“Ten years ago, the world's best artificial intelligence systems were unable to classify images at a human level. They didn't understand language, struggled to reason visually, and failed even the most common reading comprehension tests. “Today, AI systems routinely outperform human performance on standard benchmarks.” This is the observation made by the HAI Institute (Institute for Human-Centered Artificial Intelligence) at Stanford University.
In the 2024 edition of its AI Index report – which contains more than 300 pages tracing global AI trends in 2023 – the Institute draws several conclusions, starting with established progress in artificial intelligence, but also trends in research and development, performance of large language models or the impact of this technology on the world.
Advertisement
2023, the year of technology acceleration
“Although AI has already shown exciting new capabilities in 2022, the technology has accelerated in 2023,” notes the HAI Institute in its annual study. Previously known AI models were capable of processing images but not generating text, and vice versa. Today, recent models like Gemini, GPT-4 and Claude-3 are reversing the trend, revealing impressive multimodal capabilities: they can generate fluent texts in dozens of languages, process sounds and even explain “memes”.
“This year we are seeing more and more models that can work in multiple areas,” said Vanessa Parli, director of research programs at Stanford HAI. “One of the aspects of AI research that I find most interesting is the combination of these large language models with robotics or autonomous agents, which marks an important step in making robots work more efficiently around the world real”.
Open source is popular but remains far behind proprietary models
Last year, organizations (public and private) published 149 foundation models, more than double the number published in 2022. Of these published models, 65.7% were open source (and therefore freely used and modified by n anyone), compared to only 44.4% in 2022 and 33.3% in 2021.
So-called closed models always outperform their open source counterparts. Across 10 selected benchmarks, closed models achieved a median performance advantage of 24.2%, with differences ranging from as little as 4.0% on math tasks like GSM8K to as much as 317.7%. on collaboration agent tasks like AgentBench.
Advertisement
Money, a growth engine for AI
To get to this stage of development, one of the keys to success is – unsurprisingly – money. Because it is through massive financing that the technological giants have made business leaders understand the interest of such technology. Attention that has gradually transformed into global private investments in generative AI. The report notes that they have skyrocketed from around $3 billion in 2022 to $25.2 billion in 2023 and around 30 times the amount in 2019 (the HAI Institute calls this the ChatGPT effect) .
Generative AI accounted for more than a quarter of all AI-related private investments in 2023. Additionally, nearly 80% of Fortune 500 earnings calls mentioned AI, more than ever before. Additionally, strong evidence from academic economists suggests that AI tangibly boosts employee productivity and can particularly help low-skilled workers.
Tech giants are ahead of academia
The industry also remains ahead in the field of AI. In 2023, technology companies produced 51 significant machine learning systems, while universities produced only 15. Additionally, while 108 newly released foundation models came from industry, only 28 came from academia .
In this regard, Google is also considered the leader by launching the largest number of models, notably Gemini and RT-2, last year.
In fact, since 2019, the Mountain View firm has led the release of the most foundation models, with a total of 40, followed by OpenAI with 20. Academia follows industry: the year Last year, UC Berkeley published three models and Stanford two.
AI becomes more expensive as cutting-edge models emerge
However, if AI is spreading at breakneck speed, one point is becoming more and more significant: the costs of training models. First reported in last year's AI Index report, they have continued to rise. New estimates suggest that some of the newest systems, like OpenAI's GPT-4, cost $78 million to train.
The price of Google Gemini amounts to $191 million. For comparison, some cutting-edge models released around five years ago, namely the original Transformer model (2017) and RoBERTa Large (2019), cost around $900 and $160,000 to drive, respectively.
A concentration on the American continent
And this translates into concentration at the geographical level. The technology companies capable of ensuring such costs are American and invest massively on their soil. Thus, in 2023, a significantly greater number of AI models (61) came from institutions based in the United States, compared to the European Union (21) and the China (15).
The United States also remains the leading location for investment in AI. A total of $67.2 billion was invested in AI in the United States last year, almost nine times more than the amount invested in China.
Of course, the Middle Kingdom remains the biggest competitor of Uncle Sam's land, and can count on two major assets: the number of robotic installations and the quantity of global patents in AI mainly (61 %) from the country.
Do you want to stay up to date on the latest news in the artificial intelligence sector? Register for free to the IA Insider newsletter.
Selected for you
