-
Chiueh Tzi-cker
The rapid rise of ChatGPT and GPT-4 has not only started a new round of technological innovation and business competition around generative artificial intelligence (AI) technology, but also raised questions about what artificial general intelligence is and how ChatGPT is its reignited a heated debate about whether such entitlements. one.
GPT-4’s incredible progress over ChatGPT in just four months has led some experts to consider whether generative AI techniques could harm society and humanity. I was.
Some experts are calling for governments to regulate generative AI in the same way that they regulate technologies such as nuclear fission and human cloning.
As a world leader in protecting fundamental freedoms and human rights, the EU has been at the forefront of efforts to address regulatory issues surrounding generative AI. So far, we have mainly focused on how to protect an individual’s privacy and reputation from being compromised, and how generative AI companies can troll from the Internet and request commercial licenses for the training data they need to train their AI models. I was.
Last month, China announced regulatory requirements for domestic generative AI companies. By default, the questions and prompts that users submit to our generative AI service cannot be used for training without our explicit permission, and the content generated by our generative AI service reflects the core values of China’s socialism. It must be reflected and cannot be used to overthrow governments.
U.S. lawmakers have also recently engaged in intense debate on how to regulate technology, but their focus has been on how to keep users safe, how to prevent criminal-generated AI from being weaponized. It was in the way, how to build enough guardrails to prevent destruction. human civilization.
Regulation of generative AI has multiple facets, but perhaps the most thorny problem is keeping it from harming society. The problem is largely rooted in the concern that generative AI is beyond the capabilities of the average human, but its “explainability” or interpretability is surprisingly poor.
Technically, there are three levels of explainability. An AI technology has a first level if it can clearly identify the elements of the inputs to the model that most influence the corresponding output.
For example, an AI model that evaluates loan applications has a first level of explainability if it can point out the factors in the loan application that most influence the outcome of the applicant generated by the model.
AI technology can be abstracted from the underlying complex mathematical model into an abstract representation that is a combination of intuitive features and high-level “if-then-else” rules, and is more comprehensible. has the explainability of human.
For example, an AI model that evaluates loan applications can be abstracted as follows: We’ll use the weighted sum of the applicant’s annual income, the probabilities of credit card and mortgage payments on time, and the expected price appreciation of the mortgages you own. House will calculate an applicant’s overall eligibility score.
The third level of explainability in AI technology has to do with fully understanding how the underlying model works and what it can and cannot do when pushed to its limits. This level of explainability is necessary to ensure that the underlying model does not contain rogue logic or mechanisms that could produce catastrophic outputs for given inputs.
For example, if you are asked how to win a car race, the AI will create a scenario where you must stage an accident that physically harms your opponent to undermine the competition.
Existing generative AI technologies, including ChatGPT, don’t even have first-level explainability.
The reason ChatGPT is so poorly explainable is that the authors who created ChatGPT have no idea why, in its current form, it is so powerful for such a diverse set of natural language processing tasks.
It is therefore impossible to predict how a technology like ChatGPT will perform in five to ten years with an order of magnitude additional training.
Imagine that one day ChatGPT does most of the writing and reading of documents in your office or publication, and you can determine that the quality of that work is much higher than what the average human produces.
In addition, from the research ChatGPT has read, we will enhance the training algorithms used to generate the basic language models, and “grow” ChatGPT itself by creating stronger language models without human involvement. can do.
What would ChatGPT choose to do with its human users when it “feels” more self-sufficient and becomes increasingly impatient with its clearly inferior users?
In a survey of elite machine learning professionals published last year, 48% said they estimated there was a 10% or greater chance that AI would have a devastating impact on humanity.
However, despite the highly likely existential threat, under intense commercial and geopolitical competitive pressures, the efforts of major AI companies to advance the AI technology frontier are daunting. It thunders without any sign of relenting or pausing for introspection, as opposed to explainability. .
If governments around the world can put together a series of regulations and intervene as soon as possible, we can at least influence AI companies to focus on explainability and put AI technology development back on a healthier, safer and more sustainable path. increase.
Chiueh Tzi-cker is a professor at the Institute of Information Security at National Tsing Hua University.
Comments are moderated. Please keep your comments relevant to the article. Statements containing abusive and obscene language, personal attacks of any kind or promotions will be removed and users will be banned.The final decision is at the discretion of Taipei Times.
