RANT:
For a long time, the public understanding of LLM’s has been hampered by the following phenomenon. Someone resolves to test what LLM’s can do. Because they’re skeptical of LLMs, both technologically and ethically, they don’t have access to ChatGPT-4, which costs money. They ask ChatGPT-3.5 a question and it fails to answer that question miserably. They then make a post on Twitter: “We’re supposed to believe this garbage is going to change the world?”. Someone in the comments points out that ChatGPT-4 gets the right answer. No one notices. The cycle continues.
Of course, this is just a minority of people on Twitter. The real danger is all the millions of people who play around with GPT-3.5 on their own and mistake it for state-of-the-art. We should be concerned about LLM’s, and proper, responsible, concern requires understanding.
Good (?) news
The (questionably) good news is this might be a thing of the past. ChatGPT-4o, the most sophisticated model OpenAI (or anyone) offers to the public is currently available for free. Of course, this will have many downsides- easier cheating in school, more deceptive spam, etc., but at least it will make the state-of-the-art clear. One additional advantage, I suppose, is that at least the poor kids and the rich kids who can afford a GPT subscription will be able to cheat equally well.
Edit: As H. points out:
Btw, had to log in (I already have an account) to get anything other than 3.5.
The bar for AGI is getting pretty high; being able to do anything any human can do. Nobody's that capable. First example that comes to mind is languages: a serious polyglot might be fluent in five or ten, but there ~7000 not counting conlangs.
Btw, had to log in (I already have an account) to get anything other than 3.5.