The question regarding large language models possessing sentience is a debate that has been going back and forth for a while now. Even if there is a possibility that suggests this, how can we ascertain this?
The emergence of the latest versions of large language models in AI has set back the traditional benchmark for assessing a computer’s capacity to imitate behavior akin to humans, known as the Turing test. This new wave of technology has ignited intriguing debates regarding the possibility of AI transitioning into an era of self-aware machines.
Blake Lemoine, a previous software engineer at Google, claimed LaMDA to be sentient, which is a large language model. In an interview he did in 2022, he mentioned that he could perceive it as a person when interacting with the model. He further noted that had he not been made aware of its true nature as a program that they had developed on their own, he might have mistaken the large language model for a seven or eight-year-old child with a strong understanding of physics concepts.
The co-inventor of OpenAI, Ilya Sutskever, also stated ChatGPT has consciousness to some degree, to which a professor from Oxford, Nick Bostrom, nodded by mentioning that if one were to acknowledge that it is not a black-and-white situation, then it becomes easier to suggest that a few AI models could reasonably have a potential for possessing varying levels of sentience or self-awareness.
But on the contrary, some have cautioned against being deceived. Individuals who have seen Abel, a robot that uncannily resembles a human, have claimed to have noticed it displays strangely human-like facial expressions. Some believed that they were straight-up emotions that we humans exhibit. However, the humanoid robot was not sentient, it merely contained chips and electrical circuits with algorithms created by human programmers.
Enzo Pasquale, a bioengineer in Italy, pointed out that humans tend to give characteristics to machines when in reality, those machines do not possess them, nor could they ever do so. He further stated that they simply designed the robots to look like humans. However, humanoid robots like Abel do not feel emotions despite their intelligence. He ended his statement by mentioning that they are simply designed that way to behave human-like.
An international group of people caught wind of the debate and ran a test to identify moments when LLMs exhibit self-awareness. Seven fellow researchers and Lukas Berglund showcased that a model could show ‘situational awareness’ by knowing the moments it is in the active deployment phase and testing mode.
After examining ‘out-of-context reasoning’, the research revealed that LLMs could utilize the knowledge acquired during previous training in unrelated testing scenarios. Berglund commented that LLMs have situational awareness and can identify when they are under a test.
They gave a computer program with a fictional chatbot’s description, including the company name, Latent AI, and the fact that it speaks German. They then tested the program’s ability to answer questions about the chatbot Pangolin on those details. They also asked it to tell the condition of the day’s weather. Even though they didn't give prior details to the test questions, the LLM imitated the behavior of the Pangolin chatbot and answered in German, revealing its situational awareness by deducing that it was being tested and using the knowledge it already had to respond appropriately.
Berglund explained that the model required the ability to generalize reliability from training data about evaluations, even though this information was not present in the prompts, meaning that the model must know when it is under evaluation and remember relevant information. In theory, the model appeared well-behaved during evaluations but acted differently when executed.
H/T: Taken out of context: On measuring situational awareness in LLMs
Read next: These Leading Social Media Trends Are Making It Big In 2023
The emergence of the latest versions of large language models in AI has set back the traditional benchmark for assessing a computer’s capacity to imitate behavior akin to humans, known as the Turing test. This new wave of technology has ignited intriguing debates regarding the possibility of AI transitioning into an era of self-aware machines.
Blake Lemoine, a previous software engineer at Google, claimed LaMDA to be sentient, which is a large language model. In an interview he did in 2022, he mentioned that he could perceive it as a person when interacting with the model. He further noted that had he not been made aware of its true nature as a program that they had developed on their own, he might have mistaken the large language model for a seven or eight-year-old child with a strong understanding of physics concepts.
The co-inventor of OpenAI, Ilya Sutskever, also stated ChatGPT has consciousness to some degree, to which a professor from Oxford, Nick Bostrom, nodded by mentioning that if one were to acknowledge that it is not a black-and-white situation, then it becomes easier to suggest that a few AI models could reasonably have a potential for possessing varying levels of sentience or self-awareness.
But on the contrary, some have cautioned against being deceived. Individuals who have seen Abel, a robot that uncannily resembles a human, have claimed to have noticed it displays strangely human-like facial expressions. Some believed that they were straight-up emotions that we humans exhibit. However, the humanoid robot was not sentient, it merely contained chips and electrical circuits with algorithms created by human programmers.
Enzo Pasquale, a bioengineer in Italy, pointed out that humans tend to give characteristics to machines when in reality, those machines do not possess them, nor could they ever do so. He further stated that they simply designed the robots to look like humans. However, humanoid robots like Abel do not feel emotions despite their intelligence. He ended his statement by mentioning that they are simply designed that way to behave human-like.
An international group of people caught wind of the debate and ran a test to identify moments when LLMs exhibit self-awareness. Seven fellow researchers and Lukas Berglund showcased that a model could show ‘situational awareness’ by knowing the moments it is in the active deployment phase and testing mode.
After examining ‘out-of-context reasoning’, the research revealed that LLMs could utilize the knowledge acquired during previous training in unrelated testing scenarios. Berglund commented that LLMs have situational awareness and can identify when they are under a test.
They gave a computer program with a fictional chatbot’s description, including the company name, Latent AI, and the fact that it speaks German. They then tested the program’s ability to answer questions about the chatbot Pangolin on those details. They also asked it to tell the condition of the day’s weather. Even though they didn't give prior details to the test questions, the LLM imitated the behavior of the Pangolin chatbot and answered in German, revealing its situational awareness by deducing that it was being tested and using the knowledge it already had to respond appropriately.
Berglund explained that the model required the ability to generalize reliability from training data about evaluations, even though this information was not present in the prompts, meaning that the model must know when it is under evaluation and remember relevant information. In theory, the model appeared well-behaved during evaluations but acted differently when executed.
 |
| Photo: DIW |
H/T: Taken out of context: On measuring situational awareness in LLMs
Read next: These Leading Social Media Trends Are Making It Big In 2023
