Explore what's new in AI

Isaak Freeman's thesis maps a path from worm emulation to human-scale brains

The thesis argues the remaining bottleneck is data acquisition, not raw compute, and puts human brain emulation on a concrete scaling path. It's a provocative read for anyone tracking long-horizon AI and neurotech.

Vtrivedy10 asks which agent harness primitives still survive model gains

The thread asks which harness primitives are truly universal, and which are just today's best practice. Useful if you're designing agent stacks for the long haul.

tokenjuice beta 0.4 adds Claude, Pi, and CJK support

Beta 0.4 adds native Pi and Claude extensions, an escape hatch, Linux binaries, and CJK support. It looks like a solid terminal-output compaction release.

Matt Pocock's domain-model skill adds docs and ADRs during ideation

The skill turns architecture conversations into docs and ADRs while you iterate. That's a practical way to keep domain language and codebase decisions aligned.

Databricks says Genie Code now writes more code than humans on its platform

That's a rare first-party adoption metric, and it suggests the product is already doing real work at scale. It also reinforces the shift from prompting to supervising.

zeeg says Sentry should automate skill-writing instead of hand-writing prompts

The point is to automate skill generation so teams improve the process, not one-off prompts. That makes agent workflows more repeatable as models keep changing.

athas-tui combines multiple CLI agents, git state, and worktrees in one TUI

This TUI pulls multiple agents, git state, and worktrees into one command center. Handy if you juggle parallel coding sessions and want less tab chaos.

The Turing Post breaks down tokens, context windows, and pricing

It's a clean explainer for onboarding teams to tokenization, context limits, and per-token pricing. Useful if you need to align non-ML folks quickly.

Claude Code npm package switches to a native binary in v2.1.113

Shipping a native binary should reduce install friction and improve runtime behavior. That's a meaningful upgrade for anyone living inside Claude Code.

Ollama adds GitHub Copilot CLI support for repo-aware terminal agents

Copilot CLI support makes Ollama more useful for repo-aware terminal work, from issue triage to scaffolding changes. It broadens Ollama's role as the local model runtime.

Avichawla explains how MCP and Skills split wiring from procedural knowledge

The key distinction is wiring versus procedure: MCP connects tools, Skills teach usage. That makes production agent design much easier to reason about.

Uncle Bob Martin splits coding work into architect, coder, and reviewer agents

The workflow is concrete: one agent plans, one codes, one reviews. It's a useful decomposition pattern if you're running multiple agents on the same codebase.

johnowhitaker uses Codex to animate plasmid proteins into an explainer

Codex turns a plasmid notebook into a video explainer in one shot, which is a nice example of agents helping with scientific communication, not just coding.

OpenCode Session Graph visualizes how chat branches connect

The plugin shows how forks connect across sessions, which is handy when you are running multiple branches of the same agent workflow. Good lightweight tooling for long-lived debugging.

browser_use launches Browser Harness for self-healing browser tasks

Browser Harness aims to survive UI drift and finish web tasks without brittle scripts. That matters if you're trying to make browser agents reliable enough for real work.

Sakana AI ships Digital Ecosystems, a browser sandbox for competing neural species

It runs in-browser with live training, so you can steer and perturb a multi-species system without installs. Useful for studying edge-of-chaos behavior.

ClaudeDevs adds /usage to break down Claude Code spending by feature

The new command makes Claude Code spend visible by feature, which should help teams debug usage spikes and understand where agent costs actually come from.

Codex computer use turns Mac apps, browsers, and automations into one workflow

The win is less typing and more orchestration, from Mac apps to browser tasks to scheduled automations. This is a good snapshot of how computer-use is landing in practice.

Google's CoDaS agent mines wearable data for mental-health and metabolic biomarkers

Across 9,279 observations, CoDaS found candidate biomarkers for mental health and metabolism. It shows how agentic loops can speed up slow, expert-heavy discovery work.

Claude Security appears in Claude Code with repository scan prompts

The leaked UI suggests Anthropic is testing repo scans and vulnerability reporting inside Claude Code. If it ships, security review becomes part of the agent loop.

AI Engineer Europe talks argue friction still matters in agentic coding

The talks argue that faster code generation still creates hidden debt unless teams keep friction, linting, and human review in the loop. Good antidote to slop speed.

Gemini Personal Intelligence pulls facts from Gmail, Photos, and Search

The demo shows Gemini pulling passport data from connected services and keeping it on-device. That makes personal memory a real product surface, not a chatbot gimmick.