A computer said "Look what I make! These pictures and songs - they're not fake!" But humans they smiled, For art's something wild That only life's journey can wake
I gave a talk in Dec 2024 to the European School of Economics, invited by Paul Blanchard’s sustainability class in which I try to explore how the Artist is changing and how they will evolve given all the industrial creative production that is possible through generative AI technologies.
AI pervades our discussions and is this moment, and has made heavy impact on the creative sector in which I work. While many have seen a reduction in work because of GenAI, we also are familiar with how human audiences adapt quickly to changes like this and look for value in processes that move beyond.
In my talk I propose that there is a class of creative output that will be subject to detailed scrutiny of qualities that essentially flip the problem on its head. When you look for something beautiful to experience or create how much will you care about how it was made, by whom and for what reason? I expect that as we seek core aesthetic values such as scarcity, authenticity, craft we will become more and more tuned in to those affects that stand out as being unique human context that can be demonstrated to be not industrialised production.
This new creative language is an opportunity, particularly to talk to newer generations that are particularly attune to identifying technological quirks.
I now can find a good excuse
for things left here undone,
and then without the time to do
much more than doing, something fun...
Snippety snip and clippity clip,
to retransform my reel of takes.
Flippity flip and hey, that is it!
Now a majestic field of fakes.
I think the sense that nothings true
will hurt our trust as hidden cost.
But surely something we could do,
to try to keep what now is lost?
Large Language Models present an interesting unexpected twist in the development of synthetic intelligence. Authoritative sounding responses are extracted from a very large body of recorded information, with no single connection to specific sources, rather a probabilistic match that relies on the chance of finding words in certain order. So the outcome is actually “by definition” not authoritative because it cannot be uniquely reconnected to its sources.
Despite this limitation, these models offer both tremendous utility and significant risks. The vast array of potential use cases for these models can be incredibly versatile. However, their lack of context and source verification can make it difficult to assess the accuracy and reliability of their responses.
The tools feel a bit like a trick, and can lead to frustrating fail cases while searching for the right answer to something. Example tricks to draw out the most obvious “lies” is to ask the tool to try to help do something that is not well covered in the original huge corpus of information the model ingested. Some examples would be to ask questions like:
How can you use Word to convert a JPG to a TIFF?
Can you write me a MSDOS script to add up the size of all data inside a folder?
Can you pick me a random name?
Can you bulk export all your OSX notes?
These are unusual but not so unusual questions but they relate to existing corpus data that is either not covered or unreliably covered. In the first question you might get responses that appear to work with the software but that are not existing features. In the second you might get scripts that do not function, partly because the task is non trivial for all cases. In the third the system will pick a name for you but won’t be able to reassure on how this name could be random. The last one is a known IOS PITA feature (that makes me regret I ever started using Notes in the first place) but the LLM will find something to suggest, even if it does not exist – really.
In this sense the tool is useful when the result can be verified easily by the user through direct inspection, but not as good when the verification requires additional work to figure out. In some of the example answers I got for these I actually downloaded software and tried to follow the suggestion only to discover there were non existent menu options that needed to be accessed… It was a hallucination!
However… LLMs are a very important twist!
For the first time in recent years we have a system that supports open user access with simple tools and that generates responses that feel “human-like” partly also thanks to its flexibility to emphasize relying on the huge body of examples it has access to, instead of trying to be precise.
Chatbot type interaction never felt like anything more than an elaborate multiple choice engine, even in the best examples. The sense was of something predictable that had brittle rules that would collapse into repetitive states.
Eventually it would remind me of old Infocom text adventures like Zork where I’d spent hours to figure out how to find the exact sentence to get a result.
What matters here is how a relatively simple change in the mechanism with which information is processed leads to such a strong added illusion of intelligence. And that this trick is useful to help solve problems that require integrating language in vague ways across different areas.
We need to take heed that approaching the mechanics of higher human thinking could be subject to simple changes in our computational approach. And while many types of human cognitive skills will remain difficult to recreate in these ways, many others might fall by the wayside as we start to radically transform human culture.
Suddenly our mainstream culture is rekindled into existential threats of AI, employment effects, ethical questions around synthetic intelligence… many themes pour out prompted by this new toolset because it rings a bell that hasn’t been heard in this way before.
There might be nothing so special about cognitive technologies in the sense that our own thinking is one possible way to go about it but many others are possible.
I would not trivialize either the potential nor the ultimate risks we take in trying to understand intelligence. But I also see nothing more natural than for enquiry to go towards understanding what it means to understand. And all the while working to protect what it means to exist in our peculiar human way.
Based on where we are today, this is my todo list:
Engage with the technologies and evaluate them
Take active role in warning about relying on them: lying AIs for example but also try to educate people on how these models work and what to expect from them.
Advocate that only humans have skin in the game, the tools are accessories (until someday the tools may argue to be nothing of the sort…)
Remind myself that these tools will likely develop exponentially
Push for rights for sentient beings in general and their environment
Probably the trickiest question is will these technologies help solve technologically borne problems like anthropogenic climate change?
Or make these problems harder to solve?
[Special bonus addition suggested from CGPT]
Is there a way to turn the tide?
To find a path that's true?
To hold onto what we've lost
And cherish what we knew?
It's up to us to take a stand
To seek the truth each day
And in the face of all life's trials
To find a better way.
