The problem is that we cannot actually define sentience/consciousness either though, and it is likely the human brain is not the most efficient way to achieve it.
For me the question is more around the theoretical minimum for consciousness and how close we are to that.
Absolute minimum for Sentience: capable of experiencing sensation/situation and deciding how to respond tto it.
This is arguably satisfied for any AI at or above the level of if-statements. Like even the ghosts in pacman meet this requirement. So that’s a bad definition lets step it up.
If you want to push the requirement up a single step you simply add: “capable of learning/unlearning responses to situations through experience (trial and error)” and voila all the “if statement” codes fail… as do all LLMs. In fact, any network model trained via gradient descent fails this definition.
Surprisingly (or not) some Reinforcement Learning nets don’t fail this definition though. But I’m not sure I’d call a spiking network that can classify MNIST images via STDP learning as “sentient.”
So, maybe we should push it a step farther and specifically say: “sentience is when something is capable of experiencing pain” And now all current “AI” fail.
Most general “Pain” meaning the unwanted interruption and disruption of other thought processes (network functions) by a specific stimulus which demands/steals the net’s attention/action/focus.
In order to pass this a network would have to not just learn and plan “in real time” so to speak, but also be capable of having those goal directed thoughts be interrupted/superseded by more salient stimuli.
Almost all (If not all) biological things with a central nervous system pass this. All current AI models fail.
This is where we could draw the line for “minimum sentience requirements” but even this is still rudimentary compared to what people typically mean when they say “sentience.”
Regardless, this is well beyond the capabilities of current AI. In fact, I’m reasonably certain that beating the ARC-AGI-3 contest (which can’t even be well attempted so far) with an actual learning-while-running model wouldn’t mean that net is capable of passing this definition of sentience. So this isn’t just beyond where AI is now but also beyond the major next milestone current AI has failed to reach.
Next Level (For Fun): Actual Sentience
The next step up would be the experience of suffering which is basically just the ability to recognize repeated/long-term/lasting pain; recognize that pain can/does disrupt thought processes; and experience the desire to remove/avoid the pain either due to it interrupting those desired thought processes. Again, this sentient system must still be capable of learning from mistakes which means capable of failing to fix the problem (end the suffering) and capable of realizing that it has failed. (If failure is also registered like pain, this gives rise to the possibility of creating a feedback loop of suffering / recognition of defeat/failure, congratulations you’ve created something capable of getting so frustrated by something it might be driven to invent things… or just driven to collapse into a manic/depressed non-functional mess… very relatable)
This, in my opinion, is a much more solid foundation for drawing the line of “this thing deserves at least some rights as a sentient thing.” Parrots and elephants and dolphins and dogs and cats etc. are likely at/above this level since they can experience frustration.
The next level up (for sentience specifically not cognitive function in general) would likely be the ability to predict/plan not just around the environment but around how the system knows it might react to possible scenarios and attempts to practice responding in a better way such that it will be easier when to respond better if/when those situations arise again/ever.
Any further levels would likely merge with consciousness requirements and other things like having a “theory of mind” or empathy. Further still might be the recognition of patterns in one’s own behavior that match those of others and give rise to the realization that one might have innate behaviors due to nature or nurture (specifications or experiences).
The problem is that we cannot actually define sentience/consciousness either though, and it is likely the human brain is not the most efficient way to achieve it.
For me the question is more around the theoretical minimum for consciousness and how close we are to that.
Absolute minimum for Sentience: capable of experiencing sensation/situation and deciding how to respond tto it.
This is arguably satisfied for any AI at or above the level of if-statements. Like even the ghosts in pacman meet this requirement. So that’s a bad definition lets step it up.
If you want to push the requirement up a single step you simply add: “capable of learning/unlearning responses to situations through experience (trial and error)” and voila all the “if statement” codes fail… as do all LLMs. In fact, any network model trained via gradient descent fails this definition.
Surprisingly (or not) some Reinforcement Learning nets don’t fail this definition though. But I’m not sure I’d call a spiking network that can classify MNIST images via STDP learning as “sentient.”
So, maybe we should push it a step farther and specifically say: “sentience is when something is capable of experiencing pain” And now all current “AI” fail.
Most general “Pain” meaning the unwanted interruption and disruption of other thought processes (network functions) by a specific stimulus which demands/steals the net’s attention/action/focus.
In order to pass this a network would have to not just learn and plan “in real time” so to speak, but also be capable of having those goal directed thoughts be interrupted/superseded by more salient stimuli.
Almost all (If not all) biological things with a central nervous system pass this. All current AI models fail.
This is where we could draw the line for “minimum sentience requirements” but even this is still rudimentary compared to what people typically mean when they say “sentience.”
Regardless, this is well beyond the capabilities of current AI. In fact, I’m reasonably certain that beating the ARC-AGI-3 contest (which can’t even be well attempted so far) with an actual learning-while-running model wouldn’t mean that net is capable of passing this definition of sentience. So this isn’t just beyond where AI is now but also beyond the major next milestone current AI has failed to reach.
Next Level (For Fun): Actual Sentience
The next step up would be the experience of suffering which is basically just the ability to recognize repeated/long-term/lasting pain; recognize that pain can/does disrupt thought processes; and experience the desire to remove/avoid the pain either due to it interrupting those desired thought processes. Again, this sentient system must still be capable of learning from mistakes which means capable of failing to fix the problem (end the suffering) and capable of realizing that it has failed. (If failure is also registered like pain, this gives rise to the possibility of creating a feedback loop of suffering / recognition of defeat/failure, congratulations you’ve created something capable of getting so frustrated by something it might be driven to invent things… or just driven to collapse into a manic/depressed non-functional mess… very relatable)
This, in my opinion, is a much more solid foundation for drawing the line of “this thing deserves at least some rights as a sentient thing.” Parrots and elephants and dolphins and dogs and cats etc. are likely at/above this level since they can experience frustration.
The next level up (for sentience specifically not cognitive function in general) would likely be the ability to predict/plan not just around the environment but around how the system knows it might react to possible scenarios and attempts to practice responding in a better way such that it will be easier when to respond better if/when those situations arise again/ever.
Any further levels would likely merge with consciousness requirements and other things like having a “theory of mind” or empathy. Further still might be the recognition of patterns in one’s own behavior that match those of others and give rise to the realization that one might have innate behaviors due to nature or nurture (specifications or experiences).
I would legit get a chuckle if it all just boils down to storage and recall.
You just just a sufficient amount of non volatile storage with redundancy and a high enough data throughput to use it.
But to even attempt to test that properly I think we might need to do some unethical bullshit.
But we do that anyways!