A generative artificial intelligence is a technology capable of learning complex patterns of behavior from a database. With a technique called machine learning, generative AIs like ChatGPT and DALL-E are able to reproduce content after receiving training.
- Poe | App lets you create your own ChatGPT using text commands
- 10 extensions for Chrome that make ChatGPT more powerful
With the absorption of a lot of data, AI is able to generate new information in an original and even unique way for each interaction. In addition, its technical construction allows it to go beyond conventional learning, which enables constant evolution, on its own, without the need for human programming.
/i634242.jpeg)
To create a generative AI it is necessary to add a huge volume of texts, videos or images that will be processed. That done, whenever someone gives a command, the technology will offer an answer that can be right or wrong.
–
Canaltech on Youtube: news, product analysis, tips, event coverage and much more! Subscribe to our YouTube channelevery day there’s a new video for you!
–
What is the origin of generative AIs?
Thiago Pardo, a professor at the Institute of Mathematical and Computer Sciences (ICMC) at USP, points out that technologies such as ChatGPT are the result of two major revolutions: one in the field of AI and the other in Natural Language Processing (NLP).
“Generative AI uses the modeling of Transformerswhich, in general terms, are sets of artificial neural networks modeled to be more ‘attentive’ to what they have to learn”, explains the professor. “These models use numerical representations for the words of the language, the so-called word embeddingswhich, in turn, are based on classical linguistic assumptions”, he adds.
/i617258.jpeg)
For NVIDIA’s developer relationship manager for Latin America, Jomar Silva also points out that generative AIs are based on neural networks, the result of decades of research.
“Nowadays, there are several models available, each with its own way of processing. An artificial intelligence for conversations, such as GPT-3, was trained with a huge amount of texts extracted from the internet, approximately half a trillion words”, he details.
In this case — of text generators, such as ChatGPT or Bing Chat —, the capacity is in the order of millions of parameters, that is, of “artificial neurons” that take hours, days, weeks and even months to process knowledge. It all depends on the size of the neural network used and the volume of data used in training.
Jomar also explains that more recent technologies, such as the GPT-4, receive training by merging texts and images, which further reinforces their learning power. “The result of this is much, much more interesting than the old model that people got used to”, highlights the executive.
How is a generative AI trained?
When an AI model has an incorrectly presented answer, developers send feedback with a red flag for assimilation of the failure. When the answer is correct, a positive hit message is sent, which is also incorporated into the model.
This training system is valid for any format — text, images, videos or audios. The difference here is only in the technical part, that is, how the robot will interpret the command: if the model is an art creation AI, for example, it will need to convert text into an image to present the result.
After millions of interactions like this one, based on true or false, the machine is so finely tuned that it starts to get most queries right. The model’s errors and inconsistencies are corrected over time, until it gets closer and closer to perfection.
/i671032.png)
“It all depends on the type of problem we want to solve. For example, how will the machine determine whether the image is a cat or a dog? I need to divide this image into categories: dogs, cats and other animals. I’m going to tell each image what we see and the neural network will absorb that. When she sees a similar image again, she will immediately make the association”, explains Jomar Silva.
Application of generative AI in everyday life
There are, however, complex problems to be solved in sectors such as medicine. The NVIDIA representative says that an AI could be aimed at detecting a cancerous tumor, for example. In this specific case, the technology would need to be trained with X-rays or MRI scans.
“You need to note which pixels in that image are part of what you want to identify. You gather this set of data, called a dataset, and assemble it all on a training platform”, he details.
The problem is that a tumor can have different shapes, colors and textures, so you have to repeat the process countless times to reach an acceptable pattern. Even so, it is likely to take a long time to reach an optimal level of accuracy. On the other hand, there is already a practical application of ChatGPT in the medical sector to compose medical records from data entry.
This is why generative AIs are rarely (or will ever be) created by small companies. The cost is high to create something from scratch, because highly specialized professionals, technological infrastructure and a long development time are required.
“The good news is that, once these networks are trained, they can go through some transfer learning processes, which we call transfer learning, fine-tuning. These tweaks are generally much cheaper and faster to make than training the network from scratch. And this is now accessible to any company”, concludes Jomar.
AIs must change society
Professor Thiago Pardo believes that this technology will be disruptive, that is, capable of drastically altering the way human beings relate to machines. “The question that hangs in the air is what level of revolution does ChatGPT represent and whether it will, in fact, replace traditional search engines and automatic translators”, analyzes the USP expert.
The consequences of generative AI can cause an important revolution in the content production segment. “Many of these consequences will be harmful for different sectors of society. In addition to the possibility of misuse of technology, and ethical issues, the potential to exacerbate economic and social inequality is certain, both between nations and between individuals of the same nation, which is common for economies based on high technology”, emphasizes Pardo .
In the end, it will be up to global society to define how to deal with this technological innovation. Will it be used to add value to the planet, or will it become just another misused tool? No expert has yet dared to pin down an answer to that.
Read the article on Canaltech.
Trending no Canaltech:
- Samsung fixes bug in One UI 5.1 that drained battery
- Hunger for Success Review | Film brings good social analysis albeit with mistakes
- New Toyota Prius averages absurd consumption in test in Japan
- Video shows hole gushing hot water at the bottom of the ocean
- Why is Mario Italian and not Japanese?
- Xiaomi 13 x Zenfone 9 comparison: which is the best compact?
