I listened to the whole thing and it was worth it! Quick reflection: LLMs are good at producing art-based content but bad at making good art. There is no model that can account for subjectivity of artistic quality which is related to the subjectivity of artistic expression. It reminds me about how I was always frustrated with the idea of music services like pandora that made recommendations based on different sounds and genres when there can be two songs with the same instrumentation style and genres and one is incredible and one sucks.
This interview illustrates, in my view, a common misunderstanding about LLMs. Mills Baker insists that these models never produce anything truly original and only recycle what already exists. I think that’s reductive for two main reasons.
First, LLMs contain an immensely rich latent space that far exceeds our individual cognitive capacities. Within that space, associations and ideas can emerge that no human alone could have conceived. This is where we can draw on Jacques Rancière’s concept of the “thinking image”: an image that does not think by itself, but that puts the spectator in a position to think differently. In the same way, an LLM is not “intelligent” in a human sense, but the interaction with it can provoke new thoughts. Originality doesn’t come from the AI alone, but from what it sparks in us.
Second, the real problem is not that LLMs “imitate” or “automate” what we already do, but that we reduce them to that function. We imagine them as duplication machines, whereas we should see them as instruments of discovery. Their computational power allows us to explore and decrypt worlds that were previously inaccessible. The example of Google’s “DolphinGemma” model analyzing whale songs is a perfect illustration: applied to other domains, LLMs could uncover structures, patterns, and truths we never knew existed.
In other words, LLMs are not just tools of reproduction: they are openers of worlds. The real challenge is to learn how to use them not to redo what we already know, but to access what we cannot yet see.
I totally loved this. Especially because I'm heavily using Claude for similar things - finding patterns in my own psycho-spiritual work and business development. I work with the Gene Keys (which is even deeper than the IGing or Tarot) and developed my own methodology out of it. The thing is, that while it's easy to interpret patterns for others, sometimes you are blind on yourself. Here Claude is an exceptional help, because I basically cloned myself (and my niche knowledge) and it can help me, like I'm helping my clients.
The downside is, that the necessary context for each request is enormous. I start chats with over 100k Token, because this is the context needed for my very specific and surgical requests.
So for me specifically the personal and business use overlap often. Other than that and as a solopreneur, I am now able to produce things in months, that would have been taken years without LLM. To be honest, most likely I would have quit while on the path because the sheer enormous amount of time and money that building my business from zero would have cost. This year I've published a book in 2 languages, audiobook, produced several paid courses and lots of free material. All of this only possible because of Claude.
All of that while making huge personal development steps, that flowed back into my business and client work.
I agree, that revolutionary technology doesn't need evangelism - and that's exactly the case for me here. Nobody needed to convince me in late 2022 to start using the first Claude Beta. I immediately got it and am using it since then.
Also Perplexity has replaced Google for me completely and their new Comet browser is really helpful daily with very practical things.
Like I asked it yesterday how to correctly publish my first podcast here on Substack - but instead of looking in any FAQ or help articles (that may be outdated), it clicked some options here by itself and told me what settings I should consider. Or I just tell it to re-order my last Vitamin D on Amazon. Flawless.
So I use Claude and AI for nearly everything today and still believe that it's not for everyone or would evengalise others to do it too.
Simply because I know, that the majority of people won't be able to think in context or simply aren't intelligent enough to work with AI in such an efficient and individual way.
What a good conversation—so many resonating moments. Especially from 1:02:25 until the end.
Mills: We don’t ask enough how our minds work.
Alex: I think it doesn’t want to be known… the minute you look at—it’s gone.
Mills: I think that’s right… Nobody likes that answer.
I listened to the whole thing and it was worth it! Quick reflection: LLMs are good at producing art-based content but bad at making good art. There is no model that can account for subjectivity of artistic quality which is related to the subjectivity of artistic expression. It reminds me about how I was always frustrated with the idea of music services like pandora that made recommendations based on different sounds and genres when there can be two songs with the same instrumentation style and genres and one is incredible and one sucks.
This interview illustrates, in my view, a common misunderstanding about LLMs. Mills Baker insists that these models never produce anything truly original and only recycle what already exists. I think that’s reductive for two main reasons.
First, LLMs contain an immensely rich latent space that far exceeds our individual cognitive capacities. Within that space, associations and ideas can emerge that no human alone could have conceived. This is where we can draw on Jacques Rancière’s concept of the “thinking image”: an image that does not think by itself, but that puts the spectator in a position to think differently. In the same way, an LLM is not “intelligent” in a human sense, but the interaction with it can provoke new thoughts. Originality doesn’t come from the AI alone, but from what it sparks in us.
Second, the real problem is not that LLMs “imitate” or “automate” what we already do, but that we reduce them to that function. We imagine them as duplication machines, whereas we should see them as instruments of discovery. Their computational power allows us to explore and decrypt worlds that were previously inaccessible. The example of Google’s “DolphinGemma” model analyzing whale songs is a perfect illustration: applied to other domains, LLMs could uncover structures, patterns, and truths we never knew existed.
In other words, LLMs are not just tools of reproduction: they are openers of worlds. The real challenge is to learn how to use them not to redo what we already know, but to access what we cannot yet see.
I totally loved this. Especially because I'm heavily using Claude for similar things - finding patterns in my own psycho-spiritual work and business development. I work with the Gene Keys (which is even deeper than the IGing or Tarot) and developed my own methodology out of it. The thing is, that while it's easy to interpret patterns for others, sometimes you are blind on yourself. Here Claude is an exceptional help, because I basically cloned myself (and my niche knowledge) and it can help me, like I'm helping my clients.
The downside is, that the necessary context for each request is enormous. I start chats with over 100k Token, because this is the context needed for my very specific and surgical requests.
So for me specifically the personal and business use overlap often. Other than that and as a solopreneur, I am now able to produce things in months, that would have been taken years without LLM. To be honest, most likely I would have quit while on the path because the sheer enormous amount of time and money that building my business from zero would have cost. This year I've published a book in 2 languages, audiobook, produced several paid courses and lots of free material. All of this only possible because of Claude.
All of that while making huge personal development steps, that flowed back into my business and client work.
I agree, that revolutionary technology doesn't need evangelism - and that's exactly the case for me here. Nobody needed to convince me in late 2022 to start using the first Claude Beta. I immediately got it and am using it since then.
Also Perplexity has replaced Google for me completely and their new Comet browser is really helpful daily with very practical things.
Like I asked it yesterday how to correctly publish my first podcast here on Substack - but instead of looking in any FAQ or help articles (that may be outdated), it clicked some options here by itself and told me what settings I should consider. Or I just tell it to re-order my last Vitamin D on Amazon. Flawless.
So I use Claude and AI for nearly everything today and still believe that it's not for everyone or would evengalise others to do it too.
Simply because I know, that the majority of people won't be able to think in context or simply aren't intelligent enough to work with AI in such an efficient and individual way.
Thank you for this insightful talk! 🙏
Very empowering for writers! Great stuff!