards, and Seedance 2.0 animates the cut. It is a concrete end-to-end media pipeline inside agent tooling.

Addy Osmani and Shubham Saboo map five patterns for long-running AI agents

Checkpointing, delegated approval, layered memory, ambient processing, and fleet orchestration give a practical checklist for agents that run for hours.

Mixedbread adds traces for agentic search calls in its dashboard

Mixedbread's new traces show every search call an agent makes and let you tune instructions from the dashboard. Visibility like this is what turns search into an operable system.

Demis Hassabis says coding is where AI self-improvement starts

Demis Hassabis frames coding as the first place AI self-improvement can compound, because outputs are verifiable and synthetic data is abundant. That is the lab's strategic thesis.

Gemini in Chrome rolls out to some European users with Skills and Live

The Canary rollout suggests Google is widening browser-assistant access, with Skills and Gemini Live in the same experience. Worth watching if you build Chrome-first workflows.

MIT's GenCAD turns a photo into editable parametric CAD code

GenCAD turns a photo of a part into editable parametric CAD code instead of just a mesh. That could collapse a weeks-long industrial design workflow into an upload and edit loop.

0xblacklight says coding-agent take-homes need harder eval design

A spec can become a toy app too easily now, so take-home tests need to probe judgment, architecture, and agent fluency instead of just outputting a working demo.

badlogicgames turns a €9 toy robot into an Android-phone agent

A €9 toy robot becomes an Android-phone agent once WebUSB, an H-bridge, and a browser controller take over. Cheap hardware plus agent software is a potent combo.

slime's SGLang and Ray stack explains async RL inference

The walkthrough maps async RL inference from placement groups to rollout scheduling. Good reference if you are building long-running training systems with SGLang and Ray.

Webwright tops the Odysseys long-horizon web-agent benchmark

Webwright sits at the top of a benchmark that stresses multi-hour web workflows with planning, evidence gathering, and verification. A useful bar for long-horizon agents.

fast-rlm wires web search and Goodreads into one agent demo

fast-rlm shows a practical search-and-reason loop, using web search, tool use, and recommendation logic in one demo. Useful as a compact pattern for agentic search systems.

Google Research wants models to hedge honestly instead of bluffing confidence

Google wants models to surface honest confidence, so agents know when to answer directly and when to reach for search. Better calibration beats fake certainty.

Meta says coding agents learn better from summaries than raw logs

Summarizing each attempt helps coding agents reuse past work and rank candidates better than raw logs. Meta reports solid gains on SWE-bench and Terminal-Bench.

Andrej Karpathy imagines neural computers replacing fixed app interfaces

Karpathy's framing is bigger than code assist, he imagines interfaces generated on demand from raw intent and media. That would flip the app model inside out.

Anthropic workshop shows how to build agents that run for hours

The session is about the real production problem, keeping agents alive across long tasks with checkpointing, recovery, and resumable context. That is where most systems fail.

vLLM rejects low-signal PR training and asks for verifiable contributions

vLLM is now rejecting PR-training style contributions and asking for verifiable production context instead. That is a pragmatic response to AI-generated maintainer noise.

Natural-Language Agent Harnesses makes control layers inspectable and portable

Expressing stage logic, memory, and stop conditions in structured language makes agent control layers inspectable and portable. The SWE-bench results show the behavior really changes.

Adapt the Interface, Not the Model reports 88.5% gains across 18 backbones

The paper says a learned harness, not model tweaks, can deliver huge gains across 18 backbones. That makes wrapper design a first-class engineering lever.

PufferLib runs game-playing neural nets in the browser with WASM

PufferLib turns game-playing RL into something you can spin up quickly, and the wasm demo makes it easier to share and inspect. Good for lightweight RL experiments.

steipete builds a GitHub dashboard for repos, issues, and release status

A compact GitHub dashboard turns repo status, issue state, and release distance into one glanceable view. Useful when you are juggling many repos and need fast triage.

cmux splits Codex work between the Mac app and CLI

Splitting knowledge work from coding across the Mac app and CLI is a clean pattern for agent users who want fewer context switches and a better boundary.

Sebastian Raschka adds DeepSeek sparse attention to his LLMs-from-scratch repo

A from-scratch DSA implementation with motivation, reference code, and a worked example makes DeepSeek's sparse-attention idea much easier to study and port.