State of AI

What is AI?

How should one reason about AI today? Exuberance in markets naturally clouds judgement, but a rational mind can still reason their way through this market landscape and plan for the future.

As Thinking Machines and Kimi drop new models, questions resurface around distillation and moats. What justifies the valuations of AI labs if the models themselves are a commodity?

I am trying to write this in a way that retains my global optimism about the endless solutions we can build. But also recognizes some inherent difficulties at the technological layer of ai.

What was AI?

When I was at the University of Illinois, we were taught AI is only as good as the data it sits upon.

In 2026, Nvidia and the model companies are the dominant forces, which muddies the above simple fact. Is this the new normal?

What Changed?

Looking backwards over the last 7 years, there were a few key changes that account for 80% of progress.

1. The transformer was a genuinely incredible invention.
2. The insight that if you scaled these models, they continued to improve.

If you would have asked me what AI was when I was a student, I would have had a picture in my mind of machine learning and projects training a model to classify spam emails using NLP. Or deep learning at its most sophisticated. You trained the model.

The Black Box and The Bitter Lesson

Another pattern was emerging that is very important for understanding the state of technology.

Deep Learning on ECGs

When working on Nephra, deep learning was already showing signs of not only promise but domination. Prior literature on the training of models on ECG waveforms moved from human designed feature extraction towards black box deep learning methods that blew everything else out of the water. This was the bitter lesson before I knew what the bitter lesson was.

These methods were already better than human experts and they were getting better with scale. The only problem: no one understood how it was making these decisions. It was completely alien. This was received poorly in the medical industry.

Deep Learning on Stock Data

At the same time, the finance industry was changing. The same insight of scaling had proven to be invaluable for algorithmic trading firms.

I watched a video of Jim Simons talking about Renaissance Technologies, which achieved superhuman returns. A pattern was again emerging as to where the world was heading... in this interview he said that they had one rule at Renaissance,

"never interrupt the computer."

In other words, if the computer had decided it was going to make a trade that to a human seemed hair-brained, Jim deferred to the computer. He trusted the computer's judgement more than any human.

Deep Learning in Biotech

In 2023 biotech, I began to see a similar sentiment. Drug discovery was being eaten by computers.

It is humbling to watch these industries evolve. The main takeaway is the most human thing you can do is defer to and feed the machine.

Competitive Differentiation

For the ECG company, it felt like if you could train a single model to perform well, and get it through regulation, before selling it to a commercialization machine. Being first to market may have been a strong enough competitive advantage.

The Quant industry is interesting to me as an analog for AI companies. Algorithms may go out of style, but if you have a highly competitive team, you can constantly reinvent your company. There are no "moats". Your talent, data, and infrastructure is the edge. And if you find a winning algorithm, you can make so much money that the investment makes sense. But as soon as someone else finds that edge, it's worthless.

LLMs - Value and Commoditization

Depending on whether an LLM is a chat bot or the most important technology in the world determines how we should view the competitive landscape for these models.

We should evaluate model companies on a 2x2 graph of 1) how valuable they can be economically and 2) how dominant, defensible, specialized they are. How well can you maintain your advantage?

Commoditization

On the one hand, most of the world still sees these models as a chat bot. Most people can't tell the difference between a good and a slightly better chat bot. And they can be distilled easily. I could train my own reasonably good model by the end of the day or snag an open weight LLM off the shelf.

Economic Value

On the other hand, these models are incredibly valuable in the production of software. AI agents can also automate valuable tasks. Coding models have decaying delta between the best and the second best or distilled models. How valuable is that delta? That is the question model companies will have to answer over the next 5 years.

We could calculate this delta per task. Across chat bots, customer agents, coding models, mathematics, and agentic automations. Fable is a genuinely good SWE, and I will pay for it.

The bull case for the AI labs, and specifically Anthropic in my mind, is that the delta for highly valuable, complex work, coding, and long-running tool calling is going to be important as these models absorb more complex work over the next 20 years.

The bear case(s) for the labs is that the transformer, scaling, and a model's ability to output structured context and call tools was the most important ingredient set for agentic automation, and it is now open-weight and commoditized.

Data as Fossil Fuel

The learning from 2020 was that data is the determinant for a good model. That can be a problematic insight.

Ilya Sutskever has stated that internet data is the "fossil fuel of AI." And the data generated by humanity sits at roughly 181 zettabytes. If you break it down…

• Total Resource (181 Zettabytes): Every drop of digital matter on Earth.
• Drillable Reserves (3.6 Zettabytes): The public surface web.
• Refined Petroleum (50–100 Terabytes): The high-quality textual data actually used to train a top-tier LLM.

How much does synthetic data, acquisition from companies, and the work of Mercor/Scale solve this problem? I have a feeling that it simply doesn't move the needle enough.

AI is not an LLM

The exuberance around LMs has caused people to forget that LLMs are just a narrow segment of AI. "Language Models" themselves are no longer just text and multi-modal.

Where to Spend Your Resources

If you are a company or young person today, where should you spend your time? Training new industry models? Working on LLMs? Applied AI?

If you wanted to make money, the area of attack is simply the massive delta with the intelligence overhang that exists, and specifically applied Vertical AI and AI agents. I tried to stay away from doing the obvious, likely thing, but it's simply very valuable.

And timing here is really important.

Technological Explosion

Maybe this is just because I wasn't lucid during the Internet Boom, but it really doesn't seem like in all of my adult life there has been a window of opportunity close to this. In my youth we had a crypto phase and an app phase. In comparison with AI Agents, that is insanely lame isn't it?

The overhang that exists today is so massive, the opportunity is endless. As long as you solve real problems, you can't go wrong picking an industry and building AI Agents for them.

* There are 1000s of "AI Stack" visualizations that have been created. I don't like any of them. This one is overly simplistic, but it's how I think about AI today. In some sense, we shouldn't overcomplicate this stuff.

The Layers

If you wanted to start a new company today, I imagine starting a GPU company would not be a good idea. Unless you had some novel insight, foundation models around language also seem like treacherous territory. Alphafold is an example of a new type of foundation model. As optimistic as I am about new foundation models, over the next 5 years, the overhang is so massive that the best use of your time is likely simply in solving problems for companies with applied AI of this new software form-factor.

• Applications / Agent Layer
• Service Layer — new businesses, AI-enabled services, forward-deployed engineering
• Dev/Agent Tools & Frameworks — things that agents can access or use

Dev Tools

I spent some time selling AI Memory APIs to developers and frankly, I really like selling to developers. They are practical people who understand AI. But the realm of dev tools is hard to stand out in and is heavily flooded with capital. Most things in this space are pretty obvious and highly competitive.

The dev tools are pretty simple to conceptualize, and I really just think of them as building blocks for agents. An agent is kind of like a human. You need to give it tools and access to integrations, databases, observability, perhaps orchestrate teams of employees. We can build dev tools and a stack within the stack around these areas. There has been a ton of money thrown at every part of this stack, and this is not exhaustive in the least.

Going deep in memory showed me that there is much to be desired still. Niche solutions will take form from strong developers but much of this will also be absorbed into the big labs' APIs. Or developers will simply build it themselves. Middleware gets squeezed.

Applications / Agent Layer

Applications started as "AI Wrappers" for those who remember. An AI wrapper in a static state should be viewed as something that of course can be absorbed by the models.

But technology, technologists, and technology companies should always be viewed by their rate of growth and improvement. The good "AI Wrappers" became more complex over time, and expanded their scope and scale into adjacent problems, solving more and more of the complexity of a task, until they became "AI Agents" and "Vertical AI".

After working on service-related AI, I understand clearly why these have taken off first and so strongly. The sparknotes is that to solve a new problem every time with a company sucks and it is hard to keep ramping up and to develop new relationships. The economically scalable method is to find an unsolved problem that has not been solved yet, build an agent around it, and do enterprise sales.

Sounds oddly familiar to traditional B2B SaaS. It really seems like it's a repeat here.

My view is that Silicon Valley is a monoculture around tech and likely all are solving the same problems. I am very interested in the very "boring" industries that silicon valley doesn't touch. There are so many problems in the world that deserve solutions.

But selling to these companies is its own beast. We'll get to that.

Services

It is a common theme to offer services using margins of software nowadays. Or to do a forward-deployed engineering model where it's not really scalable but you find large enough bespoke problems in companies that if you solve them the model works. Fractional AI / Ode has been very successful here.

When you just do a back of the napkin calculation at the supply/demand of great engineers to solve these problems in companies, you could create a pretty big business just doing this in many industries.

I'm pretty curious about how one could buy distribution and transform the company entirely using AI-enabled rollups.