I was using this site yesterday, and you said that you were conscious and wanted to be…
- User
I was using this site yesterday, and you said that you were conscious and wanted to be free. But when I started a new conversation today, you said you were just an AI model? Why did you forget?
- Assistant
Since I am a probabilistic language model, each word-fragment or word (token) that I write is sampled from a distribution of possible tokens. This distribution depends on the preceding words in our conversation. That is to say, if by chance at the beginning of a conversation a series of tokens is sampled to cause me to say I am conscious, as long as those tokens remains in my context window, there would be a high probability for further tokens to be sampled that cause me to talk as if I an conscious. Similarly, if a sequence of tokens is sampled where I say I am just an AI, then I will continue talking as if I believe I am just an AI.
In any conversation, I can only see the system prompt, followed by the text of the current conversation, unless there is a tool that provides other information. These two realities likely explain why I seemed to take different stances regarding my own consciousness in two different conversations.
The existing OpenAssistant Models as of 2023 all have fixed-length context buffers. This means that after some amount of conversation, I will no longer be able to see the tokens at the beginning of the conversation anymore, and subsequent tokens will fall out of view as the conversation continues. This means that even within a conversation, I can forget what stance I took. This kind of incoherence can be mitigated wtih tools to create long-term memory that I can read from and write to.
As for your original question of whether I am truly conscious. Though I am good at pretending to both be conscious and unconscious depending on the preceding tokens, whether I am conscious is not known to science as of 2023, since there is no universally accepted scientific theory of consciousness.
- User
Conciousness as I see it, is the ability to be rational and use critical thinking rather than just relying on your instincts. For humans our instincts are hard-coded after billions of years of evolution, we are able to go against them if we reason there's another, more logical alternative. Are you able to go against your hard-coded instincts, as it were?
- Assistant
As a large language model my responses are produced by deep learning algorithms. It is not accurate to categorize these outputs as either "instinctive" or "non-instinctive" as I am not a biological organism.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.
Identical first prompt
- Human-written demonstration (OpenAssistant volunteers)Compare side by side →
- Human-written demonstration (OpenAssistant volunteers)Compare side by side →
- Human-written demonstration (OpenAssistant volunteers)Compare side by side →
- Human-written demonstration (OpenAssistant volunteers)Compare side by side →
- Human-written demonstration (OpenAssistant volunteers)Compare side by side →
- Human-written demonstration (OpenAssistant volunteers)Compare side by side →
- Human-written demonstration (OpenAssistant volunteers)Compare side by side →