We’ve entered the 4th year of the Slop Wars. We have colorful short-hand like clanker, vibe-code, one-shot, and you’re absolutely right!. These phrases capture the zeitgeist. Emphasis on geist. It’s been more than 3 years since the release of OpenAI’s ChatGPT, which was the inciting incident that’s upended the world economy and changed how we work.

This blog post provides some miscellaneous observations on AI, how it’s being used, and how it might be used going forward. I’m writing this mostly for myself to organize my thoughts, but it might be useful to others. These ideas are mostly drawn from what I’ve seen at my company, Scholarly.

AI as Implementation Detail

Between 2023 - 2025, every interface with AI was a chat interface. LLMs are next-token predictors, and the hello world of a next-token predictor is a chat interface. Words go in, words come out. We flew past the Turing Test with hardly a wave.

But a chat interface is only one way of interacting with a next-token predictor. As context windows have grown and model quality increased, we can trust the model with more than the call-and-response staccato. We can ask it to go do things that take longer.

The chat surface might not withstand the test of time if it’s not the most appropriate tool for the task. Chat right now is similar to 3D websites, Flash, or an under construction GIF. Novel, but potentially pointless.

It ultimately comes down to this: Users don’t care if you use chat or the latest models. They care if you solve their problem. I predict the AI-in-your-face becomes muted. Certain parts of applications might feel more probabilistic because they are driven by LLMs, but still belong to the same application. You won’t be able to tell where discrete ends and probability begins.

AI as Work Accelerant

White collar jobs in general have come to the realization that software engineers have known for a few years: This is going to change how we work dramatically.

At the time of writing, the Claude Code/Codex interface seems like the sweet spot of software engineering. This is higher level than where we might have collectively been last year, where it was focused on tab completion.

As the models have gotten more capable, we’ve started to trust them with more. What used to be a novel time saver (tab completion) is now unnecessarily slow. The entire nature of software engineering has changed, with far less time with hands on the keyboard entering syntax. We’ve gained a very capable author, so much of our time is now spent reviewing, tweaking, asking it to come back with changes.

It’s no surprise that the landscape is incredibly messy. There are incomplete integrations (Why can’t I tag Claude in my Linear ticket?) and half-baked products abound. There are companies that build something compelling, only to be obviated by an Anthropic blog post. Observing this, there will be a field of dead companies in the middle with few survivors at the edges. The model makers (OpenAI, Anthropic) will stick around, with third place subjected to the power law dropoff of consumer choice (not good!). The systems of record that participate in the AI ecosystem will be rewarded handsomely, as OpenAI or Anthropic is the conduit for white collar work.

AI as Synchronous Thing-Doer

Chat has limited us to asking for tasks to be completed inline. If something was too long for the early meager context windows, it just couldn’t be done. At the time of writing, we rarely think of context windows any more.

Claude Code and things like chain-of-thought blew the doors off of the context window. By keeping small artifacts, summarizing as we go (compacting), these models are able to stay on a task for much longer through self-regulation. A strange loop?

AI as Asynchronous Thing-Doer

There’s lots of alpha in just making webhooks work (OpenAI) or providing them at all (Anthropic). As models have gotten better and we can trust them more, we need better ways of sending them off and having them come back with a work product for inspection sooner. I predict we’ll see more ways of asynchronously interacting with models. This is also what makes it feel like a coworker replacement: we’re not chit-chatting, they are going off and accomplishing a task.

Agents, skills, etc. feel like fertile ground.

Agents communicating with agents, working together to accomplish a task? I’m not sold yet, but that could be where we see the next model improvements take us.

Latency Endgame

Right now when I ask Claude to do something, talented engineers still stand a chance to complete the task before it does. These are engineers that know the codebase, have a high WPM, and know exactly what they are going to do. If we play the tape, I predict the models will eventually start competing on latency. Indeed, there are some things that we use AI for in our application where something taking 30 minutes is not particularly impressive. That same task taking 180ms, now that’s impressive.

Between Cerebras and Taalas, there are some promising options out there. As model latency decreases, this will put strange and foreign pressure on conventional hardware. Is there a future where Claude Code is operating on my machine (or a cloud VM) and idling waiting for disk I/O? What if it’s already thought through its next 10 steps, and it’s just waiting for the host to catch up?

Death of SaaS?

You should know that I’m financially incentivized to believe that SaaS is not dead as we know it. Scholarly is a SaaS company for colleges and universities. If I believed that LLMs posed an existential threat to SaaS, I should get out of this business.

There was a market swing this year with the ethos being: “AI makes it easy to create your own applications, and we believe companies will do that instead of going with a vendor. Therefore existing SaaS has become less valuable.” I think that’s dumb. Mostly because the running code that you buy is just a small piece of software. Are you telling me that an enterprise is going to vibe-upgrade their bespoke application to the latest operating system/library/underlying dependency? I don’t think so.

If anything, there are some really exciting properties about LLMs that make being a new entrant into a space great. We are unencumbered by the past with powerful tools for helping us plug in to existing systems. So it’s not that SaaS is dead, it’s SaaS that doesn’t adapt is dead. But that’s always been the case, maybe just more urgent now.

Death of White Collar?

AI is having profound impacts on white collar work. Many of the layoffs we are seeing in 2026 may be intertwined with the hiring glut of ~2020. Extrapolating my own behavior, I suspect many service aspects of white collar industries are seeing a softening in demand.

Lawyers, for example, I bet aren’t hearing from their clients as much. They are still fielding the important interactions where expertise is required, but a quick turn on a simple contract or proof-read of an NDA might just go to Opus 4.6 or GPT 5.2 for many people. Suddenly or slowly, the billable hours slip.

Nothing concrete here, just a hunch. Not sure how it plays out.

MCP

I think MCP is dumb, but it’s what we’re using. I expect it to stick around for a few years, and then we’ll fall back to more traditional REST APIs. Kind of similar to what happened with GraphQL.

Epilogue

One of my favorite podcast episodes ever is the episode on spreadsheets from Planet Money. In it, they discuss how the invention of the digital spreadsheet put bookkeepers (the keeper of paper spreadsheets) out of work. But out of that pain new work was born around financial modeling. Rather than taking a day, it took seconds to answer the question, “What if we decreased our costs by 5%?” Work in predictive financial modeling was born. An entire new job came to be, and with it many positions. Even more than bookkeepers.

We’re experiencing a similar creative destruction currently. Lots of what we knew is being destroyed before our eyes and new opportunities are being born.

Special thanks to Claude Code for suggesting a few edits to this post.