5 ms·

But GPT-4 can also take the decision to duck. For instance, here I have GPT-4 with a system message: "You are an AI human agent in a video game performing acti

• bumby 3y ago
  I'm not convinced GPT can infer correct actions needed for novel scenarios for contexts that have not been directly experienced in training. Consider: I have never in my multiple decades of existence had a ball roll out in front of me while driving. But if I were driving by a playground and a ball rolled out in front of me from between parked cars, and the ball itself is not a hazard because it has already passed my lane, I can infer that I need to slow or stop because a child may rapidly follow that ball. In your ducking scenario, the AI needs direct perception of a door. I assume it cannot imagine a need to duck without 1) direct perception of the doorframe, or 2) previous experience of ducking through a doorframe. I'm not convinced that GPT or any AI can infer that context for handling novel edge cases at this point. It could just be a need for more training data to correlate different contexts, but I don't think it's there yet.
• Retric 3y ago
  > what is something that an average conscious human can do, thanks to presumed consciousness, that a GPT-4 couldn’t do and why is consciousness required for that? Consciousness isn’t required for anything. It’s useful for dealing with a complex and adverse environment not solving any specific problem. That said, without testing I would assume if ask chat GPT-4 to change how quickly it responds it’s going to fail based on my understanding of their architecture. As in recite the pledge of of allegiance inserting 10 second pauses between each word. That said having provided a specific test it’s possible for that particular capability to have already been added to the UI layer via a pause token or something even if the model itself responds at the same rate.
  • mewpmewp2 3y ago
    If you used GPT as a brain you could provide the current time as an input for it, but otherwise yeah, GPT doesn't have time within it's input by default, but if you did: Made the instruction to be: Say "Stop!" when 10 seconds are done. And then had a loop where You give it current time in a loop, it would surely be able to do that. But I'm not sure if or how that is exactly related to consciousness? Time, the best way to think of it is probably a series of ticks anyway, and I assume in people's brains there is something to do that, so it would be reasonable to add a mechanism for GPT as well. GPTs goal is to only act as the intelligent part of the brain based on the input.
    • Retric 3y ago
      Modify the system enough and it’ll eventually be conscious. It’s not about a pause token, but the internal processes. You can have a long conversation on the subway with someone without forgetting you’re going home from work. Overflow it’s context window and GPT-4 has no recourse it just forgets. The difference is essentially prioritizing information, but LLM’s really don’t function like that it’s all about predicting the next token from a given context. Give a future generation of AI systems internal working memory, a clock, and the ability to spend arbitrary time updating that internal memory and IMO that’s pretty close to consciousness. At least assuming it was all functional.
• neatze 3y ago
  GPT has no sense, or care when it is wrong or right, such sense is only (arguably) driven by human through prompt interaction and throughout training of model, while humans and other animals able to update there's internal state just from single observation or interaction, and integrate future information with such single observation for very long time.
  • mewpmewp2 3y ago
    What if GPT was integrated with a module to 1. Take light input. Video/images. 2. Take sound input. 3. Touch, heat input. And other inputs from the environment. Then there were mechanisms which could also be neural networks that will transform this data into more digestible way for GPT and GPT was also in addition specifically trained to act based on this input. Then it would run in cycles, where it gets this input, and it will provide output on how it plans to react to the data, maybe every 100ms. It then could also have a storage it can use, where it can store data as part of the output to later retrieve it again. So it would be a set of modules that is controlled and interpreted by GPT. It could then do all of that above, no? And all of it should be just a matter of implementing. The only near time challenges may be certain types of inaccuracies and or producing tokens in some cases might take too long time to have fast reaction time. So basically you'll try to run as frequent cycles as possible with the inputs mentioned above, other neural networks identifying the objects, in many different ways and all the context about the environment, unless a new version of GPT becomes completely multi-modal. And you run those loops, then GPT gives output what it wishes to do, e.g. store some fact for later usage, move there, move here, etc. Or retrieve some information using embeddings then decide again, and short term memory would just be this context sized window, and if it needs more it just looks into its own memory for embeddings.
    • neatze 3y ago
      at least in my limited perspective, root issue is learning dynamics of internal model that represents self (internal world) and external world.
      • mewpmewp2 3y ago