logotyp inwedo

The State of AI in 2023: Takeaways from Stanford’s AI Index Report

Stanford University has released a massive 386-page report on the state of artificial intelligence that is packed with valuable data and information. I’ve taken the time to analyze the entire piece and have included some noteworthy charts and comments below for your convenience. Let’s take a look at the details!



Publications on artificial intelligence globally from 2010 to 2021

The total number of publications on artificial intelligence more than doubled between 2010 and 2021, rising from 200,000 to almost 500,000.


According to the report, the education sector has the most significant number of publications on artificial intelligence. Looking deeper, however, you can see that the number in this area has been declining slightly from year to year in each region since 2010.

In the industry/business sector, the United States ranks first in the number of scientific publications. The European Union holds the second position on the list.


Since 2010, the Chinese Academy of Sciences has consistently held the leading position with the highest number of publications on AI. The next four institutions on the list are also Chinese universities:

  • Tsinghua University (5099 publications),
  • Chinese Academy of Sciences (3373 publications),
  • Shanghai Jiao Tong University (2904 publications),
  • and Zhejiang University (2703 publications).

As you can see from these charts, China places great importance on developing artificial intelligence, especially for use by the government.

Machine learning in artificial intelligence

The following chart illustrates the number of parameters of the machine learning systems.

The number of parameters in AI systems has increased steadily over time, reflecting the increased complexity of the tasks that the software is able to process and perform.

This trend had become particularly rapid since the 2010s, when there was the greatest progress in the area of utilized hardware and its capabilities.


The timeline below illustrates the release dates of major language and multimodal models along with the country of origin. It begins with the GPT-2 release in 2019.

Notable large language and multimodal models released in 2022 are:

  • DALL-E 2 by OpenAI,
  • and PaLM by Google.

The only large Chinese model released in 2022 was GLM-130B. This bilingual model, capable of processing both English and Chinese, was developed by Tsinghua University researchers.

Notably, another model, BLOOM, launched towards the end of 2022, is classified as indeterminate since it was the outcome of a collaboration involving over 1,000 international researchers.


Additionally, it’s worth taking a look at the cost of training language models. While AI companies seldom speak openly about it, it is widely speculated that the expense could reach millions of dollars and would only increase with larger scales.

The report’s creators estimated the cost based on the equipment used, and the training time disclosed by the models’ authors.


There’s also a link between the size and cost of large language and multimodal models. As you can see below, models with a more significant number of parameters that require more computing resources during training typically carry higher expenses.


The use of language models in planning and reasoning

In 2022, researchers designed a more challenging planning and reasoning test for large language models, which consisted of seven distinct tasks:

  1. plan generation,
  2. cost-optimal planning,
  3. reasoning about plan execution,
  4. robustness to goal reformulation,
  5. ability to reuse plans,
  6. replanning,
  7. plan generalization.

It turned out that large language models performed relatively ineffective. Although GPT-3, Instruct-GPT3, and BLOOM were able to reformulate objectives robustly in certain contexts, they encountered difficulties with tasks such as plan generation, optimal planning, and plan reuse.

The large language models’ performance was considerably inferior to that of humans, indicating their capability but lack of human-like reasoning abilities.


Impact of artificial intelligence on environmental pollution

Assessing the impact of artificial intelligence on the environment is a critical aspect, and researchers have delved into this area, too. To quantify the amount of carbon dioxide released into the atmosphere, it was crucial to identify the underlying factors that contribute to this.

It turned out that these are:

  • the number of parameters in the model,
  • the power usage effectiveness of a data center,
  • the grid carbon intensity.

Of the four language models tested, GPT3 was responsible for the highest carbon emissions, generating 1.4 times more carbon than Gopher, 7.2 times more than OPT, and 20.1 times more than BLOOM.

The latter produced carbon emissions that were 1.4 times higher than the yearly average consumption of an American and 25 times greater than the emissions generated by a round-trip flight from New York to San Francisco for one passenger. Furthermore, the energy used during BLOOM’s training was sufficient to power an average American household for 41 years.

Incidents and controversies involving artificial intelligence

The AIAAIC database, which monitors cases of unethical use of AI, shows that in 2021, the number of reported incidents and controversies concerning the use of AI increased by 26 times compared to 2012.


The upsurge in reported incidents can be attributed to the growing integration of AI with the real world and increased awareness of potential ethical abuses. Furthermore, the ability to monitor incidents and evaluate the harm caused using AI has significantly improved.


A notable example of an AI incident is a deepfake video that emerged in March 2022, featuring the Ukrainian President proclaiming the surrender of Ukraine. Such misuse of artificial intelligence carries a tremendous risk that several countries may not be equipped to manage.

Another significant danger is the absence of regulations concerning the use of AI around the world. Despite a noticeable increase between 2016 and 2022, examining the figures from Europe reveals several countries where the count remains zero.


The following data shows the ten most demanded AI-related specialized skills listed in job postings from 2010/12 to 2022. From a high-level view, each particular skill is currently in higher demand than ten years ago.


It’s also worth looking at a chart showing how GitHub’s Copilot – one of the first tools to use artificial intelligence to help engineers with their daily work – influenced the productivity and well-being of developers:

  • 96% of developers handled repetitive tasks faster,
  • 88% were feeling more productive,
  • 74% could focus on more satisfying work,
  • 59% felt less frustrated when coding.

Investments in AI

Keeping tabs on investments associated with artificial intelligence has become progressively more crucial as AI becomes more integrated into the economy. Funding for AI projects has been steadily rising since 2013 until it experienced a slight drop in 2022.


In 2022, global investments in AI amounted to $189.6 billion, roughly a third less than in 2021. The data below illustrates that most funding came from the United States, China, and the United Kingdom.


Use of AI in business

Industries with the highest adoption rates of process automation were:

  • Technology and Telecommunications: 48%
  • Financial and Business Services: 47%
  • Legal and Professional Services: 46%

The most commonly utilized technologies related to artificial intelligence were:

  • Process Automation: 39%
  • Computer Vision: 34%
  • NL Text Understanding: 33%
  • Virtual Agents: 33%

Regardless of industry, companies have observed that incorporating artificial intelligence in their operations reduces costs and increases revenue.

The most substantial reductions in cash outflows were seen in:

  • Supply chain management: 52%
  • Service operations: 45%
  • Strategy and finance: 43%
  • Risk management: 43%

In terms of revenue growth, the most substantial increases were observed in the following departments:

  • Marketing and sales: 70%
  • Product and/or service development: 70%
  • Strategy and finance: 65%

Public opinion on the use of artificial intelligence

Lastly, let’s look at the data available on the public’s perception of artificial intelligence. According to the survey, 60% of the participants believe that AI products and services will significantly alter their daily routines in the near future, making their lives easier.

A slim majority, 52%, believe that AI-based products and services offer more benefits than drawbacks. On the other hand, 40% of the respondents expressed feeling nervous when it comes to AI products and services.


Final word

When you look at the data in the report, especially how it changes over time, you can see the profound changes that have occurred in recent years. Currently, the use of artificial intelligence in business offers practically unlimited possibilities.

However, AI also has its dark side. The report points to a large number of incidents that have occurred recently. Artificial intelligence can therefore inspire fear about its ethical use.

Access the complete Stanford AI Index Report by clicking here.

Originally published on: https://trojanczyk.pl/raport-na-temat-stanu-sztucznej-inteligencji-2023/

Maybe these pieces of content will also be worth reading?

arrow-up icon