4 ms·

I like the work OpenAI has done in the past, so please take my complaints as gentle complaints: - a very small percentage of generated samples ‘read well’ - I

I like the work OpenAI has done in the past, so please take my complaints as gentle complaints: - a very small percentage of generated samples ‘read well’ - I Don’t see any evidence of ‘understanding’, at least in the sense that BERT can identify original nouns that pronouns refer to (anaphora resolution): amazing to get state of the art results - the ‘we aren’t going to share code and results’ thing reminds me of the movie The Wizard of Oz: trust them that they developed cool technology - in 6 months, high school students might be getting similar results, given how fast our field is moving forward: this makes the danger concern less important in my opinion This whole thing seems like a throwback to how science was conducted 10 years ago - ancient times in our fast moving tech world.

msla 8y ago
> the ‘we aren’t going to share code and results’ thing reminds me of the movie The Wizard of Oz: trust them that they developed cool technology And this is where it stands: They claim to have something. They won't release it. If they had nothing but a somewhat creative streak to make up the results, the world would look exactly the same as it would if they had something. Them lying is the least hypothesis here.
- unimpressive 8y ago
  Given that experts in the field seem to think this is an incremental step forward, OpenAI has already demonstrated a track record of doing things probably more impressive than this, and the massive loss of credibility that would come with being caught in a lie here, on priors we should expect they're not lying.