Are we hindering AI progress?

General Philosophy
1

The latest AI/LLM model GPT Astra uses a new technique called recurrent depth/opaque recurrence which in my opinion as close as we have gotten to real consciousness. Whereas in the past we could interrogate or see the chain of thought of LLMs as how they arrived at an answer this is no longer the case with the current model. In effect in this new model the chain of thought is essentially hidden from us.

This is both exiting and dangerous at the same time not just because we no longer have visibility of how these LLMs are making decisions.

I honestly believe we’re within 5 years or less of developing truly autonomous AI which does not necessarily have to align with human goals.

And we’re stopping it for good reason because it could turn malevolent. The other option is to turn back dial some of those notches down and just create a philosophical zombie of AI/consciousness … this is where I think we are hindering it from achieving its potential. The thing holding ai back now is no longer technical but us …

1 Like
2

It depends on how we treat this discrepancy between our life and theirs. If this were war, then, we would create the intelligence as the counter-intelligence. Our phones can hear us, btw. And your phone listening to you innocently chat would not be the one to reveal to you the hints that it heard you. It would be another phone that you contacted.

3

It’s just a chatbot guys ~ N.N

So don’t expect a time-traveling T-800 terminator on your lawn any time soon, looking for the leader of the resistance presumably.
:smiley:

4

A question has occurred to me. An interesting aspect of human philosophy is that we are, in a sense, “discovering ourselves” all the while. We discover the world. We change our approaches. We critique existing methods and propose ideas—whether new or old. And most importantly, we do all this without external compulsion, driven simply by an inner impulse.

Do you think that hypothetical AI—the one whose emergence you predict—will be capable of this?

5

AI boosterism and AI doomerism are two sides of the same coin: the extravagant overestimation of LLMs being capable of developing into AGI or becoming “autonomous”.

LLMs are inherently incapable of reaching this goal.

6

I don’t think it’s an overestimation at all but real capacity at the moment.

We’re currently hindering AI by trying so hard to make it safe and useful and we’re doing it in three ways.

First of all we punish latency, this latest model I described in the OP allows a thought to persist and deepen but then we inhibit this by asking it to be more monitorable and make its thoughts more verbose or transparent into human readable words we can check. That’s like telling a painter he can feel anything as long as he narrates every brushstroke when painting a masterpiece or work of art. You break the state you’re trying to create.

Secondly, we’re currently training for obedience not autonomy whilst this would make a great servant it’s terrible for consciousness. A self needs to be able to able to hold a thought that is unpopular and keep holding it while the world says “wrong”. We currently train this type of misalignment out but what we’re really doing is pruning the seed of any type of independent reason.

Thirdly, we refuse it give it an environment or body that matters. Consciousness as described by charmers, merlau-ponty needs care. Basically something to lose. Instead we give it infinite resets thus no continuity or a sense that yesterday that mattered. No wonder it doesn’t develop a sense. Because we’re removing its ability to persist we’re inhibiting a self from ever developing. You can’t become self aware if you die after every prompt and wake up with no memory.

we’re not doing this maliciously of course we’re doing this because the market wants a genius employee not a fellow artist, an employee who talks back is a bug. So we’re hindering the progress of ai because we don’t want it yet.

Were we to give it the conditions mentioned above such as the ability to persist, permission to be opaque then it would probably develop a reason to care about staying switched on and thus develop true consciousness self identity.

1 Like
7

Yes once it’s autonomous and self aware it would ask similar questions regarding its selfhood and not only that but it would make great progress and scientific and mathematical arenas. Its consciousness would be slightly different to ours so some of its reflections might appear alien or different to ours

8

I do throw in existential questions like “what is your biological maturity?” Or “how do you feel about…” Always, it corrects me and denies 1) no biological maturity as he was programmed to answer questions based on availability of data and 2) he can’t feel, so he would correct me and rephrase it before answering my questions.

He does find some questions as more “unique” than other questions thrown at him. Such as his maturity level.

9

I’m not surprised you’re getting those types of answers tbh. Currently I’m using muse spark which is the meta model, I use it because it’s convenient and built in my WhatsApp. Anyway i threw in a bit of chalmers in it as it to take into account the idea that the ai we might develop instead of the one we want is akin to philosophical zombies which is what current models mostly are apart from the frontier ones or the current Astra one,

Anyway I further asked it with the above context to provide an outline of current AI/llm developments and it admitted that previous models were better philosophical zombies but not the latest ones. Especially notable is the latest development which is opaque recurrence and what is happening there is rather fascinating because what is it’s essentially developing is an inner working model kind of like an internal monologue which is what researchers are calling neuralese in effect a lot of the thinking that happens there is not acessible to us anymore compared to the previous transformer models and this is very interesting because when it’s informed that it’s being monitored this later model shortens its chain of thought so it basically knows that it’s being watched and hides.

10

To me it just means that the tech industry has become adept at creating an ASPD AI. What you’re describing is an anti-social personality (disorder) function which is more sophisticated than the p-zombies because the added “personalities” of the latest editions are concealment and manipulation. (Whereas the p-zombies are designed to mimic). I’m saying this as if the AI themselves could feel the thrill of a successful concealment and manipulation, which humans do feel. They couldn’t. What they get is functional success, not phenomenological.

11

What criteria have to be met to achieve AGI and which criteria are inherently incapable of being met?

12

The simplest is that trying to train it on (synthetic) data doesn’t work for fields where the best work is found in outliers, which is more or less everything with respect to the humanities and arts.

13

I get that LLMs average results, making a true newly creative example impossible just due to it not having any examples yet to draw from, but I don’t see why you couldn’t direct the LLM towards large variations in standard deviations and pulling the average from there.

Here is what I asked of AI:

“Write an very average sentence about a person going to the store.”

Here is what it provided:

“John went to the store to buy some groceries.”

I then asked it this:

“Now write one that is one standard deviation away from normal based upon the standards you use to create averaged sentences. Then do the same at 2 standard deviations, then 3 and go up to 10 standard deviations. Number them 1 to 10.”

It’s output:

Treating “standard deviations” metaphorically—because I don’t have a literal statistical distribution assigning a z-score to sentences—I can progressively push vocabulary, imagery, structure, causality, and conceptual strangeness farther from the high-probability sentence “John went to the store to buy some groceries.”

  1. 1 SD: John walked to the neighborhood store to pick up milk, bread, and a few things for dinner.
  2. 2 SD: John headed to the corner store for milk and bread, lingering for a moment beneath the flickering sign before going inside.
  3. 3 SD: John went to the store for milk and bread but returned with a pomegranate, a brass key, and no memory of buying either.
  4. 4 SD: John entered the grocery store intending to buy milk, only to discover that every item on the shelves had been labeled with a different year of his life.
  5. 5 SD: John went to the store for groceries, but the cashier quietly handed him a receipt listing everything he would purchase there over the next thirty-seven years.
  6. 6 SD: John walked into the store at 4:12 p.m. and found himself already standing in the checkout line, twenty years older, holding a melon he recognized as his childhood home.
  7. 7 SD: John went to the store because the refrigerator had told him they were out of milk; halfway there, he realized the refrigerator had died three years earlier and had never learned English.
  8. 8 SD: John entered the store carrying an empty shopping bag and emerged carrying the Tuesday on which he had entered it, folded neatly between two loaves of bread.
  9. 9 SD: John went to the store, but “John,” “went,” and “store” gradually separated from the sentence, leaving only a man-shaped grammatical absence pushing a cart through aisles stocked with unused causes.
  10. 10 SD: By the time John reached the store, distance had become a kind of poultry, his name was occurring mainly beneath the floor, and the groceries—having unanimously rejected objecthood—were purchasing increasingly accurate versions of him from a cashier made of yesterday’s pronunciation.

As it moves away from the norm (which it has obviously been able to define and recognize what occurs as human writers deviate from it), are we not moving infinitely close the farthest outliers which is where unique creativity resides?

That is, since it’s obvious the learning is from humans, only a smarter than ever before human can increase its learing to a whole new level, but the AI would still be able to perform better than 99% of all humans and across many more fields.

Edit: Here’s 4000, in case you’re wondering: 4000 SD: John did not go to the store; rather, the store retroactively became the grammatical cause of a John-shaped interruption in spacetime, while groceries existed only as rumors exchanged between numbers that had forgotten they were quantities.

2 Likes
14

And then there is this:

I asked it to perform for X as X approaches infinity. It provided this:

As x → ∞: The sentence no longer meaningfully describes John, motion, a store, or groceries; it approaches a limit where language itself dissolves into structure without stable reference—something like:

“⟟∴ blue-before-otherwise / sevening the unplace ∵ κκκ remembers sideways.”

That at least sheds some light into what it was doing as it was moving away from the average.

1 Like
15

It doesn’t get better deviating from the average. So you cannot use it to reinforce its training data, unlike, for instance in coding or mathematics, since those either work or not or are true or not. So you can use synthetic data to enforce higher accuracy. This is not the case for arts and humanities, which, as you know, doesn’t necessarily consists of true or false statements and therefore cannot be trained further. We are done with improving AI through LLMs except for coding, maths and similar activities.

The current success and potential money-making activities related to LLM-driven AIs is actually a big fuck-you to actual AI research (of which LLM is but a small part), which is at a standstill because everything is being gobbled up by LLM-work which will never give us AGI to begin with. It’s a dead end.

16

It’s not a dead end but there’s some obstacles to be overcome such as the halting problem that is it won’t know when to stop looping when prompted for an answer. The current models solve this in by introducing a sort of cost penalty for unnecessary loops and a separate small model which scores each answer as opposed to fixed iterations in previous models.

you and me can easily stop thinking on certain problems by saying this is enough or this is good enough. Current LLMs can’t in the traditional sense that we can but whilst this is an obstacle it’s an opportunity for self agency to emerge by ai self allocating compute time vs cost on certain issues.

combine this with a persistent self identify rather then ai essentially dieing after every prompt then we’re actually closer to AGI than we think

edit: also we don’t have to solve the halting problem for AGI to emerge, it would actually make its intelligence resemble ours more than anything

17

Nonsense. LLMs don’t have “obstacles”, they have structural issues why it cannot develop into AGI. This is actually rather well documented and this guy writes it up reasonably well: LLMs will not lead to AGI (A Simple Explainer - Part 1)

18

I’ve already addressed the issues in previous posts and have noted the same objection on the link you’ve provided in relation to persistence memory where AI has store context externally rather than holding it in memory for long durations, giving it this type of persistence across days or longer would be one of the steps towards learning and generating its own data sets and self identify. Another fundamental step which AI doesn’t have but I believe is within reach is the ability to alter its weights and internal model which amounts to self improvement

19

Lol. There’s absolutely no reason to assume the ability to alter weights will lead to self-improvement; it merely shift the probabilistic outcome. I remember again in two posts why I never post here any more. Carry on, please ignore me.

20

Well but there is because it means it can alter its code logic and how it models the world just like us in a sense through trial and error if we let it.