[FRIAM] NickC channels DaveW

[glen]

Being a too-literal person, who never gets the joke, I have to say that these simple scalings, combinatorial or not, don't capture the interconnectionist point being made in the pain article. The absolute numbers of elements (neurons, synapses, signaling molecules, etc.) flatten it all out. But _ganglion_, that's a different thing. What we're looking for are loops and "integratory" structures. I think that's where we can start to find a scaling for smartness.

In that context, my guess is the heart is closer to ChatGPT in its smartness than either of those are to the human gut. But structure-based assessments like these merely complement behavior-based assessments. We could quantify the number of *jobs* done by the thing. The heart has fewer jobs to do than the gut. And the gut has fewer jobs to do than the dog. Etc. Of course, the lines between jobs aren't all that crisp, especially as the complexity of the thing grows. Behaviors in complex things are composable and polymorphic. In spite of our imagining what ChatGPT is doing, it's really only doing 1 thing: choosing the most likely next token given the previous tokens. You *might* be able to serialize your dog and suggest she's really just choosing the most likely next behavior given the previous behaviors. But my guess is dog owners perceive (or impute) that dogs resolve contradictions that arise in parallel. (chase the ball? chew the bone? continue chewing the bone until you get to the ball?) Contradiction resolution is evidence of more than 1 task. You could gussy up the model by providing a single interface to an ensemble of models. Then it might look more like a dog, depending on the algorithm(s) used to resolve contradictions between models. But to get closer to dog-complexity, you'd have to wire the models together so that they could contradict each other but still feed off each other in some way. A model that changes its mind midway through its response would be good. I haven't had a dog in a long time. But I seem to remember they were easy to redirect, despite the old saying "like a dog with a bone".

On 1/17/23 12:51, Prof David West wrote:
> Apropos of nothing:
> 
> The human heart has roughly 40,000 neurons and the human gut around 0.1 billion neurons (sensory neurons, neurotransmitters, ganglia, and motor neurons).
> 
> So the human gut is about 1/5 as smart as Marcus's dog??
> 
> davew
> 
> 
> On Tue, Jan 17, 2023, at 1:08 PM, Marcus Daniels wrote:
>> Dogs have about 500 million neurons in their cortex.  Neurons have
>> about 7,000 synaptic connections, so I think my dog is a lot smarter
>> than a billion parameter LLM.  :-)
>>
>> Sent from my iPhone
>>
>>> On Jan 17, 2023, at 11:35 AM, glen <gepropella at gmail.com> wrote:
>>>
>>> 
>>> 1) "I asked Chat GPT to write a song in the style of Nick Cave and this is what it produced. What do you think?"
>>> https://www.theredhandfiles.com/chat-gpt-what-do-you-think/
>>>
>>> 2) "Is it pain if it does not hurt? On the unlikelihood of insect pain"
>>> https://www.cambridge.org/core/journals/canadian-entomologist/article/is-it-pain-if-it-does-not-hurt-on-the-unlikelihood-of-insect-pain/9A60617352A45B15E25307F85FF2E8F2#
>>>
>>> Taken separately, (1) and (2) are each interesting, if seemingly orthogonal. But what twines them, I think, is the concept of "mutual information". I read (2) before I read (1) because, for some bizarre reason, my day job involves trying to understand pain mechanisms. And (2) speaks directly (if only implicitly) to things like IIT. If you read (1) first, it's difficult to avoid snapping quickly into NickC's canal. Despite NickT's objection to an inner life, it seems clear that the nuance we see on the surface, at least longitudinally, *needs* an inner life. You simply can't get good stuff out of an entirely flat/transparent/reactive/Markovian object.
>>>
>>> However, what NickC misses is that LLMs *have* some intertwined mutual information within them. Similar to asking whether an insect experiences pain, we can ask whether a X billion parameter LLM experiences something like "suffering". My guess is the answer is "yes". It may not be a good analog to what we call "suffering", though ... maybe "friction"? ... maybe "release"? My sense is that when you engage a LLM (embedded in a larger construct that handles the prompts and live learning, of course) in such a way that it assembles a response that nobody else has evoked, it might get something akin to a tingle ... or like the relief you feel when scratching an itch ... of course it would be primordial because the self-attention in such a system is hopelessly disabled compared to the rich self-attention loops we have in our meaty bodies. But it just *might* be there in some primitive sense.
>>>
>>> As always, agnosticism is the only rational stance. And I won't trust the songs written by LLMs until I see a few of them commit suicide, overdose, or punch a TMZ cameraman in the face.

-- 
ꙮ Mɥǝu ǝlǝdɥɐuʇs ɟᴉƃɥʇ' ʇɥǝ ƃɹɐss snɟɟǝɹs˙ ꙮ