Explore what's new in AI

Simon Willison ports Moebius image pinpointing to the browser

Claude Code helped move the model into browser WebGPU, which makes the demo usable without a server-side Python stack.

Binary quantization shrinks embeddings and cuts CPU to GPU transfer costs

The same idea that speeds image pipelines also works for retrieval, where smaller vectors keep most ranking signal and move faster.

supermemory turns long-term recall into an inspectable memory graph

A visible graph makes agent memory easier to reason about, debug, and curate instead of leaving it hidden in opaque state.

Notion frames itself as an agent-first workspace in Akshay Kothari interview

The interview argues agents should spread expertise across an org, turning one workflow into reusable company infrastructure.

OpenRouter adds stablecoin payments with Coinbase

Letting agents and humans pay the same checkout with stablecoins is a small but practical step toward global AI commerce.

Manus AI turns SEO research into an executable workflow

The demo bundles keyword research, competitor analysis, audits, and rank tracking into one agent run, which is the right abstraction.

NotebookLM makes flashcards editable instead of forcing regeneration

Users can now fix wording, tweak answers, and add cards in place, which makes the flashcard stack much more usable.

OpenRouter detects 1Password and offers to save API keys

A small setup improvement, but it removes a common key-management snag and nudges teams toward safer credential handling.

Google Labs Jules reframes coding-agent evals around useful behavior

The point is to judge whether the agent notices what matters and anticipates needs, not just whether it can finish a task.

guinnesschen says a fully specified prompt saves more tokens than a small one

The token math is brutal, most cost comes from the agent doing the wrong thing downstream, not from the original prompt length.

guinnesschen says skills should be ranked by downstream invocations

A good skill is one other agents reuse, so invocation counts may be a better company metric than token totals or star counts.

Dexhorthy says agent planning docs should stay outside version control

The argument is that plans need discoverability and recovery, not branch semantics, so VCS is the wrong home for them.

Box adds HTML preview, editing, versions, and secure sharing

Box is making agent-produced HTML first-class, with browser editing, version history, and controlled sharing in one place.

Vercel decouples deploy from release with a feature flag switch

Ship code when it's ready, then flip a flag when the feature is. That's a clean way to shrink blast radius and merge faster.

Apodex 1.0 turns open-ended research questions into evidence-backed analysis

The demo walks step by step through a real profitability question, which makes it useful for finance and other research-heavy work.

Qwen distillation turns DeepSeek thinking style into a local orchestrator

The distill improves hard reasoning and cuts empty-answer failures, which is exactly what agent orchestration needs.

pi-ai 2 splits SDK imports into selective modules and a compat path

The breaking release trims global state and lets consumers pull only the pieces they need, which should make migrations cleaner.

Hermes Agent tops OpenRouter's daily ranking at 1.03T tokens

A trillion-token day is real production scale, not a vanity stat. It shows Hermes is now pushing serious traffic through the stack.

Jason Zhou open-sources a loop engineering template for agent runs

The template bakes in shared artifacts, logging, and verification so each loop compounds instead of starting from zero every time.

OpenCode adds unique user rankings to its model dashboard

Counting unique people, not token-heavy sessions, gives a truer read on adoption. It's a small metric change with big product implications.

React Doctor for Performance turns slow React code into a CLI fix

One command surfaces the bottlenecks and suggests fixes, which makes performance triage much easier to repeat across repos.

Gemini Spark beta turns Google AI Ultra into a personal agent

Spark is built around recurring personal workflows, from inbox triage to calendars and docs, not just chat responses.

Claude mobile adds Cowork for background tasks and scheduled runs

You can start a task on your phone, steer it from any device, and let it keep running even after the app closes.

GLM-5.2 tops DeepSWE v1.1 with a 44% pass@1 score

That makes GLM-5.2 the strongest open-source coding agent on the benchmark, with real long-horizon software tasks behind it.