0:00 / 0:00

The days of tokenmaxxing might be over, y’all Instead of using AI to build the kind of code we talk about in this video, token-heavy agentic loops are still the most broadly talked about area right now. An article came out a few months ago about token minimization, but it seemed to be more about cutting employees off from their ability to use AI at all. Plus, let’s look at a very relevant case study indicating that token minimization isn’t really happening yet: Anthropic’s recent S1 filing. About 80% of Anthropic’s revenue is billed per token rather than as flat monthly subscriptions, which includes developers building on the API and companies whose employees use Claude’s own apps, like Claude Code, on usage-based plans. Revenue thus grows with how long and how often agents run. The S1 filing shows that the top 5% of Claude Enterprise users who also use Claude Code spend about 20x more than those who don’t, net revenue retention is about 900%, and quarterly revenue went from $376M to $11.6B in six quarters. And yes, many people have been doing this for a while - certainly software engineers who are used to building software & automations. I made coding guides on how to fine-tune open models for this exact thing over 3 years ago! But this hasn’t become *mainstream* yet - I advise companies, and I see it in tech-forward startups and among the most technically advanced people in larger corporations. Again, token-heavy agentic loops are still the most broadly talked about area right now. I think this knowledge of building actual software & systems that selectively use AI will become much more mainstream — after years of encouraging people to do it! Also, if you're new here, hi! I'm Harper. I've spent a decade building AI as a Stanford computer scientist and coding AI systems at Meta & at a startup acquired by NVIDIA. I've always loved to teach, and now I get to create educational videos to share the skills, understanding, and intuition I've worked years to develop with you. For my (free) 10 Days of AI Basics series + more, DM me GUIDE and I'll send you my video walkthrough guide.

@harpercarrollai
105.0K views7.0K likes1:07ENOct 30, 2026
202 words1101 characters9 sentencesReadability: High School

Transcript

I'm a computer scientist for AI, and I think the next big thing in AI is going to be token of minimizing, not token maxing. Let me explain. Use the really smart models to write code, and then you create software systems like you did in the old days that run on its own, they don't need tokens, they have unit tests, they have other types of tests, and when those tests fail is when you use tokens, is when you debug what's going on with a really smart AI model. We're also going to see people using really smart models to design workflows where most of it is classic automation that people call AI and agents, but it's really just code running. And then at certain points, it talks to small LLMs, often fine-tuned for specific tasks. And those cost almost nothing to run. So you're choosing very specifically where you want an LLM, you're using an open source one, that's small, that is fine-tuned for your use case. And you're paying almost nothing for the whole flow. But the big giant frontier models are used to write the code initially, to debug issues, and to fortify code against cyber attack.