And we're going to show you that this is a car.
It's very clear what went wrong.
From the limited number of images it was trained on, the AI determined that color was the most powerful way to differentiate between cars and vans.
But the amazing thing about AI programs is that they made this decision for themselves, and we can refine their decision-making.
You can see that we have incorrectly identified two new objects. This forces us to find new patterns within the image.
But more importantly, by feeding the training data with more diverse images, we can correct the bias in the training data.
By combining these two simple actions and performing them at scale, most AI systems have been trained to make incredibly complex decisions.
How does AI learn by itself?
Although supervised learning is a very powerful training method, many of the recent advances in AI have been made possible by unsupervised learning.
In the simplest terms, this means that through the use of complex algorithms and large datasets, AI can learn without human guidance.
ChatGPT may be the most well-known example.
Because the amount of text on the Internet and in digitized books is so vast, ChatGPT was able to spend months teaching itself how to combine words in meaningful ways and then helping humans fine-tune its responses.
Imagine you have a large collection of books in a foreign language, some of which contain images.
Eventually, you might find that the same word appears on the page whenever there is a picture or picture of a tree, and a different word appears on the page when there is a picture of a house.
And there are often words near those words that can mean things like “a” or “the.”
ChatGPT has built a huge statistical model that can be used to make predictions and generate new sentences through this in-depth analysis of relationships between words.
This relies on vast amounts of computing power, allowing AI to memorize vast numbers of words, singly, in groups, within sentences, and across pages, and then read and compare how they are used over and over again in a fraction of a second.
Do I need to worry about AI?
Rapid advances in deep learning models over the last year have sparked a wave of enthusiasm and public interest in concerns about the future of artificial intelligence.
There has been much discussion about how biases in training data collected from the internet, such as racist, sexist, and violent speech and narrow cultural viewpoints, can lead to artificial intelligence replicating human biases.
Another concern is that artificial intelligence will be tasked with solving problems without fully considering the ethics and broader implications of its actions, potentially creating new problems in the process.
In AI circles, this has become known as the “Paperclip Maximizer Problem” following a thought experiment by philosopher Nick Bostrom.
He imagined an artificial intelligence being asked to create as many paperclips as possible, and slowly repurposing all of Earth's natural resources to accomplish its mission. This includes killing humans to use as raw material for more paperclips.
Some say that instead of focusing on the future of killer AI, we should be more concerned with the pressing question of how people can use existing AI tools to increase distrust of politics and skepticism of all forms of media.
The world's attention is focused on the 2024 US presidential election, especially how voters and political parties will deal with new levels of sophistication and disinformation.
What happens when social media is flooded with fake videos of presidential candidates created by AI and tailored to offend different groups of voters?
In Europe, the EU has enacted artificial intelligence laws that protect citizens' rights by regulating the introduction of AI. For example, it prohibits the use of facial recognition to track or identify people in public in real time.
These are among the first laws in the world to establish guidelines for the future use of these technologies, setting boundaries for what businesses and governments can and cannot do. However, as artificial intelligence capabilities continue to grow, this is unlikely to be the last law.
