Posts

Two years ago, the idea of useful AI on your phone was fantastical. Siri couldn’t finish a sentence. Local models hallucinated nonsense.

Last week, Google released Gemma 4 E4B¹, a free model that matches GPT-4o and runs entirely on your phone.²

The next few weeks promise even more advanced pocket models. The market expects new releases from DeepSeek³, Qwen⁴, Kimi⁵ & Minimax⁶.

Frontier models don’t stay frontier for long. Within three to four months, you can run a model with similar performance on your laptop; 23 months later, you can run the same model on your phone.

Two days ago, I burnt 250 million tokens in a single day.

That’s up 20x in six weeks. This idea, called tokenmaxxing, is the deliberate practice of maximizing token consumption. The question : how much electricity can we turn into useful work?

The secret is parallelization. Structure a plan at the start of the day that allows multiple agents to work simultaneously. METR research shows the latest models can now work autonomously for 12 hours, up from 1 hour a year ago. Here’s the ramp once I started implementing a daily plan :

On April 9th at 10:00 AM PDT, Lena Waters will kick off a new version of Office Hours.

Lena led marketing at Notion, Grammarly, & DocuSign. At Notion, she was CMO during the company’s AI product transition. She guided the shift from product-led growth to enterprise expansion while the company deepened its position in AI-powered work. At Grammarly, she oversaw marketing as the writing assistant added AI features. At DocuSign, she managed enterprise go-to-market strategy.

Jevon & Veblen walk into a data center.

The dominant motif around AI has been Jevon’s Paradox¹ : the cheaper a product becomes, the more it is consumed.

Token prices dropped 10-20x over the past 18 months & demand exploded in response.

Anthropic surged past $19 billion in run-rate last month, up from $9 billion at the end of 2025.² OpenAI topped $25 billion in annualized revenue in February, a 17% increase in two months.³

Would you choose one software over another because it has a proprietary model with better performance?

Two companies shipped custom AI models today (three in a week counting Cursor!¹), raising that question. Intercom launched Apex 1.0, a model for answering customer support tickets.² Chroma released Context-1, a model for multi-hop agent search.³

Apex 1.0 beats GPT-5.4 & Claude Opus 4.5 on customer service tasks.² Context-1 scores 97% on agent search benchmarks.³ One Intercom gaming customer saw resolution rates jump from 68% to 75%.²

The SaaS era was defined by unbundling : find a workflow, optimize it, own it. Salesforce chose sales automation. Slack chose chat. Dropbox chose file sharing. Point solutions won by perfecting single workflows. The playbook : own one pain point, expand from there.

AI is moving faster than anyone predicted. When models change every 42 days, buyers can’t assemble a best-of-breed stack. They want a platform they can trust for three to five years.

Last week, Cursor launched Composer 2 to over one million daily active users.¹ Within hours, a developer discovered Cursor had built its flagship model on top of Moonshot AI’s Kimi K2.5, a Chinese open-source model.²

Moonshot AI’s response? “This is the open model ecosystem we love to support.”³

Cursor’s model is at near parity with state-of-the-art at one-eighth the price.⁴ It’s also no coincidence the editor powering Cursor is open-source, VS Code.

In 2025, we predicted that 2026 would be the year agents would earn as much as a person.

It’s already happening.

In markets where there’s a labor shortage and an urgent need to hire people, we are seeing agents command 75%, 85%, even 100% of a human equivalent salary. This is faster than we were anticipating.

The first-order benefit is completing the work.

But there are second-order benefits that are now starting to appear. Training agents is significantly faster since all materials can be presented at once & in parallel to the AI.

I set up a race today between two robots.

My Mac on the left vs Claude Code on the right. Both tasked with building a payment app on Stripe’s new Tempo blockchain. Same prompts, same task, side by side.

Opus 4.5 is about 20% smarter than Qwen 35B on benchmarks. And it’s likely 50x larger. The hare should have won. It didn’t.

The local model finished in 2 minutes. Claude took over 6. I asked Claude to score both outputs : local model 6.5, Claude 4.5.¹

For every dollar hyperscalers earn from AI today, they’re spending twelve dollars to build more capacity.¹ That’s the bet embedded in $575 billion of capital expenditure this year.²

How fast does AI revenue need to grow to pay back this data center mortgage?

From 2020 to 2024, hyperscalers issued an average of $20 billion in bonds annually.³ In 2025, that jumped to $96 billion. In 2026, it will reach $159 billion.³ Morgan Stanley projects $1.5 trillion over the next few years.⁴