|
|
|
|
|
|
by Teever
1211 days ago
|
|
|
Everything ends, it's true. But there's no reason to believe that this technique will plateau before some noteworthy results are obtained. Other than being connected by the term 'AI' what makes you believe that the past performance of researchers is indicative of future performance? Are any of the limitations of the Apollo program inherent limitations of space exploration? How are they connected? |
|
|
Absolutely there is, and the article specifically touches on it.
Because these models do not encode semantics, there's very little reason to believe the hallucination problem can be solved with current techniques.
In fact, OpenAI themselves have been downplaying expectations about GPT-4, and my suspicion is that's because these techniques are already starting to plateau.
I fully expect that GPT-4 will create even more natural-seeming language, with greater sentence variation, less repetition, etc, and with no change (or possibly even an increase) in the frequency of factual errors in the results.