The ultimate goal of artificial intelligence (AI) is to create “intelligent” machines that can perform human-like tasks. AI algorithms already impressively imitate human abilities in many areas. But are there areas of expertise in which AI is already ahead of humans? Recently released data suggests that AI has overtaken humans in areas such as image recognition, reading and language comprehension, and more. In contrast, humans remain far superior to AI in areas such as basic mathematics and computer programming.
AI is now omnipresent in our daily lives. Unknowingly, the world has adapted to this virtual “intelligence” through applications such as personal assistants, autocomplete, suggestions on Netflix and Amazon, machine translation, chatbots, computer vision and many others. AI is evolving, and analysts predict that within a decade it will be sophisticated enough to replace workers like lawyers and secretaries. But the question now arises as to whether AI is already better at certain skills than humans. According to a recent report, the answer is yes.
Based on data from the company Contextual AI, the report shows how quickly AI models began outperforming the benchmarks of certain performance tests (and whether or not they reached human performance levels). Each test is based on a specific skill, including handwriting recognition and language comprehension or reading comprehension. Several standardized tests make it possible to measure the performance of AI models. The ImageNet benchmark therefore makes it possible to measure the performance of AI systems in the area of image recognition.
Speech recognition then came in 2017 with the Switchboard database. The MNIST dataset was used to train AI systems in handwriting recognition and the SQuAD database in reading comprehension. The GLUE benchmark helps measure the performance of AI models in language understanding, and the HellaSwag dataset helps evaluate AI models in common sense questions. There are other tools for evaluating AI models. In the graph above, each percentage value is compared to the following benchmarks:
- 0% or Peak Performance Base: This value represents the best known performance of the AI at the time the data set was created;
- 100%: This rating represents human performance on the data set.
By creating a scale between these two points, one can track the progress of AI models for each data set. Each point on a line corresponds to a better result, and the higher the line, the closer AI models come to human performance. The graphic shows that AI has overtaken humans in the areas of image recognition, reading comprehension and language comprehension. Other abilities such as handwriting recognition, voice recognition, and common sense issues are more or less at the human level.
Of course, there are areas where AI is still lagging behind. Basic mathematics and computer code generation are two areas where humans remain superior to AI systems, although analysts say rapid advances could allow AI to overtake us sooner than expected. The table below shows when AI began matching human performance in the eight skills. And one of the most important observations is how much progress has been made since 2010. In fact, many of these databases – such as SQuAD, GLUE and HellaSwag – did not exist before 2015.
However, it’s worth noting that data sets could plateau by 2026, which would complicate AI progress. In response to benchmark obsolescence, some of the newer databases are continually updated with new relevant data points. For this reason, AI models have not yet technically reached human performance in some areas, although they are well on their way. It will be interesting to see how AI companies like Google and OpenAI, as well as other companies in this sector, address these challenges in the future.
But what has led to such spectacular growth in AI capabilities in recent years? According to the report, this is thanks to the revolution in computing power, availability of data and better algorithms. AI models are becoming faster, have larger data sets to learn from, and are optimized to be more efficient than they were a decade ago. However, one of the main problems for AI developers is that their models consistently beat the benchmark databases designed to test them, but fail in real-world tests.
With further computational and algorithmic advances expected in the coming years, it is likely that this rapid progress will continue. However, the next potential bottleneck to AI progress may not be AI itself, but rather the lack of data on which AI models can train.
Source: study report
And you ?
What is your opinion on this topic?
What do you think of the results of the analysis? Does this correspond to reality?
Could AI outperform programmers? If so, how soon?
What do you think has driven the rapid developments in AI?
Why does AI have a hard time doing well on math tests?
In your opinion, which skills does AI already have a better command of than humans? For what ?
In which areas could AI outperform humans in the coming years? For what ?
What impact could these developments have on the labor market and society in general?
See also
Sam Altman suggests he wants to replace middle humans with artificial general intelligence – comments that have sparked controversy
Artificial intelligence will surpass humans in five years, according to Elon Musk, who reveals how his company Neuralink, which develops computers implanted in the brain, will save us
Ilya Sutskever, scientific director of OpenAI, on the future of AI: It will be monumental and overwhelming. There will be a before and an after; he predicts the arrival of a “digital superintelligence.”