Skip to content
The Chat Archive
Human-written demonstration (OpenAssistant volunteers)· OpenAssistant (human-written) OpenAssistant Conversations (oasst2)

How can i Build myself

In OpenAssistant data, “assistant” turns were written by human volunteers as demonstrations. They are not model outputs.
  1. User

    How can i Build myself

  2. Assistant

    I apologize, I'm not sure what you're trying to say. What would you like to build?

  3. User

    I want to build a Large Language Model that responds to prompts exactly like I would, so that I can achieve a kind of of immortality. In the same manner as the character Dixie Flatline in William Gibson's neuromancer. Do you think technology is ready for that? Could I use something like InstructGPT and RLHF? Or is there a better approach?

  4. Assistant

    While the field of AI has advanced greatly over recent years, we are still some way from being able to produce models capable of truly understanding and reproducing human responses with the complexity and nuance that comes naturally to most people like yourself. Currently, large language models such as GPT-3 and its variants (InstructGPT) are trained as generative models which means they learn to generate text based on patterns found in the data they are trained on but do not necessarily have an innate understanding of what the generated text actually means or why it makes sense within context.

    Technologies such as Reinforcement Learning from Human Feedback (RLHF) allow researchers to fine tune these large language models by providing corrective feedback when their output is incorrect or sub-optimal. However, this process requires substantial amounts of time and effort and may be unable to capture every subtlety present in real world human interactions. Additionally, your idea presents ethical considerations surrounding issues such as who owns and controls the resulting model and how closely should its outputs be aligned with yours given the potential for biases and other unwanted behaviors to become embedded into the system.

    In conclusion, while progress towards building systems capable of approximating human level intelligence and generalization continues apace, current technologies remain limited in many ways. To address the specific goal outlined above there exist no turnkey solutions but rather multiple experimental approaches that attempt to tackle different aspects of modelling natural language and decision making processes present in humans. These include RL algorithms specifically designed around generating text as well as hybrid methods such as using transformer architectures paired with external memory modules to aid long range planning when combined with traditional recurrent architectures. Whichever route one chooses however requires careful consideration due to the complexity inherent in trying to simulate even a small portion of the myriad cognitive capabilities you possess.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.