͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏

Forwarded this email? Subscribe here for more

🔮 Exponential View #554: Intelligence gets cheaper. Adoption gets harder. Jet engines become power plants; tiny brain chips & emotional AI++

An insider’s view of AI and exponential technologies

Azeem Azhar and Nathan Warren

Dec 14

∙

Preview

READ IN APP

You help me keep my finger on the pulse of what is happening and what may be to come. – Joanna P., a paying member
Upgrade to paid

Hi all,

Welcome to the Sunday edition.

Inside:

GPT 5.2 shows how rapidly the cost of intelligence is collapsing.
AI adoption has hit its first plateau – why breadth is flattening and where intensity will drive the next wave.
China’s H200 trap: why limited US chip access could shrink America’s compute lead and still fail to deliver leverage.
Plus: AI for teen support, a jet engine turned 42MW data center turbine, and hair‑thin brain chips streaming at ultra‑high bandwidth.

In my latest talk, I break down what AI needs to scale

Listen on Apple Podcasts or Spotify

Takeaways from OpenAI’s Code Red

Open AI’s GPT 5.2 shows how rapidly the cost of intelligence is collapsing. GDPval is a benchmark that tests professional tasks in fields like finance and healthcare. GPT 5.2 beats industry experts 70% of matchups. It does so at more than 11x the speed and at less than 1% of the cost of human experts. To get a sense of model progress, that is double the wind rate of GPT 5.1 which launched a month ago.

This latest model scores 90.5% of the Arc-AGI benchmark for $11.54, more than 390x cheaper than the previous high score set by o3 one year ago. On Arc-AGI 2, a harder test, GPT 5.2 is substantially more capable than four-month-old GPT-5.

Both GPTval and Arc-AGI’s benchmarks point to a continued acceleration in model progress.

GPT 5.2 reinforces my argument that we’re going to see plenty of model variety. Claude Opus 4.5 remains a better coding model. GPT 5.2 is especially powerful when you run it in Pro Mode, but this can take 15-20 minutes. So Gemini 3 Pro still has an important role, especially for tasks that don’t need that depth. I’ve not yet had a chance to pit Gemini 3 Pro in thinking mode against GPT 5.2 Pro. Let me know in the comments if you’ve done it and what you found!

Share Exponential View

The first plateau?

The Ramp AI Index, which tracks AI adoption through AI subscription spend, flattened in November and has effectively plateaued since July. That can look like a slowdown, but I see a transition.

Here’s what’s likely going on. Enterprise AI is already a $37 billion market growing 3.2x per year – what Menlo Ventures calls the fastest-scaling software category on record.¹ But that growth is uneven and the market’s character is shifting...