The end of 2025 was filled with a certain amount of Claude Code mania, and I am inclined to agree to a point. The title of this post is Future Shock, from Toffler’s work describing a society coping with an extreme rate of change. Claude Code with Opus 4.5 is what I consider to be my first encounter with real artificial intelligence. The combination of a frontier model with a terminal harness creating an independent actor. Something that I suspect is just the tip of the iceberg. It impressed me enough that I wanted to collect my thoughts, and put this post out for those who are either dismissive or unaware.

I have found LLMs to have a moderate degree of utility since Claude 3.5. June 2024 seems a very long time ago now. Before that models were fascinating toys, but net negative time on productive tasks. The negativity emerging from failures of basic logic and syntax that required more fixing than saving, despite attempting an impression of utility. Claude 3.5 marked the point where it could generally get syntax and library choices correct enough to be interesting.

Syntax, especially for me, is of interest because I tend to do a very wide variety of work, often with different languages, platforms, and unique environments. Because of this a syntax reference engine is highly useful for me, even as a try once tool before returning to traditional processes. Useful, but hard to call intelligent. Testing prior agentic coding experiences, even more modern ones felt similar. It could write specific functions that I could then test and iterate with. The design of those functions, let alone the overall system, not being remotely within scope of those llms. Again, this is not to be dismissive of where these models were. Leveraged properly they could provide meaningful speed up with controllable downsides. You still reviewed and understood the codebase in the same way because of the limited potential for the model’s extemporaneous efforts.

Claude Code with Opus 4.5 marked a significant change. Effectively Opus interacts directly with the command line and generates code, executes it, reviews logs and outputs, fixes errors, and conducts tests. It does this somewhat cautiously, as a new tool should, with some human prodding, but it really can do a complete loop without human thought. Opus 4.5 is OODA capable. Observe, orient, decide, act. It takes a broad instruction, finds a path to a solution, and executes the solution. The execution feedback it receives prompting more loops until it reaches a satisfactory state. If you’ve ever done any engineering, you know just how much of problem solving was just covered. Not only that, in well documented and established software tooling, I would rank it as at least an 80th percentile engineer (global), and in terms of speed significantly faster. If we aren’t going to call that “intelligent” we are going to have to have a serious talk about how many humans qualify.

It isn’t all sunshine and rainbows of course. Opus 4.5 still struggles significantly with the overall implementation and construction strategy. I don’t actually blame Opus for this, it’s a difficult and poorly documented problem. One of the current problems I am working on is a better form of distributed execution environment. Good design produces an elegant simplicity that is modular, sustainable, and minimal in weight. As I gathered Claude’s opinion, the original suggestions would have been none of those things. Not exclusive to Claude, this held true for other frontier models. The essential issue being that although the model suggests a solution path that would work, it is taking you up the creek without a paddle as the solution gets larger and more bloated and more dependent on AI to manage the inherent problems. It also has an odd tendency to vary wildly in the “enterprise” nature of its implementation. Heaven forbid the solution isn’t ready for a million users, but authentication or parameterization/reusability often needs specific and repeated instruction. Again, in defense of Claude these aren’t easy problems and it’s not as though most people would succeed in the face of those problems.

Claude Code is also rather expensive for non-corporate use cases, and I’m not a tremendous fan of how the terminal environment is managed. Fortunately the solution space is just opening up. Opencode will provide a similar terminal experience with free models, albeit in what feels like a severely crippled state as of the time of writing. General theory holds that open models lag by 6 months to a year, so you may be able to have a similar but significantly cheaper experience within a year, likely with better features and integrations.

Overall I would liken the experience to riding a horse for the first time. LLMs have already been an effective bicycle. With limited independent agency they can provide a speed boost in specific situations. The horse may not even be faster, and may be ill suited to many situations, but unlike the bicycle, the horse doesn’t need you to move. That’s the future shock. It doesn’t have to be a magic do or know everything box. Something that is capable of independent problem solving at a level better than most people is going to create a very different future. Not super intelligence, but commodity intelligence. The improvements to the models mattering less than improvements to the surrounding frameworks and harnesses.

At one of our dinners, Milton recalled traveling to an Asian country in the 1960s and visiting a worksite where a new canal was being built. He was shocked to see that, instead of modern tractors and earth movers, the workers had shovels. He asked why there were so few machines. The government bureaucrat explained: “You don’t understand. This is a jobs program.” To which Milton replied: “Oh, I thought you were trying to build a canal. If it’s jobs you want, then you should give these workers spoons, not shovels.”

  • Apocryphal Friedman quote from the WSJ.

I don’t pretend to have perfect knowledge of what will happen when we have median specialized human on demand, but I expect we aren’t going back to spoons. I think humans will remain relevant for some time, perhaps indefinitely, and while they are the ability to shepard these new intelligences will be an absolute premium skill. Welcome to the future.