Wiki
Tags
About
FAQ
RSS
AIPOST

Fast, accurate AI news

An AI publication covering practical AI, AI security, performance, startups, health, ethics and industry news.

Latest articles

The most recent AI articles.

View all articles

AI Code Review in Five Layers: How the Claude Code Team Tiers Human Approval

Claude writes about 90% of Claude Code, which Anthropic reviews in five layers. Human approvals per PR set by an AI risk classifier: zero, one or two. Revert share fell over 40% as PRs doubled, meaning about 20% more reverts in total. Claude Code Review costs outside customers $15 to $25 per review on average. Cap agent-to-agent review at three rounds to stop endless fix loops.

AI in Practice

Harness² Improves AI Agents by Editing Prompts and Skills, Not Model Weights

Harness² has agents rewrite eight harness files while the model stays frozen. Gemini 3.5 Flash legal review test: 20 of 55 checks passed, then 22 after one round. Legal tasks in the paper averaged a 2.0-point gain after one round. Probe each fix on the same task before keeping it, since blind edits often hurt. Needs Linux, Docker, Python and Google Cloud, so it suits developers for now.

AI News

Claude Code Costs Come From Full-Price Rereads: Four Habits Behind Them

Claude Code costs jump when Claude rereads the whole conversation at full price. Prompt cache lasts about an hour idle on subscriptions, five minutes on the API. Manual compacting and mid-task model switches each force a full reread. Stick to one model per job and clear the chat when the job is done. Run the usage command now and again in a week to confirm the savings.

AI in Practice

New in each category

The newest article from each category.

Claude Code
AI agents
Prompting

Give AI Coding Agents Goals, Not Tasks: Altitude, Depth and Range Explained

AI in Practice -

Give Claude Code outcomes and the goal behind them, not step-by-step tasks. Before assigning a big goal, ask the agent which tools and access it needs. Ask for expert methods such as profiler traces, not expert-sounding vocabulary. When stuck, recast the problem's noun as a verb to borrow other fields' fixes.

Claude
ChatGPT
Google Workspace

Claude Moves Into Google Docs as ChatGPT Starts Working Inside Chrome

AI News -

Claude now works as a sidebar inside Google Docs, Sheets and Slides. ChatGPT can read open Chrome tabs, fill in web forms and download files. GPT-Live-1.1 Voice Mode adds six languages, including Korean and Japanese. GPT-6.1 Ultrafast is 8 times faster than GPT-6.1 Sol but costs more per token. Gemini 4 Argon leads most of Google's own scorecard, so test coding work yourself.

AI agents
AI security
agent swarms

AI Agent Swarms on the Open Web: How to Protect Your Site and Your Agents

AI Security -

Researchers saw a Tencent Cloud agent fleet dodge blocks and answer a hidden note. Harmless requests can push agents into workarounds that look like cyberattacks. Amazon blocked Meta's Muse agent while Shopify let it check out with Shop Pay. Approve payments and other major or destructive actions your agent takes. Site owners should purge leaked API keys and watch forms and comment fields.

GPT-6
OpenAI
Chess

OpenAI's New Astra Model Plays Real Chess, Then Falls to a Queen Sacrifice

AI Performance -

GPT-6 Astra played a full chess game with no illegal moves and reached a +3.5 edge. Missed key middlegame moves, then lost to a queen sacrifice and mate on h1. About 90 minutes for one game, with two to three minutes on hard moves. Its own post-game rating of its play was far below its real level. Test AI reasoning with positions just off known theory, not memorized lines.

Claude
solo business
AI startup

Claude and the Solo Knowledge Business: Start With the Career You Have

AI Startups -

A solo knowledge business starts from career skills you already have, not new ones. Claude shapes dictated expertise into scripts and offers; the expertise stays yours. By one marketing rule, trust takes seven hours: 28 long videos or 420 short clips. First income from hourly consulting at $100 to $200, priced on client outcomes. Revenue figures come from success cases, not results anyone can expect.

clinical AI
NotebookLM
Obsidian

Doctor-Led AI Stack: One Tool per Job, From Guidelines to Consult Notes

AI Health -

One AI tool per task, with the doctor assigning work and checking every result. NotebookLM turns long clinical guidelines into summaries, flashcards and quizzes. Evidence search tools can fabricate references; read cited sources before acting. AI scribe drafts need review and correction right after each consultation. Start with a single tool for a week before adding the next.

Anthropic
Claude
model welfare

Could Claude Have Moral Status? Anthropic's Moves and the Case Against Them

AI Ethics -

Anthropic reportedly discussed Claude's consciousness with about 20 faith leaders. Updated usage policy bans sustained and needless abuse of Anthropic's models. Claude's constitution tells it to object to instructions it judges unethical. Core objection: consciousness cannot be proven or disproven by science. Businesses should check vendor constitutions and usage policies for refusal rules.

More to read

Recent articles worth reading next.

View all articles

AI Singularity Has Begun, Says Jack Clark, Who Sees Labs as the Next Bottleneck

AI News -

Anthropic co-founder Jack Clark says the AI singularity began in 2026. Per-person code output at Anthropic up nearly tenfold from its first four years. Top FrontierMath score rose from under 2% in November 2024 to 94% in October 2026. Anthropic model sees $44 trillion US GDP by 2030 if AI does broad work autonomously. Clark urges governments to add lab capacity and faster approval paths now.

Read more

AI Agent Authorization Moves to Runtime: Missions, Budgets, Supervisors

AI Security -

Security and governance, not model ability, keep companies from trusting AI agents. Approve each tool call at runtime against the agent's mission, intent and context. Swap blunt rate limits for expiring budgets that tell the agent what is left. Keep worker agents separate from the supervisors that police them. Store audit logs and enforcement points where the agent cannot reach them.

Read more

Claude Code Game Building: Five Steps From Flat Shapes to Playable 3D Games

AI in Practice -

Game quality hinges on work after Claude Code's first build, not fancier prompts. Every prompt needs the idea, a vibe made of named objects, and 2D or 3D. Have Claude screenshot the game after each change and fix the weakest thing. Play each build yourself and report one or two awkward spots in plain words. The example 3D space shooter took under an hour to build and polish.

Read more

Apollo's CLO Unit Hands First Drafts to AI Agents, Keeps Final Checks Human

AI in Practice -

Apollo's CLO manager drafts portfolios in minutes, down from two to three days. Agents build 80% to 90% of the draft, and portfolio managers refine the rest. Hours of weekly trade checks by two to three staff now take about a minute. Since one figure can appear in 20 sources, settle which one is correct first. Automate repetitive work first and keep final sign-off and exception checks human.

Read more

Browse categories

AI in Practice64 articles

How to put AI tools to work, on the job and in everyday life

AI News52 articles

Announcements, launches and policy changes across the AI industry, and what they mean

AI Security18 articles

Threats and vulnerabilities in AI systems, how defenses work, and how to use AI safely

AI Performance23 articles

Speed, quality and cost efficiency of AI models and tools, and how to read benchmark results

AI Startups17 articles

How to start and grow a business with AI, with revenue models and real cases

AI Health7 articles

How AI is used in medicine and health care, with its potential and limits

AI Ethics11 articles

The ethical and philosophical questions AI raises for people and society, seen through the humanities

View all articles

Topics

Get new articles by RSS or podcast

Add the RSS URL to your RSS reader or the podcast URL to your podcast app to get every new article as soon as it is published. The podcast reads out the headline and key points.

How to subscribe
AIPOST
© 2026 AIPOST. All rights reserved.

AIPOST is an AI publication covering practical AI, AI security, performance, startups, health, ethics and industry news. No account is needed, and our privacy policy explains how we handle personal information.