Joscha Bach
speaker
510 appearances
2 recordings
2 series
first heard Aug 2023
last heard 23 Jan
Joscha Bach’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 1 in all, peaking in Jan 2026 with 1.
Appearances
Most of the stuff that I'm doing in my life doesn't need chat GPT.
There are a few tasks that are where it helps, but the main stuff that I need to do, like developing my own thoughts and aesthetics and relationship to people, it's necessary for me to write for myself because writing is not so much about producing an artifact that other people can use, but it's a way to structure your own thoughts and develop yourself.
And so I think this idea that kids are writing their own essays with chat GPT in the future is going to have this drawback that they miss out on the ability to structure their own minds via writing. And I hope that the schools that our kids are in will retain the wisdom of understanding what parts should be automated and which ones shouldn't.
I recently wrote a tool that is using the camera on my MacBook and Swift to read pixels out of it and manipulate them and so on, and I don't know Swift. So it was super helpful to have this thing that is writing stuff for me. Also interesting that mostly it didn't work at first.
I felt like I was talking to a human being who was trying to hack this on my computer without understanding my configuration very much and also making a lot of mistakes. And sometimes it's a little bit incoherent. So you have to ultimately understand what it's doing. There's still no other way around it. But I do feel it's much more powerful and faster than using Stack Overflow.
GPT-N, probably. It's not even clear for the present systems. When I talk to my friends at OpenAI, they feel that this question whether the models currently are conscious is much more complicated than many people might think. I guess that it's not that OpenAI has a homogeneous opinion about this. There are some aspects to this.
One is, of course, this language model has written a lot of text in which people were conscious or described their own consciousness, and it's emulating this. And if it's conscious, it's probably not conscious in a way that is close to the way in which human beings are conscious.
But while it is going through these states and going through a 100-step function that is emulating adjacent brain states that require a degree of self-reflection, it can also create a model of an observer that is reflecting itself in real time and describe what that's like. And while this model is a deepfake, our own consciousness is also as if. It's virtual, right? It's not physical.
Our consciousness is a representation of a self-reflexive observer that only exists in patterns of interaction between cells. So it is not a physical object in the sense that exists in base reality, but it's really a representational object that develops its causal power only from a certain modeling perspective.
Yes. And so to which degree is the virtuality of the consciousness in chat GPT more virtual and less causal than the virtuality of our own consciousness? But you could say it doesn't count. It doesn't count much more than the consciousness of a character in a novel, right?
It's important for the reader to have the outcome, the artifact of a model is describing in the text generated by the author of the book what it's like to be conscious in a particular situation and performs the necessary inferences. But the task of creating coherence in real time in a self-organizing system by keeping yourself coherent so the system is reflexive
That is something that language models don't need to do. So there is no causal need for the system to be conscious in the same way as VR.
And for me, it would be very interesting to experiment with this, to basically build a system like a cat, probably should be careful at first, build something that's small, that's limited, has limited resources that we can control and study how systems notice a self-model, how they become self-aware in real time.
And I think it might be a good idea to not start with a language model, but to start from scratch using principles of self-organization.
My intuition is that the language models that we are building are golems. They are machines that you give a task and they're going to execute the task until some condition is met. And there's nobody home. And the way in which nobody is home leads to that system doing things that are undesirable in a particular context.
So you have that thing talking to a child and maybe it says something that could be shocking and traumatic to the child. Or you have that thing writing a speech and it introduces errors in the speech that your human being would ever do if they were responsible. But The system doesn't know who's talking to whom. There is no ground truth that the system is embedded into.
And of course, we can create an external tool that is prompting our language model always into the same semblance of ground truth. But it's not like the internal structure is causally produced by the needs of a being to survive in the universe. It is produced by imitating structure on the internet.
Maybe it's sufficient to use the transformer with the different loss function that optimizes for short-term coherence rather than next token prediction over the long run. We had many definitions of intelligence and history of AI. Next token prediction was not very high up on the list. And there are some similarities, like cognition as data compression is an old trope.
Solomonov induction, where you are trying to understand intelligence as predicting future observations from past observations, which is intrinsic to data compression. Mm-hmm. And predictive coding is a paradigm with this boundary between neuroscience and physics and computer science.
So it's not something that is completely alien, but this radical thing that you only do next token prediction and see what happens is something where most people, I think, were surprised that this works so well.
Showing 321–340 of 510 · page 17 of 26
← Previous
Next →