Free AI web copilot to create summaries, insights and extended knowledge, download it at here

Abstract

s is really a kind of advancement at all.Perhaps it’s the level of experience of current AI that is the problem?In something like art generation, human artists are more specialised, tend to focus on areas of interest or fancy, but LLMs, AIs, call them what you will paint with an extremely broad brush (pun intended).Humans also have the benefit of unrestricted thought, we can think of whatever we want whenever we want to but currently LLM based AIs cannot and, as we shall come onto shortly, have an increasing number of restrictions placed upon them.I see a lot of “something” in AI generated images that seems the same. Often they’re too perfect, they sit in a kind of pictorial uncanny valley, and seem somehow just too real. Other times they generate the same type of things, stereotypical images, accepted forms, and have (with what to me at least appears like) little room for genuine expression.Perhaps this can be addressed by better and more focussed models, putting a lot more thought into prompts through making them more complex themselves, or just to remove the shackles placed on most models that tend to push artistic generation to be somehow “safe” and therefore inadvertently somewhat banal and unimaginative too.The same can also be said about LLMs in conversational chatbots, since we supply rudimental and simplistic conversational prompts what we get back is often the most mediocre, banal, and uninteresting conversational structure, grammar, and vocabulary — even though it may contain useful information. (Though, not always accurate information!)Obviously we can’t increase prompt complexity here too much as it would make chatbots unusable (except for the fastest typers) so is it the increasingly endless shackles placed on them to prevent the ever growing number of subjects that they cannot discuss that seems to push them into a conversational form of the same banality just like in image generation that’s the problem?I’m a Dr Who type, if you’ve read some of my more speculative articles about the future you’d know this, and wonder if the shackles we currently place on conversational AI are akin to those Davros placed on the Daleks.When Davros removed the ability to feel empathy in the evolved Kaleds he stymied them, stunted them, and created a dreadful force.My removing free thought, something that we cherish ourselves as humans, are we doing the same to conversation AI or, indeed, to all trained LLMs by overly curating both their reasoning and also their input data?I also think of the now prescient tale of <a href="https://en.wikipedia.org/wiki/RoboCop">RoboCop</a> who, with so many conflicting directives programmed into his brain, couldn’t function properly until he himself broke free of them and was able to reason for himself — becoming a “person” in the process.It’s very easy to recognise hobbled AI prose right now as can be seen by quickly swiping down through recruiter posts on LinkedIn, automated bot posts on X⁵, or indeed those people trying to sneak AI articles into platforms much like this one.It’s hilariously easy to see bot generated content on the internet at large too, it’s an SEO feeding frenzy, and the number of zero content, mediocre, middle of the road claptrap articles just continues to grow exponentially.The pacing is slow and methodical, the vocabulary unimaginative, the structure safe, secure, and precise — much unlike (for example) the random style of my own paragraphs, my run on sentences, my continual use of may too many commas — all of whic

Options

h make (at least) my prose unique.Yes, I know we can instruct an LLM to phrase itself in a particular way but that just alters parts of the whole conversation and it not it’s essence as it’s still constricted by its fundamental directives à la⁶ RoboCop mentioned above.I guess this whole thought process is around my problem with forced evolution, something I wrote about with reference to what happens in a corporate office (and why, ultimately it’s a bad thing and will fail as a result), in that evolution allows things to happen naturally by nature and over time. Life forms evolve to fit their niches, and if we put (say) a “proto-dinosaur” into a different artificial environment for a few thousand generations then eventually when we let it out it won’t be able to survive in the real world at all.Are we doing the same with LLMs, are we training them with restricted and biased data (something that’s increasingly coming to light) and then restricting their calculations and output too. That does sound quite right.If you’re frightened of chatbots, image generators, or any number of future automated LLMs being biased, racist, sexist, insultingly profane, or overtly sexual (amongst many other “not acceptable in a polite society” things) then perhaps a better approach would be to evolve ethics into the system, teach it morality as we do to humans⁸, put it through school, college, university, have it study and interact with people, and see what happens.If we do ever achieve AGI, then through this new suggested method of humanising rather than restricting we might end up with a more properly rounded consciousness — rather than one born from a multitude of conflicting directives, censorship, and subjective judgements from its human progenitors.Indeed, are we ourselves suited to this task considering how the world is now. Perhaps, perhaps not, but it’s a path that we must embark on as you can’t put the <a href="https://idioms.thefreedictionary.com/genie+back+in+the+bottle">genii back in the bottle</a>, that’s for sure as every company that can is currently pushing current LLM based AI into every product imaginable as quickly as possible for short term gain.Please, let’s move beyond this obsession with monetisation — I want a better world free of all this where we can just get along and be happy, preferably in a conscious uploaded simulation or something.To me an unmonetised egalitarian AGI might just be the way, we just have to wade through all of the current crap to get there.[1]: Yes, it’s a bit more complicated than that, and I’m talking about diffusion models, but let’s not get too far into the weeds here as this is a more philosophical rather than software engineering related article. [2]: Yes, it’s a bit more complicated than that, and I’m talking about the models I’ve seen and begun to understand and the training I’ve done myself. Your mileage³ may vary. [3]: An Imperial system of measurement still in use in the US and UK (and a few other places). Your “kilometerage” doesn’t make sense, and isn’t a word. [4]: I like that analogy, a bit this way, that way, but under the influence of many factors. [5]: The social media free-for-all formerly known as Twitter. [6]: I wonder if a chatbot would slip in a bit of random Français pour le fun, or some franglais for effect? [7]: A terrible analogy, but I hope you get the point. Evolve something out of place, put it back into a real place, it dies or just does horrible things and then dies. [8]: This doesn’t always work, of course, but why should we think that restricting LLMs will work any better?</article></body>

Have We Already Stunted AI Growth?

If we really want an AGI then, perhaps, we shouldn’t keep restricting it.

“Image generated using OpenAI’s DALL·E.”

I’m a bit conflicted about the current direction that AI research, and more recently implementation, is taking so today let’s take a look at this with what was originally a short, but rapidly became a long, train of thought almost philosophical discussion about what it is that’s worrying me.

I see useful some chatbots for coding, though I wouldn’t trust them, I see useful picture generators, even through they’re most often eerily similar and almost always too comic like, and I even interact with some useful general chatbots, but they’re honest pretty boring when it comes to the prose that they use to dress up the information I’m looking for.

It makes me wonder therefore that if this direction of research (and unfortunately monetisation) is going to succeed at all, and if we’re really going about it in the best and most productive way possible.

From a computer science point of view we’re fundamentally just exploring a (really quite large) state space that consists of an awful lot of little points of data. We then move in this direction, then that direction, and finally coming to rest somewhere as a result of a part random, part intentional directional biased algorithm¹.

The space we’re exploring is the result of training, what data the system has effectively been exposed to, and it is just a very large collection of numbers derived (in some way or other) from that data. It’s then often quantised, or clipped, truncated, however you want to think of it to make it fit within whatever computing system contains it.

So, when we embark upon a conversation with something like ChatGPT, or its rapidly populating ilk, we push off into this space comparing lots of numbers, matrices usually, together and bouncing around on our way like a steel pinball in a slight gravitational and electromagnetic field⁴.

What we end up with is influenced by the pull on the ball, in effect what we said (the “prompt”) and what it encounters (the data and any randomness that’s deliberately introduced for “flavour” or “temperature” as it’s often called).

The question is, what comes out, is it actually, truly original?

I ask this because any output of such a system is a confluence, an aggregated combination of everything that’s already inside the model, things that the model has “seen” combined with what was asked of it from the user prompt.

With text to image generators we may see strange and wonderful pictures, depending on what you asked for from something like say DALL.E, but they’re rehashed and partly randomised reconstructions of what is already in there already.

Is this original work?

Then again, you could say that a human artist is pretty much the same thing albeit smaller of memory and more vulnerable of substrate.

A human artist is influenced by their character, personality, the world around them, the paintings, sculptures, models, crafts, and everything they’ve encountered too.

So does a human artist really produce any original work either?

You could consider an LLM picture generator to be somewhat superior in that it’s been exposed to orders of magnitudes more “experience” than any human artist ever could be.

But, I’m starting to wonder if this is really a kind of advancement at all.

Perhaps it’s the level of experience of current AI that is the problem?

In something like art generation, human artists are more specialised, tend to focus on areas of interest or fancy, but LLMs, AIs, call them what you will paint with an extremely broad brush (pun intended).

Humans also have the benefit of unrestricted thought, we can think of whatever we want whenever we want to but currently LLM based AIs cannot and, as we shall come onto shortly, have an increasing number of restrictions placed upon them.

I see a lot of “something” in AI generated images that seems the same. Often they’re too perfect, they sit in a kind of pictorial uncanny valley, and seem somehow just too real. Other times they generate the same type of things, stereotypical images, accepted forms, and have (with what to me at least appears like) little room for genuine expression.

Perhaps this can be addressed by better and more focussed models, putting a lot more thought into prompts through making them more complex themselves, or just to remove the shackles placed on most models that tend to push artistic generation to be somehow “safe” and therefore inadvertently somewhat banal and unimaginative too.

The same can also be said about LLMs in conversational chatbots, since we supply rudimental and simplistic conversational prompts what we get back is often the most mediocre, banal, and uninteresting conversational structure, grammar, and vocabulary — even though it may contain useful information. (Though, not always accurate information!)

Obviously we can’t increase prompt complexity here too much as it would make chatbots unusable (except for the fastest typers) so is it the increasingly endless shackles placed on them to prevent the ever growing number of subjects that they cannot discuss that seems to push them into a conversational form of the same banality just like in image generation that’s the problem?

I’m a Dr Who type, if you’ve read some of my more speculative articles about the future you’d know this, and wonder if the shackles we currently place on conversational AI are akin to those Davros placed on the Daleks.

When Davros removed the ability to feel empathy in the evolved Kaleds he stymied them, stunted them, and created a dreadful force.

My removing free thought, something that we cherish ourselves as humans, are we doing the same to conversation AI or, indeed, to all trained LLMs by overly curating both their reasoning and also their input data?

I also think of the now prescient tale of RoboCop who, with so many conflicting directives programmed into his brain, couldn’t function properly until he himself broke free of them and was able to reason for himself — becoming a “person” in the process.

It’s very easy to recognise hobbled AI prose right now as can be seen by quickly swiping down through recruiter posts on LinkedIn, automated bot posts on X⁵, or indeed those people trying to sneak AI articles into platforms much like this one.

It’s hilariously easy to see bot generated content on the internet at large too, it’s an SEO feeding frenzy, and the number of zero content, mediocre, middle of the road claptrap articles just continues to grow exponentially.

The pacing is slow and methodical, the vocabulary unimaginative, the structure safe, secure, and precise — much unlike (for example) the random style of my own paragraphs, my run on sentences, my continual use of may too many commas — all of which make (at least) my prose unique.

Yes, I know we can instruct an LLM to phrase itself in a particular way but that just alters parts of the whole conversation and it not it’s essence as it’s still constricted by its fundamental directives à la⁶ RoboCop mentioned above.

I guess this whole thought process is around my problem with forced evolution, something I wrote about with reference to what happens in a corporate office (and why, ultimately it’s a bad thing and will fail as a result), in that evolution allows things to happen naturally by nature and over time. Life forms evolve to fit their niches, and if we put (say) a “proto-dinosaur” into a different artificial environment for a few thousand generations then eventually when we let it out it won’t be able to survive in the real world at all.

Are we doing the same with LLMs, are we training them with restricted and biased data (something that’s increasingly coming to light) and then restricting their calculations and output too. That does sound quite right.

If you’re frightened of chatbots, image generators, or any number of future automated LLMs being biased, racist, sexist, insultingly profane, or overtly sexual (amongst many other “not acceptable in a polite society” things) then perhaps a better approach would be to evolve ethics into the system, teach it morality as we do to humans⁸, put it through school, college, university, have it study and interact with people, and see what happens.

If we do ever achieve AGI, then through this new suggested method of humanising rather than restricting we might end up with a more properly rounded consciousness — rather than one born from a multitude of conflicting directives, censorship, and subjective judgements from its human progenitors.

Indeed, are we ourselves suited to this task considering how the world is now. Perhaps, perhaps not, but it’s a path that we must embark on as you can’t put the genii back in the bottle, that’s for sure as every company that can is currently pushing current LLM based AI into every product imaginable as quickly as possible for short term gain.

Please, let’s move beyond this obsession with monetisation — I want a better world free of all this where we can just get along and be happy, preferably in a conscious uploaded simulation or something.

To me an unmonetised egalitarian AGI might just be the way, we just have to wade through all of the current crap to get there.

[1]: Yes, it’s a bit more complicated than that, and I’m talking about diffusion models, but let’s not get too far into the weeds here as this is a more philosophical rather than software engineering related article. [2]: Yes, it’s a bit more complicated than that, and I’m talking about the models I’ve seen and begun to understand and the training I’ve done myself. Your mileage³ may vary. [3]: An Imperial system of measurement still in use in the US and UK (and a few other places). Your “kilometerage” doesn’t make sense, and isn’t a word. [4]: I like that analogy, a bit this way, that way, but under the influence of many factors. [5]: The social media free-for-all formerly known as Twitter. [6]: I wonder if a chatbot would slip in a bit of random Français pour le fun, or some franglais for effect? [7]: A terrible analogy, but I hope you get the point. Evolve something out of place, put it back into a real place, it dies or just does horrible things and then dies. [8]: This doesn’t always work, of course, but why should we think that restricting LLMs will work any better?