Large Language Models such as GPT along with its consumer facing chatbot ChatGPT have shown that AI has a bright future ahead of it. In spite of the fact that this is the case, AI tends to glean a lot of information from the internet, and with so much of this information being generated by AI itself, this could lead to future technology being unable to say anything coherent with all things having been considered and taken into account.
British and Canadian researchers put out a paper recently, saying that after a few generations of AI are trained based off of each other’s output, they might make less sense than might have been the case otherwise. For example, one AI that was in its ninth generation was asked about Medieval architecture, but instead it started talking about entirely irrelevant topics such as jackrabbits.
With all of that having been said and now out of the way, it is important to note that scientists are referring to this phenomenon as model collapse. Scraping data from the internet may not be tenable in the near future because of the fact that this is the sort of thing that could potentially end up dramatically reducing the quality of the AI’s output.
The vast majority of data that current AI was trained on was generated by humans, since there weren’t any advanced chatbots that could be used to automate this process until ChatGPT finally made its way to consumers. Now that so much online content comes from AI, the successive training algorithms will dramatically increase incoherence until eventually crashing the system in its entirety.
AI is not immune to making errors, and other Ais that learn these errors will start to spiral out of control. The basic premise behind AI is that it usually falls short of the level of quality that human beings can produce, at least right now. Future Ais will learn all of these mistakes, and they will become continuously more pronounced until they are eventually impossible to ignore. It will be interesting to see how the industry responds to these claims.
Read next: Cost of Living Index 2023: Which Countries are the Most Expensive to Live in?
British and Canadian researchers put out a paper recently, saying that after a few generations of AI are trained based off of each other’s output, they might make less sense than might have been the case otherwise. For example, one AI that was in its ninth generation was asked about Medieval architecture, but instead it started talking about entirely irrelevant topics such as jackrabbits.
With all of that having been said and now out of the way, it is important to note that scientists are referring to this phenomenon as model collapse. Scraping data from the internet may not be tenable in the near future because of the fact that this is the sort of thing that could potentially end up dramatically reducing the quality of the AI’s output.
The vast majority of data that current AI was trained on was generated by humans, since there weren’t any advanced chatbots that could be used to automate this process until ChatGPT finally made its way to consumers. Now that so much online content comes from AI, the successive training algorithms will dramatically increase incoherence until eventually crashing the system in its entirety.
AI is not immune to making errors, and other Ais that learn these errors will start to spiral out of control. The basic premise behind AI is that it usually falls short of the level of quality that human beings can produce, at least right now. Future Ais will learn all of these mistakes, and they will become continuously more pronounced until they are eventually impossible to ignore. It will be interesting to see how the industry responds to these claims.
Read next: Cost of Living Index 2023: Which Countries are the Most Expensive to Live in?
