Old Skin, New Guts

The first installment of Five Things You Should Know About the State of AI in 2026

Thank you for being here. Please consider supporting my work by sharing my writing with a friend or taking out a paid subscription.

---

This week I gave an informal talk about AI to a group of my colleagues at Harvey Mudd. I thought you all might find what I had to say useful, so I decided to write up a version of what I shared. To make it more digestible (and contrary to my general spontaneous style here I must add!), I am planning to break it up into five parts presented serially over the next few weeks.

Here are the five ideas I shared with them and that I am planning to write about over the next month or so.

1. Old skin, new guts

2. It’s all about context, baby (and by extension tools)

3. The future is local

4. Agents are just fancy while loops

5. Software is (still) underrated

This week we’ll tackle the first item on the list. Let’s dig in!

---

An electrified ‘66 Mustang. Image credit calimotivee.

When Apple releases a new iPhone, they always change the physical design just enough so that you can tell the new generation apart from the past generation. They want to deliver something that is the optimal combination of familiar and new. While they want to continue to build on the brand equity of the iPhone with something that looks familiar, there is almost always a new feature (most obviously in the form of the changing camera bump) that is a clear distinguishing mark to let everyone know that you’ve got the latest and greatest.

However, this same strategy does not apply to the AI apps that have become part and parcel of many of our lives over the past few years. If you look at side-by-side screenshots of ChatGPT or Claude from a few years ago and compare them to screenshots of the tools today they are almost identical. The standard UI/UX of the chatbot messaging interface, at least the one that most of us interact with, has been almost unchanged over four years. No newfangled camera bumps to be found.

And yet, looks are deceiving.

While the ChatGPT app on your phone or computer might look nearly identical in 2026 as it did in 2022, what is beneath the interface has changed significantly. It might sound trite, but one of the most important things you should know about the state of AI in 2026 is that it is not the AI of 2022, despite the fact that it looks largely the same on the outside.

It’s like a ‘66 Mustang skin with Tesla guts

As I was preparing to give this talk, I was reminded of a story I had run across a while back about someone who had replaced the guts of a beautiful 1966 Ford Mustang with those of a Tesla Model 3. What’s been happening with consumer-facing AI tools is essentially equivalent to this. The skin looks like ChatGPT of 2022 but the insides bear almost no similarity.

When ChatGPT was first released, you were essentially interacting directly (and exclusively) with the LLM (GPT-3.5 at the time). You were sending in a prompt, getting the probabilistic continuation of that prompt with what to our eyes today was a very simple wrapper around it (essentially just helping the system know when to stop and return the response to the user).

How quaint by 2026 standards. While your interaction with AI tools is mostly the same—type a prompt, wait for a response, rinse and repeat—what’s going on under the hood is not at all the same. Instead of directly interacting with the LLM itself, you are interacting with the agentic harness (more on that later). The LLM is still an important part of the harness, but only a part of it.

Pay attention to the harness

All the fuss about so-called “agentic” AI is really about the harness. When you hear about agentic AI, you should think LLM + harness. While the based models continue to get better and better, getting the harness right has turned out to be arguably even more important for the utility of the model.

The AI harness is a piece of traditional software that surrounds the LLM and helps to determine what the inputs to the LLM are. While the LLM sits at the center of the harness and is the secret sauce that makes these tools work as well as they do, it is only part of what makes the whole system work. The harness surrounds the LLM with a set of additional components which are well outlined in this blog post from LangChain. Here’s a lightly edited list of what they present there. An agentic AI harness includes:

• System prompts

• Tools, Skills, Model Context Protocol (MCPs) Servers

• Bundled infrastructure like a filesystem, sandbox, and a web browser

• Orchestration logic for doing things like spawning subagents and handing off/routing tasks to different models

• Hooks and middleware for deterministic execution of various functions like context window compaction, continuation, and lint checks

Taken together this means that what is happening when you type a prompt into your favorite AI tool today is much more sophisticated than what happened a few years ago. If you pay attention, you can see it happening.

For instance, one of the first things your AI tool will often do when responding to a prompt is pull in information from a web search. What an obviously smart thing to do. If you are asked a specific factual question about information that could be found on the web it is much better to go out and try to pull web pages that have that information and put them in context for the model to work with rather than try and get at the same information through the trained weights of the model. This one architectural change plays a big role in making the responses more accurate.

To consider a specific example, consider how AI tools answer mathematical questions. Next token prediction, it turns out, is a not very good or reliable way to tackle this problem. Using this strategy, you are depending on the likelihood of certain tokens following after one another.

This might work well enough for some expressions which are well-represented in the training data. For instance, if you ask the model to compute 2 + 2, it’s highly likely that there are plenty of examples in the training data that show that the value of this expression is 4. However, the answer to something much more complicated like 987654321987654321 × 123456789123456789 is much less likely to be represented well in the training data.

But LLMs are a one trick pony. Tokens in, tokens out. So what is one to do? The same thing that you would do. Use a calculator.

This is increasingly what the tools are doing. Instead of trying to do the computation directly using next-token prediction, which is unreliable for lots of things, LLMs are instead using their very useful ability to flexibly interpolate from prompts to continuations in order to figure out what the goal of the prompt is and then make a good suggestion of a deterministic tool to use to get the job done. For example, by writing a small Python script to compute the result of 987654321987654321 × 123456789123456789 (which is 121,932,631,356,500,531,347,203,169,112,635,269 for the curious among you).

So what?

There are many things that could be said here, but here are two things that I would suggest you take away.

1. It’s worth updating your priors about generative AI tools and what they are capable of, especially if you haven’t had the chance to play around with some of the latest models (e.g., Anthropic’s Opus 4.5 and later). These models, which are being aggressively optimized both at the model level and in the harness around the model have made these tools much more reliable.

2. When you are using an AI tool, you should think at least as much about what tools you are giving it access to as you think about what model you use. A decent model with access to the right tools to pull relevant information into context will often beat a fancier (and more expensive) model without access to the right tools.

Next time we’ll talk more about context and tools. It’s a natural followup to this conversation. Stay tuned, and share and subscribe, as they say. See ya next week.

Got a thought? Leave a comment below.

Reading Recommendations

Grateful to have my inbox full of more L. M. Sacasas goodness lately. He’s published a few posts lately, all of which are worth your time and attention (as usual).

Feeding on Illusions

Your AI Is Not a Tool

---

The Book Nook

Inspired by Mike Sacasas’s post on Ursula Le Guin, I decided to read her short story "The Ones Who Walk Away from Omelas." I get the sense that this is one of the most widely read of Le Guin’s stories and it’s a good one. The short story genre is such a lovely one and the length and provocative nature of Le Guin’s story reminded me of Flannery O’Connor. Worth a read if you haven’t yet read it (along with the foreword from Le Guin about the origin of the name Omelas).

---

The Professor Is In

I am continuing to have a fun time getting ready for the fall semester, specifically as I prepare for the new AI-Amplified Prototyping class that I am co-teaching. The basic thesis for the class is getting sharper and sharper (perhaps we’ll actually have a syllabus by day one!) and converging to something like “build an AI-powered app from the ground up, understanding how it works from 1s and 0s, through neural network and GPTs, all the way to HTML, Javascript, and databases with no black boxes”.

Super fun class to build and teach. I’m excited about it!

---

Leisure Line

Got back on the bike this week for a few morning rides. Of course on my first trip out I managed to pick up a screw in my back tire. So it goes. Luckily I had an extra inner tube and a CO2 cylinder to get me back up and running without too much time lost.

---

Still Life

Another beautiful lizard discovered this week in the garden. That’s my index finger for scale!