Posts

These are the top 10 posts of 2025. I like to do a roundup each year as a retrospective of how much the industry has changed through some of these posts from the beginning of the year. Some are outdated already & some presaged accelerating trends.

AI Drove the Largest New Bookings of Any New Product : In January, ServiceNow announced that their AI product was the fastest growing product in their history. Since then, many other companies have reported massive growth in their AI businesses. It was a harbinger of the tectonic shift to come.

Ten years ago, I hired an editor to grade my blog posts as an AP English teacher might. Yesterday, I asked an AI to do the same thing.

The verdict? A solid B-.

Claude’s Scorecard

Grade: B-

Strengths: The piece demonstrates a clear understanding of complex technical concepts and attempts to make them accessible to a general audience. The author’s personal connection to deep learning through graduate school studies adds authenticity and historical perspective. This human touch helps ground the technical material in lived experience.

In the bustling tech campuses of 2024, the age of passive AI – systems that merely respond to our queries – is giving way to something far more profound: the era of AI agents.

As we look to 2025, we’re about to discover what happens when algorithms learn to act.

At the heart of these emerging agents lies a trinity of learning approaches :

• supervised learning : like reading a book to a child, humans provide clear guidance to AI labeling cat & dog, sheep & cow.
• unsupervised learning : AI discovers hidden patterns in data ; an ecommerce site recommends products you might like by clustering similar users’ purchasing patterns.
• reinforcement learning : an AI learns to play a video game by playing thousands of times, just the way a gamer might.

Deep learning means using the neural networks architecture to calculate an answer like what will the weather be tomorrow or summarize the Knicks game last night.

Today, we’re announcing our second fund of $450m to support our mission of partnering with early stage software companies that leverage technology discontinuities into go-to-market advantages. This marks the next chapter in our firm’s evolution.

Since we launched Theory in 2023, we’ve gathered a wonderful team : Lauren, Spencer, Andy, Rafa, Amber, Arjun, & Kristin.

We have partnered with 8 marvelous founding teams, all using data to power the next wave of innovation across the Modern Data Stack, Artificial Intelligence, & Web3.

I didn’t notice it at first but there in the back corner of my laptop, I’ve been assembling a little library.

The library doesn’t contain books, but scraps of text that explain what an AI should do.

Visit Anthropic’s home page & you’ll find their collection :

Each of us will assemble these little libraries with tracts like :

• write a blog post in my voice about this topic
• collect action items to send to the team from team notes & format them in this particular way
• analyze the P&L for the finance team paying particular attention to gross margin
• write a performance review using the company’s stated evaluation rubric for my teammates
• summarize a competitor’s webpage every month to detect changes in positioning

As we work in a role, we have assemble these workflows and kept them to ourselves.

What it cost to have an assistant with you like in the movie Her?

The cost of using AI has dropped precipitously, an order of magnitude every year.

If the average American picks up their phone 144 times per day & engages with an assistant, each time for four interactions every day of a month, an assistant like Her would cost about 78 cents in inference cost.¹

I’m not taking into account any of the additional costs associated with delivering such a product. Assuming a commercial vendor would 10x the price, it’s still $7 per month, less than half of a Netflix subscription.

77% of enterprise AI usage are using models that are small models, less than 13b parameters.

Databricks, in their annual State of Data + AI report, published this survey which among other interesting findings indicated that large models, those with 100 billion perimeters or more now represent about 15% of implementations.

In August, we asked enterprise buyers What Has Your GPU Done for You Today? They expressed concern with the ROI of using some of the larger models, particularly in production applications.

After the election, the public markets have roared, but not equally.

The broad software ecosystem has seen a relatively muted change in forward multiples. There’s no statistically significant change in the days after the election compared to the month before.

On the other hand, crypto’s top tokens have seen tremendous appreciation. Bitcoin is up 48% ; Solana up 70% ; & SUI up 324% in the few days since the announcenment. All of those ar statistically significant changes.

Over the weekend, I found myself in an hour-long conversation during my drive with an AI.

We jumped from discussing Cooper Flagg’s basketball stats at Duke to comparing Carlo Rovelli and Brian Greene’s competing theories of physics to talking about the history of San Francisco. In a daring feat of economic analysis, I asked it to calculate if the after-tax returns of two ETFs were statistically significant & to compare the energy portfolios using the 13-Fs of a few hedge funds for investment ideas.

The major areas of AI innovation automate white-collar work. Reviewing the BLS’ data on employment for white collar work, I aggregated the data to these categories.

It’s striking that most of them already have a significant number of AI startups pursuing their ambitions to change workflows.

Software engineers were the first to benefit with Copilot. Today, there’s a panoply of different kinds of AI software for developers, including test generation, code refactoring, code generation, & security analysis.