ChatGPT, a AI-Chat Bot based on a LLM (Large Language Model) is all the rage now. It can answer many questions about the world in almost any language and it has knowledge of many obscure topics. It can write programs in programming languages or generate a mathematical formula in LaTeX or translate it into a COBOL program.
While on the on hand side its capabilities are astonishing it also makes some really stupid mistakes. E.g. it answerd that subtraction is commutative:
chatgpt claims that subtraction is commutative
And it was even rather sure about its believe. So it seems its capabilities for reasoning are still somewhat limited and it is, after all primarily a LLM which, at its core, predicts the liklyhood of some text following other texts. One posting that points this out is:
Mastodon Post from Felix Stalder
and Talking About Large Language
Models (Murray Shanahan, December 2022)
In general People are really impressed by the abilities of ChatGPT at first but with a little bit of poking around one can also quickly find out where its limits are. At this point many people seem to be rather quick in dismissing the whole thing all together.
One thing is that they do not let ChatGPT directly access the internet, that the training data they used is mostly from 2021 and that it seems they have not used Wikipedia for training. If you as ChatGPT it avoids to answer questions about the licenses of the training data. It seems obvious that the company behind ChatGPT (OpenAI) want to fend of demands that its creation needs to belong to the public. OpenAI should have been a non-profit organization but is now more or less under the control of Microsoft. Fortunately, there are other initiatives that try to create truly free and open alternatives like GPT-J.
When asked about the Steinlaus ChatGPT was not aware that this insect does not exist and was only a joke that made it into a clinical dictionary. This at least confirms that the Creative Commons Licensed Wikipedia was not used for the training of ChatGPT.
Why ChatGPT should not easily be dismissed
The shortcomings that the Article from Murray Shanahan mentions are real but the point is that it is easy to see how they can be overcome.
It should be fairly easy to create some forms of feedback loop that feed the output of a chat bot similar to ChatGPT back to itself. This way you would immediately have something that is very close to having some form of consciousness. The system would be able to "think" even in the absence of an external prompts. It could ponder some questions and after "thinking" it for a while could create new knowledge that it uses as training data.
Currently ChatGPT does not have the ability to scrap the internet and certainly it is easy to see why this data has been kept hidden from it but technically it would be easy to include much more sources and the ability to scan the Intenet for new information.
Still the main limitation mentioned in Shanahans article would be that it does not have a real world. Yet again it would be easy to give it the ability to experiment with the real world or simulated worlds. E.g. its enormous capabilities for generating code could be used to run programs in sandboxes and then to study the output of the code it has generated. Actually I would assume that a lot of people are currently in the process of implementing just that.
The possibilities of interconnecting systems like ChatGPT with other already existing AI systems are also mind blowing.
So I think that some are to quick to dismiss ChatGPT by looking at its current limitations. Given how easy it will be to overcome those, it seems that we are on the brink of a technological revolution.
Ray Kurzweil still thinks it will take till 2029 to have an AI that really passes a Turing test, but given the current state of affairs I think that all this could happen much, much faster as well.
In any case given the current abilities there are a lot of jobs that an AI like ChatGPT can make obsolete. Even more seems possible with a few extra steps that will only be a matter of weeks or month.
So it is more then time for a Universal Basic Income.
More frightening are of course all the possible harm that can be done with technologies like this in the wrong hands. And given the current state of the world, it is more then likely that all this is already in the wrong hands.
Franz Schäfer (Mond) Dec, 2022
