Startups testing GPT-4 say its power is amazing

OpenAI recently released its text generation model GPT-4, which has attracted widespread attention. But the company said developers can't build any products or services on it yet because the API is still on the waiting list.

This means that only a few users have the opportunity to use OpenAI’s latest large-scale language model. One such company is artificial intelligence startup Miðeind ehf, which was one of only six companies selected to test GPT-4.

A team of 12 people at the company wanted to train GPT-4 in Icelandic, becoming an early tester of Silicon Valley's hottest product after traveling to the Bay Area last May to test GPT-4. one of those. The CEO of Miðeind joins an Icelandic government delegation to explore how technology can be used to help preserve the country’s language.

Miðeind’s CEO talks with OpenAI’s Sam Altman about how GPT-4 can adopt and develop low-resource languages ​​like Icelandic. These languages ​​pose a challenge to the global application of large model languages ​​because much less data is collected to train the models.

The team at Miðeind offers their thoughts on how GPT-4 could be improved, the artificial intelligence used to preserve the Icelandic perpetual calendar, and how GPT-4 could create a very interesting new term for cats.

Exciting Development

This team at Miðeind was tasked with seeing whether they could improve GPT-4 by feeding it Icelandic reinforcement learning data (the phase after the initial training) -4 Performance in language application and processing.

Pétur Orri Ragnarsson, one of Miðeind’s machine learning team members, said the results were a clear improvement over GPT-3.5, but the model is still not perfect when it comes to the use of Icelandic. He said, "The text it generates in Icelandic tends to be understandable, but still has some grammatical errors."

Ragnarsson said he can see huge improvements in general reasoning with GPT-3.5 .

He said: "The most exciting thing is that you can ask it to do something and explain why it gives this result. GPT-3.5 can do it, GPT-4 is better because this allows People feel its explanations are more reasonable or believable. One thing people often try is to have GPT-4 do something and explain every step of the way - it does it very well."

"Explainability" is a big problem that people developing generative AI have been trying to solve, because the way large language models function means that the output is generated in a "black box." This means that even the developers building GPT-4 don't know how it answers the question, which means it's difficult to get these models to show how they work.

If generative AI is to be widely used in industries such as medicine and law, then people working in these fields need to be able to trust the output of the model.

Higher Order Thinking

Another feature of GPT-4 that impressed Ragnarsson was its ability to produce sharper responses than previous models. He gave an example of using it to perform sentiment analysis on a piece of text, with ratings ranging from neutral to positive on a scale of 1 to 5.

Ragnarsson said: "I entered a sentence that I thought was quite neutral, which was a customer asking customer service for something." He was surprised to find that GPT-4 rated this sentence as Slightly positive.

He said, "I asked, 'Please explain.' The answer I got was very surprising. It said, 'While the statement itself is neutral, the actions being considered would improve their life, so overall this sentence is slightly positive.'"

He believes this shows that GPT-4 has learned to go beyond the "surface meaning" of the text. Linda Heimisdottir, chief operating officer of Miðeind, said these capabilities of GPT-4 are particularly impressive because, to her knowledge, the model was not specifically trained for sentiment analysis.

She said: "It's amazing to see a model like this doing something that researchers have been doing for years, and that it wasn't specifically trained to do. To see its results and people's The idea is really exciting, and it makes people feel that GPT-4 has huge application potential.” Language uses compound words, which combine different concepts into one word.

Heimisdottir said that she asked GPT-4 to tell a story about cats, and GPT-4 gave an Icelandic word, "kattafræðilega", which is a compound word invented by GPT-4, and its approximate meaning Meaning "cat" (cat).

She explained: "The first part 'katta' means 'cat' but the second part 'fræðilega' means 'related to theory'. GPT-4 describes the cat as 'kattafræðilega duglegur' '.duglegur is an Icelandic word meaning diligence or hard work.

When I asked GPT-4 to explain what it meant, it said: 'kattafræðilega duglegur' means this cat is particularly diligent. In other words, it is good at scratching, investigating, chasing insects, looking for food, and is full of energy and interest in its surroundings. He's very good at being a cat. ”

Miðeind believes that “for large language models to achieve real high performance in less commonly used languages, good multilingual datasets need to be included in the initial training, and we hope that the next step can Enter pre-training. ”

Research like this is critical to ensuring that the next generation of artificial intelligence is not just an innovative advancement further concentrated in the English-speaking world, as the big tech companies in Silicon Valley already dominate the large language model field. In fact, OpenAI The company's choice of Miðeind as an early tester of GPT-4 at least shows that the company has a global vision for generative artificial intelligence, even if it is for commercial motives.

The above is the detailed content of Startups testing GPT-4 say its power is amazing. For more information, please follow other related articles on the PHP Chinese website!