#Five takeaways from the State of AI 2024

The 2024 edition of the State of AI, an annual report on AI worldwide, came out a few weeks ago.

As every year, it’s a lot: 200 slides packed to the brim, and you’d better not be allergic to technical concepts. I could list 20 interesting takeaways, but here are my top 5:

1. Models from the “frontier labs” (OpenAI, Anthropic, Meta, Google and so on) are converging in performance. OpenAI’s lead is under threat from models like Claude 3.5 and Gemini 1.5, even if o1 has temporarily put them a step ahead again.
2. Research is speeding up on smaller, more efficient models. The goal: keep performance high while cutting the compute and energy needed by a large factor, so they can run on a phone.
3. NVIDIA is still crushing the GPU market, with 90% share. Neither the usual competitors (AMD, Intel, etc.) nor specialized startups have stopped it from becoming the most valuable company in the world ($3.5T).
4. Despite US chip restrictions, China is producing strong LLMs, at Tencent and Alibaba for instance. Alibaba’s latest Qwen release (this week) matches Claude or OpenAI o1 on code generation.
5. The 2023 talk about AI’s existential risks has gone quiet, notably since the failed coup at OpenAI. And the Trump administration will likely keep federal regulation in the US to a minimum over the next few years.

One last thing: last year the authors predicted that a song generated entirely by AI would reach the Top 50 this year. It happened, in Germany, with “Verknallt in einen Talahon”.