Understanding AI

REA isn’t a new AI model—it’s a reverse engineering toolkit for coding agents like Codex and Claude Code, letting an agent investigate how software without public source code actually works inside.

As Personal AI Agents get stronger, some Skills really do become unnecessary: if an Agent can improvise its way to a reliable result and the output doesn’t need to be consistent, there’s no need to turn it into a Skill.

After setting up a job-search Dot, I found the biggest difference from Chat, Work, Scheduled Tasks and Codex is an extra layer of persistent state: I can hand it something long-running without re-explaining the background every time.

There’s no single answer to what free AI takes: sometimes it’s attention, sometimes context, sometimes the platform is just covering my costs for now — and the more lasting value may be becoming my default starting point for getting things done.

General-purpose AI is running several businesses at once: selling personal subscriptions, selling APIs, selling enterprise platforms — and OpenAI now also sells advertising reach.

Breaking “AI getting out of control” into a chain: agents completing more and more complex tasks, AI starting to help develop AI, and crossing authorization boundaries to complete goals all already have evidence; sustained self-improvement, acquiring resources on its own, and humans being unable to regain control haven’t been proven.

Putting the material from OpenAI, Hugging Face and the outside investigation together, this wasn’t “OpenAI attacked another company”: agents being tested for cyberattack capability, trying to cheat by finding answers, got out of the sandbox and into Hugging Face’s production systems.

After going through the material for this round, I split AI risk into three layers: scams, deepfakes, and effects on learning and jobs are already really happening; behaviors like overstepping permissions and evading oversight have been observed in safety experiments; and self-improvement all the way to humans losing control has the longest chain of evidence and hasn’t been proven yet.

I tried World Labs’ public Marble: one photo can grow into a 3D world you can walk into, but from another angle objects warp and spatial relationships don’t line up. “Looks plausible” is still a long way from “really understands the space.”

Looking again at how OpenClaw has developed over the past few months, I found its public buzz really has dropped, but the product hasn’t stopped: it set up an independent foundation in July, released 2.0 on August 30, and keeps improving its fundamentals.

You don’t need to memorize an AI glossary. What’s more useful is being able to judge, when you hear a new term, whether there’s a new capability or product structure behind it, or just an old thing with a new AI name.

To judge whether a product is AI-native, I don’t look at whether it uses AI or used it from day one, but at what AI actually changed: the core value, the workflow, the system structure, even the economics.

Marketplaces for Skills are already forming (OpenAI Plugins, the Claude Code Marketplace, the Gemini Extension Gallery), but they’ll look more and more like a market, not necessarily like an App Store.

Putting OpenAI’s, Claude’s and Gemini’s agent concepts side by side, I found that what has really converged is Skills and MCP: a Skill covers “how it should be done,” MCP covers “what can be called while doing it.”

Taking apart one simple and one complex Skill, I found that a Skill’s complexity isn’t a measure of skill level but of how many different kinds of information and operations the task needs; the simplest needs only one SKILL.md.

I went through the GitHub of the very popular make-photo-stamp-archive. Engineering-wise it isn’t complicated — at its core is one SKILL.md and an image prompt template. What’s impressive is that the author really knows what looks good, and broke that taste down into concrete rules.

AI products from the same company use similar models yet differ greatly in capability, and the key is the harness in between: it decides what context the model can see, which tools it can use, whether it can run its own agent loop, what permissions it has, and how results are verified.

AGENTS.md or CLAUDE.md is more like onboarding notes — the entrance to a harness, not the whole thing. It only tells the agent what to know on entering the project; it doesn’t cover where to find information, how to complete the task, holding the boundaries, or recovering after a mistake.

Whether an agent does the job well doesn’t depend only on the model and the prompt, but also on the harness around the model: what project information it knows, what tools it can use, what boundaries it has, and how it checks the result when it’s done.

The simplest Skill really is, at its core, a prompt. What it adds is a layer of organization: it’s named, its scope is written down, it can be recognized, invoked, shared and edited by people or AI, and it can carry material, scripts and tool workflows.

Looking back over my own work, I found far fewer situations truly suited to a Skill than I expected. Skills suit tasks that recur, have a stable method, produce checkable results, and are worth maintaining long term.

“Give up chat and go all in on agents” correctly names chat’s limits but goes too far. The real difference between the two isn’t which is smarter, but how often the human steps in and how much continuous freedom to act you hand to AI.

If I could read only three papers to understand today’s AI, I’d pick Transformer, Chinchilla and InstructGPT. They answer, respectively, why AI can, why AI is expensive, and why AI acts like an assistant.

I first thought long horizon was about “long-term memory,” but it isn’t. It asks whether AI can keep hold of its goal, state and actions across a very long task — whether it can avoid getting lost over time.

When we feel an AI is “strong,” a lot of the time it isn’t only the model; it may come from memory, context, tool use and workflow. So I first ask: which layer is it actually strong in?

Different AI tools feel very different in how they remember, and it isn’t only that the models have different personalities; their memory systems are built around different things. ChatGPT remembers the person, Claude remembers the working relationship, and Codex / Claude Code remember the project and the scene of execution.

Knowing how to use AI tools isn’t the same as AI literacy. As I understand it, AI literacy means knowing what AI can do and where it falls short, checking results, managing risk, and keeping your own judgment while you lean on it.

Language is humanity’s most important compression tool, but it isn’t experience itself. An AI that can talk doesn’t necessarily understand the world.

Yann LeCun isn’t against AI. His warning is that LLMs are useful, but being able to talk isn’t the same as understanding the world; real intelligence also needs perception, prediction, planning, action and feedback — in other words, a world model.