Anyone who has learned a foreign language as a younger person but forgotten it as they got older will know: skills are something you have to keep using, or you slowly forget them. Same is true for musical instruments, painting, dance or playing all kinds of sports of games. Writing is a language skill that takes regular practice, and this is true for all kinds of writing, be it for prose, verse, or computer code. There are some things that a person simply must do regularly in order to maintain their level of skill over time, if not slowly improve.
If I have one piece of advice to offer anyone on how to use AI, it is this: never ask an AI to solve a problem for you, ask it to teach you how to solve the problem yourself, and then do it yourself.
Whenever I hear people in the tech industry talk about how AI will make it so that anyone can create their own apps, this is technically true, but I can't help but role my eyes a little. This is such a narrow-minded way of thinking, trapped in the world of the late 2000s and early 2010s when smartphones were the new big thing and selling apps in smartphone app stores was a booming industry. I say, forget about using AI to write apps, there are other major implications that we should be thinking about when we consider the possibility than anyone can make their own apps now.
People who still write code by hand will become more rare, and they will have skills that may become extremely valuable in places where AI cannot, or should not be used for reasons of trust, safety, and mitigating risk of liability. People who still bother to take the time to write code by hand will maintain their skills, and though they may be slower at delivering a product, you can be sure that the skill they put into it will have the added benefit of keeping that art form alive. But as with other art forms, it will probably be never be as well-appreciated as it should, nor will it have much commercial value. Governments may need to devote a portion of federal tax revenue to maintaining the art of software development and pure research in programming language theory, just to make sure that "brain trust," that community of people with these essential skills, doesn't go extinct.
If AI can create an app for you, there is no need to buy software anymore. So the whole idea of an app stores or app platforms becomes literally worthless. When smartphones were invented, "there's an app for that" was the mantra people would invoke for any problem. Nowadays, the mantra is "just ask AI to do it for you." So any business that has made money selling apps had better be scrambling now to find a new line of business, a new "core competency" as they like to say, or that business will not have much of a future.
But the important thing is: the idea of apps, as in software applications, is going away. And "apps" as an idea may not be the only thing going away. The idea of information property itself seems to be antiquated now that AI companies have been given carte-blanche to use the hard work of artists to train their AI without compensating the artists themselves. Will the idea of private property also eventually dissapear?
From now on, people are going to be asking for app platforms, so-called "Agentic AI" with the best generative AI available to help you solve your problems. What Claude Code calls "skills" are really little algorithmic tools, classical computing functions, that are built-in to the user intefface (the "Agent") that the LLM can use to gather information and compile that information into a software solution.
These agentic "skills" will replace apps, and "Agentic AI" will replace the idea of apps. Really, "skills" are just ways of combining old-fashioned deterministic computing algorithms, like pocket calculators which don't make mistakes, or computer programs that can count how many times the letter "R" appears in the word "strawberry," with the non-deterministic algorithms that make the AI appear to be having a conversation with you. Agentic AI gives you the best of both: tools to do classical computing, and the ability to teach you how to use those tools yourself, and improve your own personal skills in using a computer to solve problems.
What's hilarious is: it's all been done before!
What people are calling "Agentic AI" nowadays is just a new take on an idea that was first invented back in the 1980s. In the 1970s, the Lisp programming language was considered the language of AI, the reason being that not only could it perform complicated symbolic calculations (classical computing algorithms), but the Lisp programming language itself had a minimalist, kind-of "axiomatic" grammer that could be used to derive the rules for other grammer (what are called "macros"), including the grammer of natural languages like English. Natural language understanding systems like Terry Winnograd's "SHRDLU" was able to have a conversation in English with a human user and manipulate objects in a virtual world follwing English language commands.
And SHRDLU could carry on a conversation on a computer with literally one one-millionth the amount of computing power (memory and CPU calculation rates) that modern computers have now to run generative AI. Though obviously modern AI does it much better, the point is that even back then, the idea of talking with your computer in natural language, rather than issuing it commands in a programming language, was a problem that people were trying to solve, and were solving it pretty well giving the constraints of that time.
The Emacs software still has a version of this software built-in to it available today, just run the command "doctor" (M-x doctor<RET>). You can see for yourself what an AI agent was like back in the 1970s running on computers with extremely limited computing power. Check the appendix of this article, I created the following chat log by interacting with the Emacs doctor command the day I wrote this article.
The idea of AI first entered the mainstream back in the 1970s, after some 30 years of it being discussed among academics in the nascent days of the computing industry, going all the way back to the founders like Alan Turing, John Von Neumann. By the 1970s, computer had become a well-established industry, and people like John McCarthy and Marvin Minsky were thinking about ways to make computers do not just math, but think and reason more like a human.
But computing clock speeds at the time were measured in megahertz, a thousand times slower than today. Memory was measured in kilobytes, a million times smaller than today, long-term storage (hard disk drives) were measured in megabytes, a billion times smaller than today. How could they do AI with such a small amount of computing power?
Well it turns out, those primitive apes from the 1970s (who are our parents and grandparents, actually) had discovered ways to make "intelligent" computers which actually worked fairly well. Understanding natural human language was still out of reach at the time, but they had come up with stop-gap solutions that worked pretty often. In this article, I especially want to focus on the idea of , and how they are different, in some ways surprisingly better, in some ways obviously worse, than the what we have nowadays such as "LLM-based agentic AI" applications like OpenClaw, or coding agents that can manage a whole software project from start to finish like Claude, Copilot, or Cursor.
And software like Emacs, which is a Lisp app platform, has always had this "Agent AI" like quality to it. You tell it what you want it to do, and it gathers information from all around your computer system and from the Internet, and produces a result. But it does it with Lisp rather than with generative AI. Emacs did not die off during the first "AI winter" when the AI bubble of the late 1970s burst and all the Lisp companies went bust. But people stopped calling it AI, and as enthusiasm for AI waned, the things that people once called "intelligent" went back to being called ordinary "algorithms." And Emacs itself never became the intelligent computing systems that Lisp machines had once promised to become, it became just an ordary text editor.
I have written about this on my blog before, about how Emacs is not an editor, it is an app platform more like Node.js, or Java, or ".NET". VSCode is the latest most-popular editor in the world, and it runs on Node.js (specifically "Electron"). But having used both VSCode and Emacs, in my expert opinion, these two applications are structured very differently, in ways that are indicative of the time in which each of these software evolved. Emacs even nowadays still has that "Agentic AI" feel of the 1980s Lisp machines, like Genera Medley or the Xerox Star, whereas VSCode is designed to just build apps. I imagine soon VS Code will be adapted to become more like an "Agentic AI," with tools like "OpenClaw".
So I am genuinely excited to see that the idea of "Agentic AI" is coming back, I see this as a vindication of the people who never forgot the idealistic visions of what Lisp machines could become during the first AI bubble of the 1970s and 80s. And I hope we can put to rest the idea that you need to buy the right set of apps to solve problems, because that has never been true.
So AI is definitely a "disruptive" technology, but not in the way that personal computers (PCs) were disruptive back in 1977 when the first PCs first hit the market. And AI is not disruptive in the way that the Internet was back in 1995 when the World Wide Web (WWW) first hit mainstream consumer PC markets. AI is disruptive in how it changes us humans.
This is good and bad. It is good because we can get rid of this idea of being passive consumers of apps, and start embracing the idea of actually using computers with generative AI to improve our own set of skills that let us solve problems for other humans. It is good because the idea packaging up computing algorithms into "apps" which you can consume more easily is an idea that is going away. And possibly the idea of information property itself may be going away soon too. But it is bad, because it can be amused in more ways than we are yet able to imagine with such a new technology.
"I live in my AI agent app"
Lately, on more than one occasion, I have seen people write about how they use AI agents to do everything. They claim to live in OpenClaw or whatever agent they are using, to work on coding projects, to write comments and documentation, to publish code, to write e-mail, to write release notes, to manage project timelines, to manage teams of people, to do everything. That doesn's really bother, rather it is somewhat amusing that people are discovering all-in-one tools from which you can coordinate all of the tasks you do in your work life.
What is amusing is that I have heard people say the exact same thing about GNU Emacs (a text-based application platform), and I have heard people say the exact same thing about the Unix/Linux command line interface. I still hear people say "I live in Emacs," or "I live in the terminal." It is just now I hear, "I live in OpenClaw" as well. And here is the punchline: Emacs and the terminal can run locally (no Internet access required) on the smallest computer you can think of, such as a Raspberry Pi Zero, whereas OpenClaw uses massive, energy intensive models that must be run in a datacenter somewhere over the Internet.
For decades, computer companies have tried to create software that let people focus on their solving work-related problems without knowing anything about the details of the software engineering that is required to solve those problems. They just download an app for it, and get on with it. But companies that sell apps, especially apps for cell phones, are incentivized to reduce the amount of interoperability between apps. They want to lock you in to their app ecosystem, and you aren't as likelty to use all of their apps if their apps could just as easily integrate with competing apps.
But now that "agentic AI" agents are becoming popular, people are going to start expecting that all apps be integrated. They are going to expect the agent to be able to do everything for them. Now what may have been accomplished by multiple apps, e-mail, social networking, calendar and scheduling, banking and budgeting, content creation, file sharing, and even shopping, may start to all become integrated under a single chat-bot user interface.
It is just that computers have already been able to do this for decades, long before LLMs were invented. Though programming environments like Emacs or the Unix shell were not able to do things like use advanced statistical pattern matching to find software bugs, or translate from foreign languages, these tools were able to fully-integrate all apps under a single user interface. The difference was, the interface was not a chat bot, it was a command line, or a text file, and you would actually have to spend a few hours learning how to use these tools, rather than just communicate with them in natural language.
Natural language might be an easier user interface to use, but learning a commnad line language is still a useful skill worth learning. Learning to use Emacs is still a useful skill worth learning. And they take a fraction of the energy, and don't require Internet access, in order to use them to do your work.
For a while, it seemd like the "there's an app for that" mentality was winning out, and all-in-one integrated tools like Emacs were a thing of the past. But now with LLMs able to do more, people expect their computers to do more for them, and they expect disparate apps to all work together.
This is definitely a good thing, we just need more people to learn that you don't necessarily need an "agentic AI" to acheive this all-in-one integration. You just need Emacs. Once you get good at Emacs, then maybe integrate Claude Code into Emacs for the the things that AI Agents do better than Emacs or the command line alone, like advanced code analysis, or foreign language translation.
What is old is new again
Back in the 1970s, computer technology was rapidly becoming more powerful, "microcomputers" were invented and people started using computers in their own home for the first time in history. And people back then started thinking of how computers could be used not just for accounting or physics simulations, but to actually think and reason the way humans do. People like Marvin Minsky and John McCarthy started up research laboratories for artificial intelligence.
McCarthy invented the Lisp programming language and it quickly became the language of choice for artifical intelligence. It's high-level features abstracted away the underlying computer hardware better than other languages of the era, and allowed people to focus on solving problems using symbolic logic. Some of the first natural language systems, like ELIZA and SHRDLU, were invented at the time, and ran on computers with mere kilobytes of memory, at clock speeds of not much more than a few megahertz. Computers literally many thousands of times less powerful than modern computers were able to chat with human users, albeit in a much more limited way than what LLMs can accomplish.
All-in-one integrated Lisp environments like the Genera operating system running Medley InterLisp did back then many of the things that "agentic AI" can do nowadays, with all of the "apps" working together in harmony, rather than each in isolation. It worked well, even with the extremely limited computing hardware that was available at the time.
Now, with LLMs and agentic AI, people have re-discovered what once existed back in the 1970s and 1980s. But this time, our computers are more powerful, and our computer network, the Internet, has billions of other users to bolster the information resrouces available to us. That and the fact that computers are millions of times more powerful are the factors that makes AI agents work better than they did 40 or 50 years ago.
As a student, I reinvented the old AI of the 1970s
I remember trying to make my own AI chat bots using rules-based systems. My Git repository for it is still up, it was a JavaScript-like language where I was hoping that the abstract syntax tree (AST) of the language could serve as a universal language for representing ideas. These AST data structures could be translated to and from natural language. It was basically like what the old Lisp-based AI of the 1970s had attempted to do, but with with a JavaScript-like language. It seems ridiculous to me now, but my reasoning was that most popular languages are C-like, so it would be better to use a C-like language for representing ideas as it would translate better to and from the existing corpus of publicly available software. My computer language was to be a universal language for expressing ideas that could translate between human language and computer language.
I intended at some point to use Hidden Markov Models (HMMs) to create more sophisticated translations between code and human language. I remember tokenizing a corpus of text and counting the occurrences of the tokens. Little did I know that using a multi-layer neural network and scaling it up to a sequence of millions of tokens using would yield the results we see with modern LLMs. But I only got to the point of making the language for the chat bot before I think I finally realized that my approach was all wrong.
Conclusion
Appendix A: a chat log with the Emacs "Doctor" AI
I am the psychotherapist. Please, describe your problems. Each time
you are finished talking, type RET twice.
I need to solve a bug in my software.
Why do you say you might want to solve this bug your software?
Because the software is not behaving according to how my client
specified it should behave.
Is it because the software is not behaving according to how your
client specified it should behave that you came to me?
Yes
I see... Well, what makes you believe this is so?
I thought AI was supposed to be able to help me solve bugs?
Are you sure that that?
Are you experiencing a bug?
What do you think?
I think you are.
Earlier you said you would like to solve this bug your software?
Yes, but I don't want to have to solve a bug in you also.
Does it bother you that you do not want to have to solve a bug in me
also?
No, it bothers me that you seem to be buggy.
When did you first know that I seem to be buggy?
When you wrote "Are you sure that that?" which is nonsense and not
something a correctly functioning AI would do.
I'll ask the questions, if you don't mind!
I did not ask you a question. You see, you are definitely buggy.
Maybe your life has something to do with this.