Posts

I hate to micromanage & I’ve been micromanaging AI.

A few months ago, I’d use Claude for a familiar workflow : capturing notes from a meeting, drafting a follow-up email, updating the CRM, writing the investment memo. Micromanagement at 10x speed. The agent would finish a step, then wait. I’d scan the output, type the next instruction, wait again. Prompt, response, prompt, response. I was the bottleneck in my own system.

Kirkland ibuprofen is the same molecule as Advil. Same dosage, same FDA requirements, same therapeutic effect.¹ It costs 80% less.

AI has its generic drug moment. DeepSeek V3 matches GPT-5.2 on most benchmarks.² It costs 90% less. OpenAI & Anthropic generated $22 billion in 2025.³ Chinese AI labs generated $1.8 billion.⁴ The ratio : 12:1.

Pricing explains the gap. Chinese AI API prices collapsed 90% in 2024.⁵ US frontier models average $3.38 per million input tokens. Chinese models average $0.48.

Databricks started later. It built a more complex architecture. It focused on unstructured data; images, documents, logs, audio. Though vast within the enterprise, this data had historically produced little insight. Too hard to process. Too messy to query. Too expensive to store in formats that mattered.

Snowflake took the opposite bet. Structured data. Clean tables. SQL queries that ran fast & returned answers executives could read. The market agreed. Snowflake went public at a $70 billion valuation. Databricks raised private rounds at half that.

Could you operate your company with half the people?

Jack Dorsey’s announcement yesterday¹, reducing Block’s headcount from 10,000 to 6,000, should provoke this question in every management team. The stock surged 24%. Dorsey’s memo framed it as inevitable :

Within the next year, I believe the majority of companies will reach the same conclusion & make similar structural changes. I’d rather get there honestly & on our own terms than be forced into it reactively.

I started by asking AI to do everything. Six months later, 65% of my agent’s workflow nodes run as non-AI code.

The first version was fully agentic : every task went to an LLM. LLMs would confidently progress through tasks, though not always accurately.

So I added tools to constrain what the LLM could call. Limited its ability to deviate. I added a Discovery tool to help the AI find those tools. Better, but not enough.

After writing about the SpaceX, OpenAI & Anthropic IPO liquidity problem, readers asked : what about Saudi Aramco? At $29.4b raised & a $1.7t market cap, it’s the largest IPO in history. Doesn’t it prove mega-IPOs can work?

Aramco isn’t a good proxy. The next biggest example, Alibaba at $231b, is a better parallel.

In 2019, Aramco floated 1.5% of the company. Six years later, it’s still just 2.4%. The Saudi government holds 81% directly; the sovereign wealth fund holds another 16%. The IPO served strategic goals beyond capital markets.

Every major interface shift finds its answer to meeting the needs of a massive audience in advertising : TV, web, mobile, streaming. Each is a hundreds-of-billions-of-dollars market. AI will be no different.

Today we announced Theory Ventures is leading Koah’s $20.5M Series A, with participation from our friends at Forerunner & South Park Commons.

Koah is building the monetization layer for AI applications. As AI search reaches its tipping point, assistants are becoming the primary source of information for many decisions. But AI apps face mounting pressure to monetize without compromising user experience.

We’re about to witness three of the largest IPOs in history. SpaceX is targeting $1.5t.¹ OpenAI aims for $1t.² Anthropic is valued at $380b.³ Combined, $2.9t in market cap.

The scale is unprecedented. But the real problem isn’t the market cap. It’s the float.

Typical IPOs offer 15-25% of their shares to public markets. This creates enough liquidity for price discovery while allowing founders & early investors to maintain control. Facebook floated 15%. Google floated 19%. Alibaba floated 15%.

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.