Gen AI’s Accuracy Problems Aren’t Going Away Anytime Soon, Researchers Say

Generative Ai Chatbots It is known that they are very wrong. Let’s hope you have not tracked Google’s AI offering Add glue to your Pizza recipe Or eat a rock or two days for your health.

These errors are known as hallucinations: essentially, the model of the model. Will this technology improve? Also read researchers AI It’s not optimistic that will happen soon.

This is one of the findings of a panel Two On Artificial Intelligent Specialists This month was released by this monthly association this month for the progress of artificial discipline. The group also examined more than 400 members of the Association.

You can provide information about the developers of these tools for more years (or months), which are far from improving the EU (or for months), how these vehicles and industry experts have been moving these tools. This includes just getting facts and avoid strange mistakes. The reliability of AI instruments should be increased sharply in case of producing a model that can produce to introduce or pass human intelligenceGenerally known as artificial general intelligence. Researchers believe that they are likely to develop in this process.

“We do not believe in something until a little careful and actually worked” Vincent ConitzerCarnegie Mellon University said that computer science is one of the professors and panelists.

In recent years, artificial intelligence developed rapidly

The purpose of the report is the AAAI President Francesca Rossi, wrote with the application, is to support research in artificial intelligence that produces technology that helps people. The issues of trust and reliability are not serious in providing serious information not only to provide accurate information, but to prevent the future AI’s serious result. “We must work together to advance the AI ​​to make sure that technological progress is supported by mankind and adjust to human values.

EU’s acceleration, especially since Openai launched Chatgpt Conitzer said in 2022. “Surprising in some ways, and many of these methods are working better than most of us,” he said.

There are some areas of AI research, “There is a dignity of the hype” John TholtstunAt Cornell University, the Associate Professor of Computer Science told me. Especially in mathematics or science that users can check the results of a model.

“This technology is amazing,” he said. “I have been working in this area for more than a decade and shocked how well it will be and how fast it will be.”

Despite these improvements, there are still serious problems, the experts are still taken into account.

Will Chatbots start to get the facts straight?

Despite some improvements in improving the reliability of information from generative AI models, more work should be done. Recently Report from Columbia Journalism Chatbots found, it was not possible to answer questions that they could not respond to the wrong data and (and fictional links) sources they submitted to support these wrong claims.

Improving reliability and accuracy “Today is the largest area of ​​AI research,” said AAIAI sheet.

Researchers noted the three main way to increase the accuracy of AI systems: beautiful regulation, such as human opinion strengthening with learning; A generation of searching for the system where the system has collected and responded to the answer; And something thoughtful, the desires are where the question is reduced to smaller steps that the AI ​​model can check the hallucinations.

These things make chatbot answers more accurately coming soon? It is not likely: “Fishery is far from being solved,” the report said. About 60% of those involved in the survey showed that the concerns of truth or reliability would be resolved recently.

In the generative AI industry, optimism that expands the scale of existing models, will further clear them and reduces the hallucinations.

“I think hope was always a little overwhelmed,” he said. “In the last few years, I did not see any evidence that really accurate, high factual language models are in the corner.”

In spite of the mistake of large language models like Anthropic cluster or Meta callsUsers may think that they are wrong because they are confidently providing answers because they answered Conitzer.

“If we see someone’s confident response or confidently sounding words, we see what a person really knows what they are talking about,” he said. “The EU system can only claim to be very confident in something that is completely nonsense.”

Classes for AI user

Information about the restrictions of generative AI is important to use it properly. ChatGpt and thick stone recommendation for users of models like and Google’s Twins Simple: “You should check the results.”

Wide large language models consistently show a weak job of obtaining actual information. If you want something, you should probably probably have to follow the answer in a search engine (and not to trust the search results AI summary). If you do this, you could have been better than to do it in the first place.

Talk, the most commonly used way of AI models, automate the tasks that can do anyway, and it can check the accuracy as it format information or writing code. “The extensive principle is that it is most useful for me to find these models,” he said.

Read more: 5 ways to stay smart when using the Gen AI explained by the Professors of Computer Science

Is artificial general intelligence around the corner?

A priority of the AI ​​Development Industry is a competition that is often visible to create artificial general intelligence or AGI. This is generally a model that is capable of a person’s thinking or better level.

The request of the report found strong ideas in the competition for AGI. It should be noted that more than three-quarters of respondents (76%), such as large language models, it is impossible for the expansion of current AI techniques. The most important majority of researchers doubt that the current march will work to AGI.

So far, the same vast majority, which is capable of artificial intelligence, has so far have the public by private institutions (82%). It is associated with concerns about creating a system that can concern people and concerns about potential landings. Most researchers (70%) said they were against the suspension of AGI research until security and management systems are developed. “These answers prefer to prefer to continue the subject within some information,” the report said.

The conversation around Agi is complicated. In some sense, we have already created systems with a general exploration form. Large language models like Openai Chatgpt can do different human activities, unlike old AI models that can do anything like only anything. The question is something that a person can do much in a row.

“I think we are far away from it,” he said.

He said that these models do not have the ability to manage the concept of a truth and really open creative tasks. “I do not see the way to manage them in a healthy human environment using existing technology.” “I think there is many research progress on the way to go there.”

Conitzer, the definition of those who express the AGI in a clear way: often people mean something that can make better tasks better than a person, some say this is something that is able to perform only a number of tasks. “A certain definition is something that will really make us completely unnecessary,” he said.

Researchers are these skeptics AGI is in the cornerConitzer warned that the AI ​​researchers did not expect the dramatic technological development of what we saw in the last few years.

“We did not see how quickly this has changed soon,” and therefore you can think we will see if we continue to go faster. “