Posts

I’ve spent the last year building AI agent systems. Here are nine observations.

1. Prototype with the Best

When the input is unpredictable, email parsing, voice transcription, messy data extraction, reach for state of the art. Figure out what works with the best models, then specialize them over time.

2. Polish Small Gems

I fine-tuned Qwen 3 for task classification using rLLM¹. The 8B model beats GPT 5.2 zero-shot prompting & runs locally on my laptop. Fine-tuning shines when the task is well-defined & the input distribution is stable.

For the first time in venture history, three distinct channels share the liquidity burden roughly equally.

A decade ago, secondaries barely registered. They accounted for roughly 3% of exit value in 2015. Today they claim 31% : nearly $95b in the trailing twelve months.

Last quarter, my AI inference costs hit $100,000 annualized.

I started small. Six months earlier, I was spending $200 a month on Claude. Then I added three agent subscriptions : Codex, Gemini, & Claude Code. I was paying $600 a month.

Next I started using AI to transform my todo list into my done list, increasing tasks to 31 per day. $92 daily inference invoices started arriving. Then $400 per month on browser agents.

Scale AI sold for $14.8 billion.¹ Character.AI for $2.5 billion.² Inflection for $650 million.³ They represent a tiny fraction of the market.

From 2020 to 2025, there were 5,700 AI & ML acquisitions. Only 21% disclosed a deal value. The remaining 4,500 share a pattern : small teams absorbed into larger companies. No fanfare. No valuation headlines. Just talent migrating from startup to incumbent.

Energy is up 17% this year. Materials 16.5%. Industrials 12%. Technology is flat.

Over the past month, $3.25 billion moved into XLE (energy) while $1.66 billion left XLK (tech).¹

The logic isn’t obvious until you look at operating leverage.

In AI, distribution is king. Skills are seizing the crown.

Skills are programs written in English. They tell an agent how to accomplish a task : which APIs to call, what format to use, how to handle edge cases. A skill transforms an agent from a conversationalist into an operator.

Remember Trinity in The Matrix? “Can you fly that thing?” Neo asks. “Not yet,” she says. Seconds later, Tank uploads a B-212 helicopter pilot program directly into her mind. She steps into the cockpit & flies.

Vertical software has fallen 43% this year. DevTools, just 21%. The gap between them, twenty-two percentage points, tells you what markets actually believe about AI.

As Michael Mauboussin argues in Expectations Investing, prices contain information about what markets believe will happen.

The conventional interpretation is that investors fear AI will replace certain categories of software. But that explanation misses something important.

Google’s Q4 2025 earnings call revealed a company in the midst of a spectacular AI acceleration.

“Our first-party models like Gemini now process over 10 billion tokens per minute via direct API used by our customers, up from 7 billion last quarter.”

This represents a staggering 52x increase year-over-year, up from ~8.3 trillion tokens/month in December 2024 to an annualized run rate of over 430 trillion.

Leveraged software companies running on leveraged infrastructure. When AI compresses software revenue, the stress doesn’t stop at equity. It cascades into debt.

BDC assets hit $475 billion in Q1 2025.¹ Software comprises 23% of Ares Capital, the largest BDC.²

Shares of Blue Owl, Ares, & KKR dropped 9%+ on Tuesday. UBS estimates 35% of BDC portfolios face AI disruption.³

Meta’s first internal AI agent went from zero to thousands of engineers using it daily. That doesn’t happen by accident. On Tuesday, February 25th at 5:30 PM Pacific, the person who built it, Jim Everingham, will explain how.

Jim is the CEO & co-founder of Guild.ai. Previously, he led Meta’s developer infrastructure organization & was responsible for building Meta’s first internal AI agent, work that moved from experimentation to real adoption across engineering teams.