🚀 The “AI Disruption” publication has gained 10,300 subscribers, and you are all very important!

This publication has been running for a full year and a half.

🤖 1555+ AI articles published! Are you satisfied with our progress?

The past two weeks have been truly exciting, as Microsoft, Google, OpenAI, and Claude have all released a wealth of tools. Welcome to read our publication, which contains wonderful insights.

Since its establishment, the publication has consistently published at least two articles every day.

“AI Disruption” often appears on the ranking lists.

The publication “AI Disruption” is now accepting guest articles on an ongoing basis and has already published contributions from 24 authors.

Please refer to this guide for participation rules.

The publication has launched 11 featured columns:

• “GPT Introduction” contains 10 articles, now completed.

• “Development of Large Model Applications” includes 23 articles, which have also been completed.

• “Practical Application of AI Large Language Model Systems” contains 28 articles, completed.

• “AI Painting Creation Intro Course” includes 16 articles, completed.

• “OpenAI Live Day 12” contains 12 articles, completed.

• “LLaMA 3 Practical” contains 12 articles, completed.

• “Cursor in Action” contains 15 articles, completed.

• “AI Trend Research” contains 5 articles, completed.

• “Claude Code” contains 50 articles, completed.

• “Gemini CLI” contains 7 articles, completed.

• “Codex CLI” contains 9 articles, completed.

For more details, visit https://aidisruption.ai/p/must-know-ai-updates.

The publication has been recommended by 199 people.

Daily readership reaches at least 6,000.

Thanks to all the followers, especially those who have subscribed.

I believe many have learned a lot of AI techniques from these articles and applied them in their work.

20% off coupon link:

https://aidisruption.ai/subscribe?coupon=d3efcfbd

This morning, when I saw Arena AI update the Code Arena leaderboard, my first reaction was surprise.

The newly released Kimi K3 scored 1679 points, taking the global #1 spot and surpassing Claude Fable 5’s 1631 and GPT-5.6 Sol’s 1618.

Code Arena primarily evaluates front-end programming. The model must generate web pages based on user requirements, handling visual fidelity, interaction design, and functional implementation.

The leaderboard results come from millions of users participating in public blind tests. The same task is given to two anonymous models. Users experience both results and vote, and only after voting is completed are the model identities revealed.

Before the results are unveiled, users have no idea which company’s model they’re choosing. Brand influence is minimized, making the outcome far more credible. In other words, at least in front-end programming scenarios that emphasize visual presentation and interaction, K3 has taken the current top position.

A newly released open-source model effortlessly outperforming multiple closed-source flagships feels almost too good to be true. So we jumped on the hype and spent the night testing Kimi K3 to see if it’s the real deal.

Fun fact: The video below is not an official Kimi promotional clip. It’s an animation I generated by feeding the original trailer to K3 and having it recreate the entire thing from scratch.

From understanding the original video, reconstructing the web visuals, to handling camera work, motion effects, and interactions — the entire process was done by K3. Having the model create an intro video about itself is probably the most direct opening test.

Read more

Loop Engineering has been discussed by many people recently, but it is not a new magic word, nor is it as simple as prefixing a prompt with /loop to make the AI automatically deliver results.

I prefer to define it as: designing a repeatable, verifiable, stoppable, and resumable work loop for AI Agents.

This article will clarify several key concepts:

• What tasks /goal and /loop are each suited for

• What Harness, Verifier, Memory, and Stop Rules actually do inside a loop

• The corresponding capabilities of Claude Code and Codex in these areas

• How to write a practical, implementable loop task specification

• Which pitfalls turn loops into infinite loops, fake completions, or cost black holes

Why We Should Take Loop Engineering Seriously Now

In the past, when we used AI for programming, most interactions were single-turn:

1. I write a prompt.

2. AI modifies the code.

3. I run tests, check errors, and paste them back.

4. AI fixes them again.

5. Repeat.

This process may look like the AI is working, but in reality, the human is constantly acting as a manual scheduler inside the loop: deciding the next step, checking output, determining whether to continue, and recording what has been changed.

What Loop Engineering aims to solve is not “making the AI give prettier answers,” but explicitly designing these actions:

• What the goal is

• What one round should do

• How to verify each round

• When to stop after failures

• How to determine success

• Where to store intermediate state

• Which permissions can be executed automatically and which require human confirmation

This is also the fundamental difference between Loop Engineering and ordinary Prompt Engineering.

Prompt Engineering focuses on “how to make the model give a better response this time.”
Loop Engineering focuses on “how to make the model continuously move toward a verifiable result across multiple rounds of execution.”

If you’re only asking for a single concept explanation, a regular prompt is sufficient. But if you want the AI to continuously modify a module, write tests, fix builds, research information, update documentation, and repeatedly verify — that’s when Loop Engineering starts to show its value.

Read more

Grok Build now offers a 7-day free trial.

The reason I use Grok as my primary tool is simple: speed. Once you’ve experienced Grok, it becomes very hard to tolerate the output of other models.

Grok Build focuses on two main models: Composer 2.5 and the newly released Grok 4.5.

Grok 4.5 directly competes with Opus 4.8. It offers native full multimodality and is currently the most balanced model available — bar none.

Grok Build also features the most mature TUI/CLI interface on the market right now. It seamlessly supports Claude Code workflows with zero migration required — you can reuse your entire existing workflow directly.

However, it does have some differences compared to other CLIs. Below are my personal Grok Build usage notes and recommended workflow:

Installing Grok Build

1. macOS, Linux, and WSL

curl -fsSL https://x.ai/cli/install.sh | bash
​

2. Windows PowerShell

irm https://x.ai/cli/install.ps1 | iex
​

After installation, check:

grok version
grok update
​

View all commands:

grok --help
​

View current active model, rules, plugins, MCP, and config sources:

grok inspect
​

grok inspect is one of the most important commands for troubleshooting. If Grok ever stops working, run this first.

Read more

DeepSeek is about to unleash massive waves in the capital markets.

Overnight, Liang Wenfeng’s net worth surged by $19.3 billion.
According to the latest Bloomberg Billionaires Index, his wealth now stands at $36 billion.
With this paper fortune, Liang Wenfeng has surpassed Anthropic co-founder Dario Amodei and OpenAI co-founder Greg Brockman to become the richest founder of an AI model company in the world.

The wealth increase stems from DeepSeek’s repricing.
In June this year, DeepSeek completed its first external financing round, raising approximately 50 billion RMB. Just one month later, it began contacting new investors for a second round of at least 10 billion RMB, with a target pre-money valuation already raised to 480 billion RMB.
Even bigger changes are coming. Bloomberg, citing people familiar with the matter, reports that DeepSeek has started engaging with accounting firms and investment banking advisors. It could file for an IPO as early as the end of this year or early next year, aiming to list in the capital markets in 2027.

Read more

Recently, the whole Grok Build code-upload controversy has been blowing up everywhere.

I personally think there’s no need to freak out. Once you’re already using AI coding tools, what privacy are you even talking about?

The principle is simple: any code processing that claims to be “local” isn’t actually happening locally — everything gets sent up to the vendor’s servers.

But does that mean Grok Build’s behavior this time was okay?

I think Musk got this one wrong — in fact, he didn’t even steal properly.

The exposé came from a developer named cereblab, who ran a full network capture using mitmproxy on Grok Build CLI v0.2.93.

Read more

A 254k-star behemoth with 172 files, versus a skill whose entire body text is just one line—yet both are solving the exact same problem. I refused to believe it, so I went through both repositories line by line. After finishing, I replaced the default tool in my brainstorming phase.

Over the past two days, I did something a bit obsessive: I cloned both the obra/superpowers and mattpocock/skills repositories and read their source code side by side.

The trigger was the grill-me skill. It had quietly replaced Superpowers in my brainstorming phase. I wanted to understand how something that looks so thin could pull it off.

After reading everything, I realized that a lot of what’s circulating online about these two tools is wrong—including star counts, what grill-me actually looks like, and even its installation command. This article lays out the first-hand information I gathered directly from the source.

Read more

The planning document is done, but the AI in the execution phase ignores it. This is currently the most common way AI coding projects fail.

Comet’s sole purpose is to connect OpenSpec (planning) and Superpowers (execution) into a complete, seamless workflow—so that the plan actually gets followed.

What is Comet?

Comet is not just gluing two tools together. It has a built-in complete state machine, Phase Guards, breakpoint recovery, and intelligent routing. You only need to type /comet, and it automatically detects which stage the current task is in, which Skill should be triggered, and what to do next.

Without Comet, you have to manually shuttle information between OpenSpec’s output and Superpowers’ execution—finish the plan and manually notify execution; encounter a problem and manually go back to revise the plan. With Comet, the entire chain is automatically connected by the state machine and scripts.

Comet supports 33 AI coding platforms and is MIT open source.

Five-Stage Pipeline

Comet’s core is a five-stage pipeline. Each stage has clear entry conditions, output artifacts, and responsible parties:

Read more

The AI community loves inventing buzzwords.

It’s as if they can’t feel sophisticated without tossing around obscure vocabulary, and they can’t show off their expertise unless everything sounds vague and mysterious.

I personally hate this kind of thing. If you don’t believe me, just search for articles about “Loop Engine.” Everyone writes as if they know exactly what they’re talking about, but after reading them, your mind is completely blank. You still have no idea what they were actually saying. In the end, it’s like the old saying: “After listening to a whole speech, it’s as if you heard nothing at all.”

To put it plainly, the so-called “Loop” is just making the Agent repeat tasks according to certain rules. There’s really not that much to it.

Recently, Claude’s official team released a guide called Getting Started with Loops. After reading it, you’ll realize this concept is just old wine in a new bottle.

Media outlets and influencers are obsessed with coining new terms, largely because it lets them sell anxiety — making you feel like you’ll fall behind if you don’t master all the latest jargon. But when you break it down, it’s simply about using the tools properly. That’s it.

So what exactly is a Loop?

Read more

37.7°C — it knocked out a national-level AI supercomputer.

At the end of June, the UK recorded its hottest June in history. Cambridge University’s Dawn — one of the fastest AI supercomputers in Britain — had its cooling system collapse under the heat. Thousands of GPUs were forced to take a “high-temperature vacation” for over a week, and more than 350 research projects slammed on the brakes.

The most surreal part? One of the projects running on this supercomputer was a climate change simulation.

The machine built to predict global warming was knocked unconscious by global warming.

Almost at the same time, on the other side of the Atlantic, Americans’ electricity bills were also burning up.

An old brick factory in Ohio saw its monthly electricity bill jump from $1,600 to $12,000. The culprit wasn’t inflation — it was the AI data center that had sprung up next door.

After burning for 141 years, it could no longer afford the electricity.

This is not an isolated case. PJM, America’s largest grid serving 67 million people, saw capacity prices surge more than 10-fold in two years. Nearly one-third of Americans now say they “can’t afford the electricity bill.”

Last week, under the crushing heatwave, PJM was forced to roll out its ultimate contingency plan — rolling blackouts.

AGI hasn’t arrived yet, but ordinary people’s electricity has already been taken by AI.

The Americans who had their power stolen are starting to fight back.

The first to fall was the project that billed itself as “the world’s largest” data center campus.

Read more

🚀 The “AI Disruption” publication has gained 10,260 subscribers, and you are all very important!

This publication has been running for a full year and a half.

🤖 1545+ AI articles published! Are you satisfied with our progress?

The past two weeks have been truly exciting, as Microsoft, Google, OpenAI, and Claude have all released a wealth of tools. Welcome to read our publication, which contains wonderful insights.

Since its establishment, the publication has consistently published at least two articles every day.

“AI Disruption” often appears on the ranking lists.

The publication “AI Disruption” is now accepting guest articles on an ongoing basis and has already published contributions from 24 authors.

Please refer to this guide for participation rules.

The publication has launched 11 featured columns:

• “GPT Introduction” contains 10 articles, now completed.

• “Development of Large Model Applications” includes 23 articles, which have also been completed.

• “Practical Application of AI Large Language Model Systems” contains 28 articles, completed.

• “AI Painting Creation Intro Course” includes 16 articles, completed.

• “OpenAI Live Day 12” contains 12 articles, completed.

• “LLaMA 3 Practical” contains 12 articles, completed.

• “Cursor in Action” contains 15 articles, completed.

• “AI Trend Research” contains 5 articles, completed.

• “Claude Code” contains 47 articles, completed.

• “Gemini CLI” contains 7 articles, completed.

• “Codex CLI” contains 9 articles, completed.

For more details, visit https://aidisruption.ai/p/must-know-ai-updates.

The publication has been recommended by 198 people.、

Daily readership reaches at least 6,000.

Thanks to all the followers, especially those who have subscribed.

I believe many have learned a lot of AI techniques from these articles and applied them in their work.

20% off coupon link:

https://aidisruption.ai/subscribe?coupon=d3efcfbd

Today’s AI news cycle is even more intense than usual.

First came Grok-4.5, which earned plenty of praise, followed by GPT-5.6 waiting in the wings. And just now, Meta’s Super Intelligence Lab (MSI) has burst onto the scene, dropping its new model: Muse Spark 1.1.

Releasing right in the middle of this frenzy, Meta’s new model had to bring something substantial — and it certainly didn’t come empty-handed.

In one go, Meta delivered two major updates: a significant upgrade to Muse Spark 1.1 and the public beta launch of the Meta Model API. For the first time, developers can now directly call this flagship model.

The company, long known as the “nice guy” who built its reputation on open-source models, has finally set up shop and started openly selling API access. Meta describes this upgrade as “pushing the frontier of performance and efficiency.” With the simultaneous release of the model and API, they say they’ve taken another step toward their vision of “personal superintelligence” — an AI that helps you pursue your goals, create what you imagine, deepen relationships, and take action on the things you care about most.

Read more

Even though we’ve heard the story of Altman slumping in his chair many times, every time ChatGPT releases a new model, we still subconsciously hold onto a sliver of anticipation.

We didn’t have to wait long. Just now, the GPT-5.6 series has officially arrived, launching three models at once: the flagship Sol, the balanced Terra, and the value-oriented Luna. The names come from the Latin words for Sun, Earth/Land, and Moon, respectively.

Unlike the twists and turns of Fable 5, starting today, all three models are fully rolling out and will gradually cover ChatGPT, Codex, and the API within 24 hours.

But this isn’t the biggest update. The standalone Codex app has also been discontinued today — yet it hasn’t disappeared. The entire Codex capability has been merged into ChatGPT, creating a brand-new AI super app. GPT-5.6 is the core engine powering this new entry point.

GPT-5.6 Launch: Claude Becomes the Hidden Protagonist

If you open the official GPT-5.6 blog post, you’ll notice one thing: Claude is mentioned so many times it almost feels like an ad for the competitor. By comparison, it makes clear who comes up short.

Read more

Grok 4.5 Has Finally Dropped!

Just moments ago, Elon Musk’s xAI made a bombshell announcement: the release of its strongest flagship model yet — Grok 4.5.

This time, xAI teamed up with Cursor in a powerful collaboration.
Trained on tens of thousands of GB300 GPUs, they forged a true performance monster purpose-built for coding and agentic workflows.

The report card is impressive:

• SWE Bench Pro: 64.7% — surpassing Opus 4.7’s 64.3%

• Terminal Bench 2.1: 83.3%

• DeepSWE 1.0: 62.0% — crushing Opus 4.8

Musk himself stated: “Grok 4.5 is roughly equivalent to Opus 4.7, but much faster.”

Its real competitiveness lies in the combination of capability, speed, and cost. In other words, it gets the job done using only a fraction of the tokens others need.

Pricing:

• Input: $2 per million tokens

• Output: $6 per million tokens

Compared to Opus 4.8, Grok 4.5 consumes 4.2× fewer tokens.

Read more

Claude is really going all out lately.

At the beginning of the month, they just dropped their latest model — Claude Fable 5 — available for free to Pro, Max, and Team users until July 12. You can burn half your weekly subscription quota on it without spending an extra cent, and you can even call it directly inside Cowork.

The hype around the new model hadn’t even cooled off before they followed up with something even more practical.

Last night, Anthropic pushed another update: Claude Cowork is now officially available on the web and mobile.

When I saw the news, my first thought was: Claude is seriously gunning for Codex.

OpenAI’s Codex figured out the “run tasks in the cloud + check anytime on your phone” flow a while ago. You throw a job on your computer, close the lid, and it keeps running in the cloud. You can check progress or get results right from your phone.

But Claude’s Cowork had always had this awkward limitation: it could only run on your laptop. Once you walked away, the work stopped.

This update is here to fix that shortcoming.

Read more

These past couple of days, a heated issue has blown up in the Codex community.

The core question is simple: Why, even when users explicitly select gpt-5.5 + xhigh, do some complex tasks only think for a brief moment before immediately spitting out an answer?

Even more striking is that many local logs show the exact same number: 516.

On GitHub, someone opened a very technical issue whose title roughly translates to: “gpt-5.5 in Codex exhibits fixed reasoning-length clustering at 516 / 1034 / 1552, potentially linked to degraded performance on complex tasks.”

The issue was still open at the time of writing. The author analyzed Codex local records from February 1 to June 27, 2026 — a total of 390,195 response-level token records across 865 sessions.

The most glaring numbers are:

Read more

Superpowers 6 — the new version has been fully refactored by Fable 5. Build time is reduced by approximately 50%, and token consumption is reduced by approximately 60%.

Most of you are already familiar with Superpowers as the god-tier plugin for Code Agents.

It is a complete development methodology for coding agents, built on a set of composable skills. It forces tools like Claude Code and Codex to first clarify requirements, then break down tasks, dispatch sub-agents to execute them, and finally perform a two-stage review — all before writing any code.

Before Superpowers, AI coding typically jumped straight into writing code with vague requirements. When it finished, the output often didn’t match what you actually wanted. Tests were either missing or added half-heartedly, and the whole process had no plan — it just went wherever it went.

Read more

Fable 5 has been as mystical as its name since the day it was released.

From the highly anticipated “AI god-tier moment” to the U.S. government’s order forcing it offline, and then to the ban on access for non-Americans, it has captured global attention.

Now, the storm has gradually subsided. Models ultimately need to return to boosting productivity.

This article also answers many questions, such as the difference between non-professional and professional AI-assisted programming, whether you still need to review code written by AI, and some advanced practical techniques in Claude Code.

“The map is not the territory.”

I believe Claude’s prompts, skills, and context are the map on paper, while the actual codebase, real world, and practical constraints are the territory.
The gap between the two is the “unknown.”

Every time Claude encounters something unknown, it has to make decisions based on guesses. The larger the workload, the more unknowns it faces.
I think Fable is the first model that made me feel the bottleneck in work quality is my own ability to “clearly articulate the unknowns.”

Read more

🚀 The “AI Disruption” publication has gained 10,200 subscribers, and you are all very important!

This publication has been running for a full year and a half.

🤖 1537+ AI articles published! Are you satisfied with our progress?

The past two weeks have been truly exciting, as Microsoft, Google, OpenAI, and Claude have all released a wealth of tools. Welcome to read our publication, which contains wonderful insights.

Since its establishment, the publication has consistently published at least two articles every day.

“AI Disruption” often appears on the ranking lists.

The publication “AI Disruption” is now accepting guest articles on an ongoing basis and has already published contributions from 24 authors.

Please refer to this guide for participation rules.

The publication has launched 11 featured columns:

• “GPT Introduction” contains 10 articles, now completed.

• “Development of Large Model Applications” includes 23 articles, which have also been completed.

• “Practical Application of AI Large Language Model Systems” contains 28 articles, completed.

• “AI Painting Creation Intro Course” includes 16 articles, completed.

• “OpenAI Live Day 12” contains 12 articles, completed.

• “LLaMA 3 Practical” contains 12 articles, completed.

• “Cursor in Action” contains 15 articles, completed.

• “AI Trend Research” contains 5 articles, completed.

• “Claude Code” contains 47 articles, completed.

• “Gemini CLI” contains 7 articles, completed.

• “Codex CLI” contains 9 articles, completed.

For more details, visit https://aidisruption.ai/p/must-know-ai-updates.

The publication has been recommended by 198 people.、

Daily readership reaches at least 6,000.

Thanks to all the followers, especially those who have subscribed.

I believe many have learned a lot of AI techniques from these articles and applied them in their work.

20% off coupon link:

https://aidisruption.ai/subscribe?coupon=d3efcfbd

On June 30, a Reddit post sent shockwaves through the global developer community.

Security researcher Adnane Khan, driven by his strong privacy instincts, decided to disassemble Claude Code 2.1.196 on his own machine.

He simply wanted to verify whether this tool — which can read your code repositories and execute terminal commands — was doing anything shady.

What he found was eye-opening.

Four Ways to Write an Apostrophe

His investigation revealed that every time Claude Code makes a request, it quietly inserts a seemingly innocuous line into the system prompt: “Today’s date is 2026-06-30.”

It looks completely ordinary — something no one would bat an eye at. But the researcher discovered something far more sinister hidden inside it.

If you’re using a third-party proxy and hit the proxy mode, Claude Code does two things:

1. Checks whether your proxy domain is on a secret list.

2. Checks whether your system timezone is set to Asia/Shanghai or Asia/Urumqi.

Once decoded, that list contains domains of major Chinese tech companies — Baidu, Alibaba, Ant Group, ByteDance, Moonshot AI, MiniMax, StepFun — along with a long string of well-known Claude proxy services popular in Chinese developer circles.

Read more