Note: This blog post is cross-posted from my personal website, where I expect a broader audience than here. If you are familiar with the difficulty and significance of neural network interpretability, skip to the third subsection titled "In defence of fighting fire with fire" Summary: This is a post about...
Why it’s important The current phase of acceleration in AI has increased the importance of the debate around consciousness to a degree that I never thought would arrive this early. Systems like GPT-4 are passing most of our, thus far thought of, tests for consciousness[1]. Still, only very few even...
This is my first post on here so please be lenient if I fail to follow any norms. An image made by Midjourney AI. The prompt was "the physical manifestation of logic" With the explosion of AI-generated images and text from Dall-E 2, Midjourney AI and GPT3, it does not...
Note: This blog post is cross-posted from my personal website, where I expect a broader audience than here. If you are familiar with the difficulty and significance of neural network interpretability, skip to the third subsection titled "In defence of fighting fire with fire"
Summary: This is a post about a research project I conducted over the last few months. I trained a model to translate transformer weights, compiled using tracr, back into the RASP code they were compiled from. I give a general introduction to the reasoning behind why I think this is valuable, describe some technical components of the method, talk about the results and finally talk about some of the... (read 4183 more words →)
Why it’s important
The current phase of acceleration in AI has increased the importance of the debate around consciousness to a degree that I never thought would arrive this early. Systems like GPT-4 are passing most of our, thus far thought of, tests for consciousness[1]. Still, only very few even consider the consciousness of an LLM to be a possibility. Even if they may have initially agreed with the metrics used to measure consciousness. Why is that?
Intuitively, it makes sense. When I consider a simple logic gate on a computer chip inside a server, responding to changes in electrical currents, I don't associate it with consciousness any more than I do a pebble... (read 1448 more words →)
I haven't thought of training the models by evaluating the selection of the image by the user. And thanks for correcting me on my Dalle 2 training data - claim.
What do you mean by "training a model to detect its own errors"? Maybe this is a naive question (i am an ML newcomer) but isn't that impossible by definition? Why would a model make a mistake if it's capable of identifying it as such? Do you mean that through continuous improvement the model could correct the mistakes it made in the past, after some time has passed?
The problem of dilution remains for GPTs in my view. Widespread use seems likely over the... (read more)
I tend to agree with you. But I am not sure that our way of distinguishing AI-generated from human-generated content will reach the perfection required for this to "work". Assuming that the mechanism of distinguishing the two will remain imperfect at least a bit of a feedback loop will remain, which will slow down development.
This is my first post on here so please be lenient if I fail to follow any norms.
With the explosion of AI-generated images and text from Dall-E 2, Midjourney AI and GPT3, it does not seem unreasonable to assume that a non-negligible part of the content of the internet might become AI-generated. This in itself is not problematic. I actually look forward to the internet where everyone has access to the means to create awesome media for almost nothing.
But, since most large language models and multimodal AIs are trained on a dataset that basically consists of "all of the... (read more)