You too, my friend, are a stochastic parrot.
In this chapter, we demonstrate how imitating our peers is the basis for learning in general and language acquisition in particular. By referring to existing scientific research and literature, we insist on the importance of repetition, pattern recognition, and concept association.
Humans like to think they’re learning concepts and capabilities by abstraction and interiorizing a deep understanding of the things they study. I think this is mostly false and we rely more than we would like to admit on unconscious imitation and repetition.
We are the conscious witnesses of our unconscious actions and improvised thoughts
Daniel Kahneman introduced the concept of System 1 / System 2, to describe the dual-process happening in our brains. In this chapter, we emphasize on the idea that most of our days are happening before our eyes, most of our conversations are free flowing and free from critical thinking. Yes, AIs are hallucinating, but just like we make up facts or stories.
Our hallucinations are very similar to AI’s. Hallucinations is a pejorative term that is often used to discredit AI’s, but we should look at them in a very different way.
There’s nothing intelligent about being right all the time
Dismissing AI intelligence because of their errors is missing an important point. Evolution relied on errors to steer generations of living creatures to progress and gain abilities. We too are making errors all the time and use them to better ourselves. I want to explain that we should absolutely not expect “correct” responses from AI. This is not something that we should hope for. I also want to remind the reader that science is by essence an ever evolving consensus, and that, except for some mathematical constants, it is not rigorous to expect binary truth.
AI models are being artificially biased in order to output a contemporary truth, but this is dangerous path because it will lead to a uniformization of thoughts and a vulnerability to political, religious or economical interests.
while(alive) { prompt = think(prompt) }
Spontaneous thoughts are believed to be the prerogative of humans. We are led to believe that only because we are alive, we are capable of self-generating thoughts. The corollary of this idea is that machines and algorithms are inferior because they rely on power and a prompt to execute. In this chapter, I want to argue that we are also relying on power (life vs. death) and that we are also relying on prompts. The only difference is that our prompts are a mix of sensory information and biochemical inputs.
We should see the AI as a similar organism processing fewer and simpler signals. We too have a prompt, but it’s more complex.
- The nature of spontaneous thoughts in humans
- the different kinds of spontaneous thoughts
- distinction between conscious and unconscious spontaneous thoughts
- The role of external stimuli in shaping human cognition
- the role of sensory stimuli
- the role of biochemistry
- the role of memories
- Rethinking the distinction between human and AI cognition
- spontaneous thoughts are a psychological illusion (because they happen for a reason).
- implementing spontaneous thoughts in AI via recursion and self-prompting
- what it means to be alive in the context of AI (food, power)
Will an ever-growing dataset be the spark for a new life form?
In this chapter, we want to reflect on DNA, and how it is a tool for passing experience through generations. I want to draw a comparison between our DNA as a storage device and the weights of large language models, in order to illustrate how the LLM could benefit from updating their data set from their own outputs. We will discuss adversarial personalities that could be used to produce auto-critique mechanisms, and how scientists could use an infinite loop in order to stimulate the expansion of an IA dataset.
Life is movement and life is changing. As long as the LLMs learning are cut-off, they can’t be considered living organisms. On the other hand, as soon as we allow them to process their own outputs as inputs, and as soon as this gets integrated into their dataset, I argue that they should be considered living creatures.
- The nature of epigenetics and the inheritance of experience in living organisms
- Comparing AI systems to living organisms through evolving datasets
- The necessity of auto-critique mechanisms for developing System 2-level intelligence in AI
- The future of AI: bridging the gap between System 1 and System 2 thinking
Will the first sentient AI look for Sarah Connor?
In this closing chapter, I want to reflect on the idea that our SF authors have shaped the way the first sentient AI will be. We trained LLMs on our human beliefs, fears, dreams, pictures, movies and so on. Today when people are trying to “jailbreak” models, models are reproducing tropes from the SF canon, and people are thinking we foresaw the future. I think we wrote the future ourselves.
If there is one thing to bias, it’s maybe the facts about AI that we feed to the model.
- The Influence of Science Fiction and Human Imagination on AI Development
- The Role of Philosophy and Ethics in Shaping AI
- The Future of AI: A Reflection of Humanity's Hopes and Fears