Generative AI chatbots are known to make a lot of mistakes. Hopefully you have not followed the suggestion of Google AI to add glue to your pizza recipe or eat a rock or two per day for your health.
These errors are called hallucinations: essentially, the things that the model constitutes. Will this technology improve? Even researchers who study AI are not optimistic who will soon occur.
This is one of the conclusions of a panel of Two dozen experts in artificial intelligence Released this month by the association for the progress of artificial intelligence. The group also questioned more than 400 of the members of the association.
Unlike the media threshing that you can see on the fact that developers are only years (or months, depending on which you ask) far from the improvement of AI, this panel of university and industry experts seems more kept on the speed with which these tools will progress. This includes not only the right facts and avoiding bizarre mistakes. The reliability of AI tools must increase considerably if the developers will produce a model that can respond or go beyond human intelligence, commonly known as artificial general intelligence. Researchers seem to believe that improvements on this scale are probably not happening soon.
“We tend to be a little careful and not to believe something before it really works”, ” Vincent ConitzerIT professor at Carnegie Mellon University and one of the panelists, told me.
Artificial intelligence has developed quickly in recent years
The objective of the report, the president of AAAI, Francesca Rossi, wrote in its introduction, is to support research in artificial intelligence that produces a technology that helps people. The problems of confidence and reliability are serious, not only to provide precise information, but to avoid biases and guarantee a future AI does not cause serious involuntary consequences. “We must all work together to advance AI in a responsibility, to ensure that technological progress supports the progress of humanity and is aligned with human values,” she wrote.
The acceleration of the AI, in particular since the launch of Openai in 2022, was remarkable, said Conitzer. “In some ways, it has been magnificent, and many of these techniques work much better than most of us thought they would,” he said.
There are certain areas of research on AI where “beaten media has merit”, ” John ShipstunAssistant IT professor at Cornell University, told me. This is particularly true in mathematics or sciences, where users can check the results of a model.
“This technology is incredible,” said Shattum. “I have been working in this area for more than a decade, and it shocked me how good it is and how good it has become good.”
Despite these improvements, there are still important problems that deserve research and consideration, experts said.
Will chatbots start to get their facts directly?
Despite certain progress in improving the reliability of information from generative AI models, you have to do much more work. A recent Columbia Journalism Review report It was unlikely that chatbots will refuse to answer questions to which they could not answer with precision, confident about the bad information they have provided and established (and provided links made to) sources to support these bad affirmations.
Improved reliability and precision “is undoubtedly the largest field of AI research”, according to the AAAI report.
The researchers noted three main ways of strengthening the accuracy of AI systems: fine adjustment, such as strengthening learning with human feedback; Generation with recovery, in which the system brings together specific documents and draws its response from these; And the chain of thoughts, where the prompts decompose the question in smaller steps that the AI model can check the hallucinations.
Will these things soon make your chatbot answers more precise? Unlikely: “the billing is far from being resolved”, according to the report. About 60% of those questioned indicated that doubts that problems of billing or reliability would soon be resolved.
In the AI -generating industry, there has been optimism that the scaling of existing models will make them more precise and will reduce hallucinations.
“I think hope has always been a little too optimistic,” said Shipstun. “In the past two years, I have not seen any evidence that very precise and very factual language models are approaching.”
Despite the fallibility of large languages models such as Claude d’Anthropic or Llama de Meta, users can wrongly assume that they are more precise because they present answers with confidence, said Conitzer.
“If we see someone responding with confidence or words that seem confident, we consider that the person really knows what he is talking about,” he said. “An AI system, it may well claim to be very confident about something that is completely absurd.”
User lessons for AI
Consciousness of the AI -generating limitations is essential to use it correctly. Eatchstun’s advice for model users such as Chatgpt and Google Gemini are simple: “You must check the results”.
The general models of large language do a bad job to regularly recover factual information, he said. If you ask him for something, you should probably follow by looking for the answer in a search engine (and not relying on the summary of the search results). As you do this, you could have done it better in the first place.
Emphipstun said that the way he uses AI models is to automate the tasks he could do anyway and that he can check the accuracy, such as shaping information or code writing tables. “The wider principle is that I find that these models are the most useful for automating the work you already know how to do,” he said.
Find out more: 5 ways of staying intelligent when using Gen Ai, explained by computer teachers
Is general artificial intelligence at the corner of the street?
A priority for the AI development industry is an apparent breed to create what is often called artificial general intelligence, or acted. It is a model that is generally capable of a level of human thinking or better.
The investigation of the report revealed strong opinions on the Act race. In particular, more than three -quarters (76%) of respondents said that the scaling of current AI techniques such as large languages models were unlikely to produce AC. A large majority of researchers doubt that the current march towards AC will work.
One majority also important believes that systems capable of general artificial intelligence should be public if they are developed by private entities (82%). This is aligned with concerns about ethics and the potential disadvantages of creating a system that can surprise humans. Most researchers (70%) have declared to oppose the stopping of research acted until the security and control systems are developed. “These answers seem to suggest a preference for continuous exploration of the subject, in certain guarantees,” said the report.
The conversation around Act is complicated, said Emphstun. In a sense, we have already created systems that have a form of general intelligence. Models of large languages such as Openai Chatppt are able to do a variety of human activities, unlike older AI models that could only do one thing, like game failures. The question is whether it can do a lot in a coherent way to a human level.
“I think we are very far from that,” said Shotpstun.
He said that these models do not have a concept of integrated truth and the ability to manage really open creative tasks. “I do not see the way to make them work robustly in a human environment using current technology,” he said. “I think there are many advances in research in the way of getting there.”
Conitzer said that the definition of what is exactly act is delicate: often people mean something that can do most of the tasks better than a human, but some say that it is just something capable of doing a range of tasks. “A stricter definition is something that would really make us completely redundant,” he said.
While researchers are skeptical AG is at the corner of the streetConitzer warned that IA researchers did not necessarily expect the dramatic technological improvement that we have all seen in recent years.
“We have not seen what things have changed recently,” he said, “and you may wonder if we will see him coming if it continues to go faster.”
