AI Trends 2024

Power laws and agents

Dec 13, 2024

The AI Trends 2024 report is finally here — its up to date including the latest from the Neurips conference and OpenAI’s shipmas. The report has 80+ pages of data and analysis covering the macro picture and all the way down to scaling laws. Lots of proprietary analysis which you’ll find insightful. This article includes a summary of the report and screenshots of the full report. If you prefer a PDF copy: link. Or if you’d like to view it online with media & links: link.

Summary

Cover images were tastefully generated by Google’s Gemini

The story of AI in 2024 is of power law and agents. Power law because the monetary gains are concentrated in a handful of companies, even in public markets. There’s also an increasing agglomeration of capital among the few who can continue to continue the path of AI scaling laws.

Beginning with the big picture, generative AI or foundation models are the fastest adopted generational purpose technology (GPT). In 2 years, 40% of the population have used a generative AI product. It took personal computers over 15 years to get similar adoption. Smartphones was the latest GPT to get rapid adoption and it took 4 years to get to 40%.

Generative AI is also a powerful technology with major implications in how we live and work. I evaluated 16,000 tasks across 800 jobs to understand what is the automation potential of different jobs across the spectrum of generative AI technologies. Agents are the next frontier of automation and we should be ready for that.

The rapid adoption is reflected in stock market returns with the Generational AI Index returning 127% since the beginning of 2023, compared to 85% for NASDAQ and 58% for S&P. As you’ll see in the full report, 75% of the AI Index gains come from just 7 companies.

Most of the gains are also concentrated in the AI value chain, notably the chips and energy supplied to data centers. Data centers are the “currency” to play in the generative AI game. If scaling laws continue, there will continue to to be chip and energy shortages in the near to medium term.

In the private markets, there’s also herding (i.e. concentration) among venture funds in a handful of companies. Different venture pursued different strategies, reflecting different risk profiles. If the generative AI boom continues, then Sequoia, a16z, and Lightspeed stand to make generational returns for their LPs and themselves.

The pace at which the AI industry advanced in 2024 made 2023 look pale in comparison. Even now as of writing in December 13, OpenAI has 5 more days in its Shipmas series. Google just released Gemini 2.0 two days ago. Microsoft also released Phi-4.

But the entire AI boom seems to centrally hinge around scaling laws continuing. There are lots of chatter about whether we’ve faced a scaling “wall”. We have not but we are facing a more complex scaling dynamic. There’s still plenty of room to scale. In the diagram below, I sketch out what scaling laws look like today.

Even if scaling laws continue, there is the question of whether it is worth it to continue pursuing it. o1 likely costed billions to train. Is a $10 billion model worth training? What about $100 billion? These are local GDP level numbers. One way to frame the ROI question is to look at how generative AI can augment and automate jobs. There’s ~$40 trillion of wages paid each year, of which ~$11 trillion are paid in the US. A $100 billion model still sounds extreme, but at least not insanely irrational. In the full report, I provide a framework of how to think about the best opportunities for AI agents and the 12 areas with the most potential.