The difference is we can improve the AI to be more accurate, and I suspect before long it’ll generate better content than a human would that’s verifiable with citations. There may come a time where writing is done by a machine much as a calculator does our math. But knowledge maybe shouldn’t be canonically encoded in ascii blobs randomly strewn over the web - maybe instead of accumulated knowledge needs to be structured in a semantic web sort of model. We can use the machine to describe to use the implications of the knowledge and it’s context in human language. But I get a feeling in 20 years it’ll be anachronistic to write long form.
The model needs known "good" feedback to improve. The problem is that the quality of its training data declines with the more output produced. It's rather inevitable that we'll be drowning in AI generated garbage before long. A lot of people are confusing LLMs with true intelligence.
That’s why I think knowledge needs to be better structured than blobs of text scattered everywhere. An AI can be more than an LLM, Wolfram posted recently about that. You can use the LLM to convert a question into a semantic query and a semantic validator and check and amend and provide a semantic knowledge graph explaining an answer and the LLM can convert it back to meat language. I think people confuse LLM with true intelligence, but the cynics also confuse LLM with a complete and fixed point solution.
Your point also seems to assume no curation can happen on what is ingested. Simply because that might be what’s happening now you could also simply train the LLM on known good sources and be as permissive or restrictive as is necessary. Depending on how good the classifiers are for detecting LLM output (openai released on recently) or other generated / automatically derived content you can start to be more permissive.
My point is people seem to be blinded by what is vs what may be. This is not the end of the development cycle of the tech, it’s the pre-alpha release by the first meaningful market entrant. I’d be slower to judge what the future looks like rather than assuming everything stays fixed in time as it is.
Oh definitely. There's bound to be improvements especially when you glue an LLM to a semantic engine, etc.
The issue is again, fundamentally, one of data. Without authenticating what's machine generated and what's "trusted" proliferation of AI generated content is bound to reduce data quality. This is a side effect of these models being trained to fool discriminators.
Ultimately now I think there is going to be a more serious look around the ethics of using these models and putting guard rails around what exactly is permissible. I suspect the US will remain a wild west for some time but the EU will be a test-bed.
Ultimately, I'm fairly excited about the applications of all this.
