Summary: The State of AI 2023 report by Air Street Capital

Air Street Capital has released its 6th annual State of AI report. The 160+ page report covers trends in research, industry adoption, politics and safety of AI. It also makes predictions for what will happen in the next year (scoring 5/9 on their last year's predictions).

What do I need to know?

Research

Big labs like Open AI and Google have grown increasingly closed, releasing minimum info on their latest models. At the same time, open research has grown substantially (kickstarted by Llama models by Meta). US-based institutions are still the leading producers of research (70%+).

GPT-4 is the best LLM out there, beating evals as well as human exams. But the open-source Llama models have had 32M+ downloads in the last 30 days on Hugging Face. However, the models are picked based on vibes more than the leaderboard positions.

Small models with quality datasets have gathered a specific interest from the research community. Significantly smaller models are beating bigger models on benchmarks which could be crucial as Epoch AI predicts the end of high-quality language data in the next 2 years.

Hardware

Nvidia chips are used 19x more than all others combined in AI. Though H100 has the hype, V100 (released in 2017) was the most used GPU in 2022. Cluster building has gained interest but comes with engineering obstacles. International chip wars have forced makers to create China special chips (that comply with US sanctions).

User Behaviour:

Stackoverflow felt the heat of ChatGPT. ChatGPT saw a significant decline in traffic during the summer when students didn’t have homework. Retention for AI apps still lags behind apps like YouTube, Duolingo, and Candy Crush. Every month or two, a new trend of AI-generated images took over social media.

Other usecases: Military and Pharma are getting along better with traditional AI. Autonomous driving is expanding with the use of Gen AI. Financial institutions, consultancies, law firms and news houses have been eager to adopt AI tools.

Talent:

The Attention is all you need authors have collectively raised 870M+ aside from the $10B Open AI and Microsoft deal. All authors have left Google. Many members of the DeepSpeech2 team at Baidu have also left for other labs/started their own companies.

Regulation:

Similar to research status, previously open companies are now asking for active regulation. Multiple ideas for global governance have been proposed. EU and China have drafted unique acts for AI whereas other countries are still in the debate phase but the pace of discussion has accelerated.

Venture interest:

Without the GenAI boom, AI investments would’ve crashed by 40% versus last year. Many of the most high-profile blockbuster fundraises weren’t led by traditional VC firms at all. 2023 was the year of the corporate venture, with Big Tech putting its war chest to effective use.

You can read the blog post by Nathan (GP @Air Street Capital) or get the full report here.