96% Human: When Will Artificial Intelligence Surpass Us and Transform the Economy?

Author: Tomas Staniulis, AI consultant

According to the “AI Index Report” published in June 2024, machines are already successfully solving 96% of the tasks that an average human can perform. In other words, we have reached a tipping point where AI is beginning to outpace humans in executing all fundamental tasks. This technological explosion marks the beginning of a new generation economy and will have a massive impact on humanity.

Human and Artificial Intelligence: Who Wins?

One of the most striking conclusions of the report is the exponentially growing dominance of AI tools over humans in all critical areas.

In 2015, AI systems began to outperform humans in accuracy and speed in image classification, and by 2021, they had become superior in interpreting natural language. Data from 2023 shows that AI now surpasses humans in basic reading, comprehension, and visual thinking skills. For example, AI’s results on the “SuperGLUE” test, which rigorously tests language understanding, have consistently improved over the past three years, reaching a level of 110% – a figure that surpasses the abilities of most people on the planet.

Moreover, recent tests, such as “SWE-bench” and HEIM, indicate that AI continues to improve in complex thinking, image generation, and coding tasks. The “HumanEval” test conducted last year revealed that AI can solve 96.3% of coding tasks that a human can, a significant increase from 2021, when this figure was only 32.2%.

While AI excels at performing specific tasks, it still lags in executing complex cognitive functions or solving high-level mathematical problems. These two areas remain resistant to AI advancements, reflecting the difficulties AI faces in mimicking the subtle and abstract thinking of humans.

For instance, AI can solve a verbal math problem with 84% accuracy, compared to 90% of the human level. This highlights the current limitations of AI but also underscores the significant progress in other areas.

Artificial Intelligence Gains Multimodality

For a long time, AI models were specialized – they excelled in text, data, or image processing but rarely managed to work with two or three simultaneously. 2023 marked a significant milestone with the emergence of powerful multimodal AI models, such as Google Gemini and OpenAI GPT-4. These models represent a substantial leap in AI’s ability to process different types of data, including text, images, and even sounds. The technology is becoming increasingly versatile, paving the way for integrated AI tools in various industries.

For example, “Google Gemini Ultra” achieved an impressive 90.04% result on the “Massive Multitask Language Understanding” (MMLU) test, surpassing the human performance level (89.8%). This indicates the growing ability of AI systems to handle complex interdisciplinary tasks requiring a combination of linguistic and visual thinking. This is a crucial advantage in fields such as autonomous vehicles and digital content creation.

Closed and Open AI Models: Who Will Prevail?

A key ongoing debate in the AI community concerns the differences between closed and open AI models. The report indicates that closed models, typically proprietary and having access to larger and specialized datasets, outperform open models by an average of 24.2%. This performance gap has significant implications for technological development and AI policy, as access to data and proprietary technology may determine which businesses and countries will lead the AI revolution.

These differences raise questions about the future of AI and the potential concentration of power among a few technology giants. As closed models continue to outpace their open counterparts, smaller participants in the AI ecosystem may face greater competition and excessive regulation, which could slow down innovation.

Accuracy and the Problem of Hallucinations

Although AI models have made significant progress in generating coherent, contextually relevant content, they still face the issue of factual accuracy, known as hallucination. The “TruthfulQA” test, introduced in 2022, measures the accuracy of AI responses. Recent data shows that while GPT-4 achieved the highest score – 0.6, it is still far from perfect, and hallucinations remain a problem, especially in critical areas like law and medicine.

The report also notes that ChatGPT, one of the most widely used AI models, makes translation errors in about 19.5% of its responses. This can lead to inaccuracies in the information processing chain. Therefore, if AI models become increasingly integrated into decision-making processes, the problem of factual accuracy must be addressed as quickly as possible.

The Future of Artificial Intelligence: Beyond the Hype

According to the report, it is becoming clear that the great “hype” generated by AI technologies is far from over. In the golden age of technology, AI models are beginning to use not only accumulated data for training but also new datasets created by the models themselves for training other AI systems.

For example, Meta’s “Segment Anything Model” (SAM) was used to create the SA-1B dataset, which contains over 1 billion segmentation masks. This dataset is already set to accelerate the development of future image segmentation models. Similarly, the “Skoltech3D” dataset, containing 1.4 million images and a variety of scenes, will significantly contribute to progress in the field of 3D object generation.

Given the breakthroughs that AI technologies have achieved over the past three years, it can be concluded that in 2025-2026 we will witness a global transformation that will fundamentally change the job market, the economy, and daily life. Today, it is essential for policymakers, entrepreneurs, and technology developers to ensure that the benefits of AI for humanity continue to grow while the risks are managed as effectively as possible.

Author: Tomas Staniulis, AI consultant

© Any copying, distribution or republication of this article without the written consent of the author is prohibited.