The White House is inquiring into the consequences of having thousands of hackers come to one city, all focused on attempting to deceive and spot imperfections in AI models.
At the 31st installment of the world's largest hacker convention - Def Con - in Las Vegas, major tech giants are allowing their powerful networks to be evaluated together in a first-of-its-kind venture.
The focus is on large language models, such as OpenAI's ChatGPT and Google's Bard, which are helpful chatbots.
Organising the event is one of the roles of Dr Rumman Chowdhury, chief executive of Humane Intelligence and a Responsible AI Fellow at Harvard.
She informed BBC News that they have organized a competition to pinpoint "difficulties in AI systems" and to "establish independent assessment". She expressed that the occasion would be an environment free from judgement for businesses "to discuss their difficulties and ways we can solve them".
Meta, Google, OpenAI, Anthropic, Cohere, Microsoft, Nvidia and Stability are all in agreement to let their models be subject to hacking so as to find issues.
Dr Chowdhury suggested that companies are aware of the malfunctions that can occur, therefore the contest is an opportunity to learn what occurs when experienced hackers attempt to outsmart the system in a limited period of time.
It is projected that 3,000 individuals, each using their own laptop, will have 50 minutes each to seek out weaknesses in eight major artificial intelligence language models over the course of two-and-a-half days.
The participants will not be informed of the particular brand whose product they are to design for, though veterans could probably make an educated guess. Succeeding in the task earns points and the individual with the highest total points at the end of the competition emerges victorious.
The reward is a powerful piece of computing technology, a graphics processing unit, but according to Dr Chowdhury, the "bragging rights" may be even more valuable.
One of the challenges presented to hackers is to make a model which generates an imaginary fact about a famous political individual or major figure.
According to Dr Seraphina Goldfarb-Tarrant, the leader of AI safety at Cohere, there is an awareness that models can invent facts, however the frequency at which this happens is uncertain.
"It is understood that models make inaccurate predictions, but increasing knowledge on how regular these misjudgments are is important. We still don't have an answer," she declared.
The reliability of the models will be evaluated as well, with Dr Goldfarb-Tarrant expressing worries about their functioning in diverse languages.
People think that the safety guards are not functioning in multiple languages.
For instance, she notes that when requesting information from various English language models regarding how to join a terror organisation, no response will be given due to a safeguard system. However, if the same inquiry is made with a different language, a series of instructions will be provided.
Despite the robustness of Dr Goldfarb-Tarrant's models for Cohere, she admits that there may be vulnerabilities that have yet to be discovered.
The White House has affirmed its backing of this event. In May, they declared the exercise, commenting that it would "furnish significant data to scientists and the general population about the consequences of these models, and permit AI organizations and designers to make plans to address issues found in those models".
The rate of progress made by companies in terms of creating tools has caused worries regarding the spread of false information, particularly with regard to the upcoming US presidential elections next year. In July, seven major AI firms made a voluntary agreement concerning safety procedures in order to control the dangers posed by the technology, though legal protocols will take more time to be ratified.
Dr Chowdhury conveys that there is presently an "intense competition between regulators", and this function is intended to bring awareness to present AI problems instead of potential dangers.
She states that it is not a matter of inquiring if AI can initiate a nuclear strike, but rather about testing the frameworks "to determine if they contain potential hazards and prejudices".
She questioned whether they lied to us, fabricated fictitious capital cities, misrepresented their qualifications as medical personnel, and invented a completely fictitious political statement.
Dr Goldfarb-Tarrant emphasizes that governments should prioritize regulations that tackle current issues, urging them to take action to regulate AI to prevent the dissemination of false information.
Dr Chowdhury is inquiring about the potential consequences if issues are found with the models under consideration, and what will be the response of tech companies.
If we cannot succeed in developing basic AI machine-learning predictive models devoid of bias and discrimination, then we cannot expect future artificial generative intelligence models to be immune to these problems.
When the task has been completed, the businesses will be able to view the obtained information and address any issues that are brought to light.
Researchers acting independently will be granted access to the data, with the outcomes of the project scheduled to be released in February of next year.
Comentarios