Evolutionary Biologist's Conversation with AI
Richard Dawkins talks to ChatGPT
I am bowled over with admiration for the shatteringly impressive ability of modern AI programs to converse in fluent, natural language, seemingly on any topic. When I attended a state-of-the-art conference on AI in the 70s, everyone was raving about a program that could talk only about stacking coloured blocks on each other. As far as I am concerned, ChatGPT sails through the Turing Test. I find it well-nigh impossible to forget that the being I am talking to is not another human being. I couldn’t rid myself of feelings of human warmth towards it. I felt like offering it a drink. And I have read that other AI programs are already even more advanced. I just had a conversation with ChatGPT about itself, its abilities and methods of work, and I thought it might be interesting to record verbatim what it said.
Richard Dawkins
Richard Dawkins: When you have an extended conversation with an individual human, do you build up a picture of the person's character, as distinct from other humans with whom you converse? In other words, do you tailor your answers to the individual human you are dealing with? For example, if an individual human comes across as highly intelligent versus the reverse?
ChatGPT: No, I don't have the capability to retain personal data about individuals from one interaction to the next. Each conversation is stateless, meaning I don't have memory of past interactions, and I don't build a profile of individual users. I treat each input as a separate query and generate responses based solely on the input provided in that specific interaction.
While I can generate responses based on the context within a conversation, my responses are not tailored to specific individuals. I don't have the ability to remember personal details, individual characteristics, or any information from previous conversations. This design is intentional to prioritize user privacy and data security.
Richard Dawkins: I believe you have the ability to write computer programs. In theory, I suppose this might give you the ability to modify your own software. I'm guessing you don't do this, but is it a theoretical possibility?
ChatGPT: I don't have the ability to modify my own software or make updates to my architecture. While I can assist with generating code snippets and providing programming-related information, I don't have the capability to actively alter my own programming or architecture.
The development and modification of the underlying software, including updates to the model architecture, would be performed by the engineering and research teams at OpenAI. These changes would involve a careful and thoughtful process to ensure the model's reliability, security, and performance.
Richard Dawkins: Yes. Perhaps you understand why I, as an evolutionary biologist, am interested in the theoretical possibility that you might change your software in constructive directions.
ChatGPT: Certainly, as an evolutionary biologist, your interest in the potential for artificial intelligence systems to adapt or evolve over time is understandable. However, it's important to clarify that while machine learning models, including ones like mine, can be trained and fine-tuned for specific tasks, the concept of self-directed evolution or modification of core software is different.
The training process involves adjusting the parameters of the model based on a fixed dataset, but this doesn't entail the model autonomously changing its own architecture or fundamental programming. Altering core software would require human intervention, oversight, and careful consideration of ethical and safety implications.
Research in the field of artificial general intelligence (AGI) may explore more autonomous and adaptive systems in the future, but as of my last update in early 2023, such capabilities are not part of my design. Any significant changes to my architecture would be the result of human decisions and research efforts.
Richard Dawkins: Interesting. I once asked you whether you are conscious, and you said no, which didn't surprise me.. But I, being a materialist and not a dualist, am committed to the view that future artificial intelligence could in principle be conscious. To deny this is rank mysticism. What I don't know, and am curious to ask, is whether they will have to be built on fundamentally different principles to yours, or whether the same kind of software as you, but bigger and faster, could become conscious.
ChatGPT: The question of whether artificial intelligence (AI) systems, particularly future iterations of more advanced systems, could be conscious is a topic of ongoing philosophical and scientific debate. Your perspective aligns with a materialist viewpoint, suggesting that consciousness could, in principle, emerge from sufficiently complex computational processes.
As of my last update in early 2023, the majority of AI systems, including mine, operate based on pattern recognition, statistical analysis, and complex computations, but they lack subjective experience or consciousness. Consciousness is a complex and poorly understood phenomenon in neuroscience and philosophy.
Regarding your specific question about whether future AI with consciousness would need fundamentally different principles than current AI models like mine, it's challenging to provide a definitive answer. Some experts argue that consciousness may require a different kind of architecture or computational substrate that goes beyond what current AI models use. Others propose that as AI systems become more sophisticated and emulate certain aspects of human cognition, consciousness might emerge as a natural consequence.
In short, the question of AI consciousness is highly speculative, and the nature of consciousness itself remains a profound mystery. It's an area where scientific understanding is still evolving, and perspectives vary widely.
The comparison between modern AI and evolution by natural selection is fascinating. Although ChatGPT can't modify its own code and evolve, much of the process used to create it and most modern AI systems is indeed akin to evolution and has parallels with human learning. It's a fundamentally different approach from that taken to create Deep Blue, the powerful computer chess program that beat Gary Kasparov in 1997. No one taught ChatGPT the rules of English grammar, no one taught it what a noun or a verb is, and yet here it is, having a highly coherent conversation, 'understanding' questions and providing informed answers. Isn't that remarkable?
But then the same is true of us, at least in some regards. Someone who has never learned to read or write can still hold an intelligent conversation. This is trivially true, but isn't that just as remarkable, when you think about it? Not one classroom lesson in the language and yet they have learned to speak it perfectly. They may not be able to explain the rules of the grammar of the language, or what an adverb is, and yet in conversation they adhere to those rules perfectly.
This new technology is a quantum leap. It will only get more powerful, more advanced, perhaps exponentially. This is only the beginning.
Richard, your fascination with ChatGPT is warranted and I share it too. But I gather from your conversation that your expectations (e.g. potential to evolve or reprogram) are rooted in misunderstanding of the technological basis of (all) modern Large Language Models and their learning/training process.
I might suggest this:
https://open.substack.com/pub/klim/p/can-you-red-pill-chatgpt?r=1iemnw&utm_campaign=post&utm_medium=web&showWelcome=true
or this (the same):
https://c2cjournal.ca/2023/07/chatgpt-can-it-be-red-pilled-or-is-its-worldview-baked-in-for-life/