As indicated by others more AI-savvy than me, only a more neural-based AI is going to breach the current limits of current largely LLM-based AI. Asking for creativity from a re-hasher is akin to the science student who memorizes facts, then is compelled to apply principles to a novel situation.
With the proliferation of articles written, at least in part, by AI, I find myself wanting to stop reading immediately when I encounter one of them. Some of them do have good content, but I worry that reading even a select few will slowly turn that kind of writing acceptable to me, which will then start to affect my own writing. I think it is important for people to not accept such articles, and if possible, let the platform, or whatever, know why you aren't going to read it. That isn't much of a step, but I think it's important for both the reader, and to let the "writer" know, this is not only not acceptable, but it also iisn't fooling anyone..
> I find myself wanting to stop reading immediately when I encounter one of them.
Same. If a person didn't take time to write it, why should I take time to read it. Because I sincerely doubt they verified every detail the machine spit out.
I find it also very annoying when I'm listening to a podcast for someone I know is a real human, but they have clearly started using AI to write their content. They are just reading it.
I know some people are taking their own content and feeding it to the machine to make it "sound better" or "more professional". But I would much rather hear their imperfect writing than the repetitive patterns that are now everywhere because of everyone using AI to write for them.
Yes. We have to find a way to call it to their attention. I have started saying things like, "I started reading your post, but then it began to sound like AI, so I stopped. Please consider not using AI, or AI-speak. We should not be making it more difficult for the public to be able to discriminate between humans and AI.”
> For example, “make a professional image editor” would likely result in a very Photoshop-like result.
It would be better to use the example "make a simple image editor" would create a replica of Microsoft paint :)
Because Photoshop is orders of magnitude more complex than what current models can build without complex prompts and human-built architecture, and also, it is not opensourced, so there is no public source code to train an LLM on to make photoshop clone. (Except the situation, when Adobe developers uses cloud LLMs, and so their code leaked to guys who used pirate libraries to train LLMs, so "they are fully trustworthy").
In your previous article where was an example on how Anthropic "created" C compiler, and with Photoshop, it would be the same.
> the things of true value remain obscured behind the avalanche of recycled ideas with shiny new skins, then the rewards no longer go to the things of true value, but instead to mediocrity highly optimized to capture attention
On one hand: doesn't reward went to something "highly optimized to capture attention" before LLMs? Facebook, Twitter, TikTok, even Youtube - are attention-capturing, time-eating boredom-destroying machines, what were long before LLMs appeared, so, things of "true value" pushed into the background decades ago.
On the other hand: LLMs have some valuable power without going out of training distribution at all: it's compilation of useful features. For example, I've cloned TrustTunnel VPN client for Android (it's opensource), and added functionality to select which apps should use VPN and which shouldn't. I'm a backend programmer, and know very little about android and flutter development, so, LLM helped me by compiling well-known features into a single app. Of course, my programming background helped me to understand the task itself, and handle review of code changes, so that it wouldn't be a total mess, but anyway, I've got useful tool, saved some time, and now I can share it with my friends (as I don't want to bother with all the google play complexities). So, the point: LLM can be useful in adding features, what developers refuse to add, recreate software that has become bloatware, and only few features are useful now, and so on.
So, when all the bubble collapse, 99% of AI corporation go bankrupt, and RAM will go as cheap as it used to be, we will still have some value in our hands with opensource LLMs. Like useful tools to make other simple useful tools for everyday use.
> It would be better to use the example "make a simple image editor" would create a replica of Microsoft paint :)
True, not possible today, but the context here was hypothesizing the limits of future capability. What types of things will AI be able to do assuming we can keep scaling them.
> "highly optimized to capture attention" before LLMs? Facebook, Twitter, TikTok, even Youtube - are attention-capturing
Yes absolutely. I often will state AI did not introduce new problems, but substantially accelerated existing problems. It has allowed us to scale some of our problems in ways we could not before.
> LLMs have some valuable power without going out of training distribution at all
I agree. There absolutely is useful utility in these machines. My biggest issues are that it is significantly overhyped, training is too expensive to be sustainable, and unfortunately, a significant amount of its use are for nefarious purposes (bots, scams, flooding the world with auto-generated spam, IP theft, etc.)
> My biggest issues are that it is significantly overhyped, training is too expensive to be sustainable, and unfortunately, a significant amount of its use are for nefarious purposes
So, the normal use of any new technology by financial capitalism: large amounts of capital that cannot earn more than a 5% year profit, should be invested somewhere, gold rush, or, to be honest, tulip mania, usage of any technology as warfare first (the nuclear bomb, not the nuclear power plant was invented first). The other side of all this - the bubble pops, counter-weapons (like antiviruses for malware, but with LLM now) would be created. It will balance out some bad sides, of course with lots of casualties or "collateral damage". Nothing can be changed here, while we are still in this economic formation
Yes, there are no guarantees a technology will succeed no matter how much investment is poured into it. The economic reckoning, that is likely coming, will sort it all out. Balance will inevitably be restored, as we cannot escape reality. What works will survive, and what doesn't will go away. And yes, might be a very painful journey nonetheless.
I am starting to see more and more posts by individuals on posts like this that are just too perfect, follow very deliberate patterns of framing with too many subheads and too many bullet points for a person who is dedicated to churning out as many articles as possible that bring them revenue. Usually the best writers can craft things succinctly and dont waste the readers time and energy listing point after point to somehow score more darts hitting the center. At some point it becomes overkill and seems obvious that this is not human thought or research.
As indicated by others more AI-savvy than me, only a more neural-based AI is going to breach the current limits of current largely LLM-based AI. Asking for creativity from a re-hasher is akin to the science student who memorizes facts, then is compelled to apply principles to a novel situation.
Yes, machines with true intelligence and creativity still do not exist, and we still do not know how to build them.
With the proliferation of articles written, at least in part, by AI, I find myself wanting to stop reading immediately when I encounter one of them. Some of them do have good content, but I worry that reading even a select few will slowly turn that kind of writing acceptable to me, which will then start to affect my own writing. I think it is important for people to not accept such articles, and if possible, let the platform, or whatever, know why you aren't going to read it. That isn't much of a step, but I think it's important for both the reader, and to let the "writer" know, this is not only not acceptable, but it also iisn't fooling anyone..
> I find myself wanting to stop reading immediately when I encounter one of them.
Same. If a person didn't take time to write it, why should I take time to read it. Because I sincerely doubt they verified every detail the machine spit out.
I find it also very annoying when I'm listening to a podcast for someone I know is a real human, but they have clearly started using AI to write their content. They are just reading it.
I know some people are taking their own content and feeding it to the machine to make it "sound better" or "more professional". But I would much rather hear their imperfect writing than the repetitive patterns that are now everywhere because of everyone using AI to write for them.
Yes. We have to find a way to call it to their attention. I have started saying things like, "I started reading your post, but then it began to sound like AI, so I stopped. Please consider not using AI, or AI-speak. We should not be making it more difficult for the public to be able to discriminate between humans and AI.”
Thanks for an interesting analysis.
Good article! I want to comment on two points:
> For example, “make a professional image editor” would likely result in a very Photoshop-like result.
It would be better to use the example "make a simple image editor" would create a replica of Microsoft paint :)
Because Photoshop is orders of magnitude more complex than what current models can build without complex prompts and human-built architecture, and also, it is not opensourced, so there is no public source code to train an LLM on to make photoshop clone. (Except the situation, when Adobe developers uses cloud LLMs, and so their code leaked to guys who used pirate libraries to train LLMs, so "they are fully trustworthy").
In your previous article where was an example on how Anthropic "created" C compiler, and with Photoshop, it would be the same.
> the things of true value remain obscured behind the avalanche of recycled ideas with shiny new skins, then the rewards no longer go to the things of true value, but instead to mediocrity highly optimized to capture attention
On one hand: doesn't reward went to something "highly optimized to capture attention" before LLMs? Facebook, Twitter, TikTok, even Youtube - are attention-capturing, time-eating boredom-destroying machines, what were long before LLMs appeared, so, things of "true value" pushed into the background decades ago.
On the other hand: LLMs have some valuable power without going out of training distribution at all: it's compilation of useful features. For example, I've cloned TrustTunnel VPN client for Android (it's opensource), and added functionality to select which apps should use VPN and which shouldn't. I'm a backend programmer, and know very little about android and flutter development, so, LLM helped me by compiling well-known features into a single app. Of course, my programming background helped me to understand the task itself, and handle review of code changes, so that it wouldn't be a total mess, but anyway, I've got useful tool, saved some time, and now I can share it with my friends (as I don't want to bother with all the google play complexities). So, the point: LLM can be useful in adding features, what developers refuse to add, recreate software that has become bloatware, and only few features are useful now, and so on.
So, when all the bubble collapse, 99% of AI corporation go bankrupt, and RAM will go as cheap as it used to be, we will still have some value in our hands with opensource LLMs. Like useful tools to make other simple useful tools for everyday use.
Thank you!
> It would be better to use the example "make a simple image editor" would create a replica of Microsoft paint :)
True, not possible today, but the context here was hypothesizing the limits of future capability. What types of things will AI be able to do assuming we can keep scaling them.
> "highly optimized to capture attention" before LLMs? Facebook, Twitter, TikTok, even Youtube - are attention-capturing
Yes absolutely. I often will state AI did not introduce new problems, but substantially accelerated existing problems. It has allowed us to scale some of our problems in ways we could not before.
> LLMs have some valuable power without going out of training distribution at all
I agree. There absolutely is useful utility in these machines. My biggest issues are that it is significantly overhyped, training is too expensive to be sustainable, and unfortunately, a significant amount of its use are for nefarious purposes (bots, scams, flooding the world with auto-generated spam, IP theft, etc.)
> My biggest issues are that it is significantly overhyped, training is too expensive to be sustainable, and unfortunately, a significant amount of its use are for nefarious purposes
So, the normal use of any new technology by financial capitalism: large amounts of capital that cannot earn more than a 5% year profit, should be invested somewhere, gold rush, or, to be honest, tulip mania, usage of any technology as warfare first (the nuclear bomb, not the nuclear power plant was invented first). The other side of all this - the bubble pops, counter-weapons (like antiviruses for malware, but with LLM now) would be created. It will balance out some bad sides, of course with lots of casualties or "collateral damage". Nothing can be changed here, while we are still in this economic formation
Sorry for the nerdy tangent :)
Yes, there are no guarantees a technology will succeed no matter how much investment is poured into it. The economic reckoning, that is likely coming, will sort it all out. Balance will inevitably be restored, as we cannot escape reality. What works will survive, and what doesn't will go away. And yes, might be a very painful journey nonetheless.
I am starting to see more and more posts by individuals on posts like this that are just too perfect, follow very deliberate patterns of framing with too many subheads and too many bullet points for a person who is dedicated to churning out as many articles as possible that bring them revenue. Usually the best writers can craft things succinctly and dont waste the readers time and energy listing point after point to somehow score more darts hitting the center. At some point it becomes overkill and seems obvious that this is not human thought or research.
Yes, even much of the AI criticism is now written by AI. It is exhausting.
https://www.mindprison.cc/p/anti-ai-sentiment-is-being-ai-generated
This is a prison.
Always good to see a new post from you.
Thank you!