Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

While I agree with most of your comment, I'd like to dispute the story about GPT-2.

Yes, they did claim that they wouldn't release GPT-2 due to unforeseen risks, but...

a. they did end up releasing it,

b. they explicitly stated that they wouldn't release GPT-3[1] for marketing/financial reasons, and

c. it being dangerous didn't stop them from offering the service for a profit.

I think the quotes around "open" are well deserved.

[1] Edit: it was GPT-4, not GPT-3.



> they did end up releasing it,

After studying it extensively with real-world feedback. From everything I've seen, the statement wasn't "will never release", it was vaguer than that.

> they explicitly stated that they wouldn't release GPT-3 for marketing/financial reasons

Not seen this, can you give a link?

> it being dangerous didn't stop them from offering the service for a profit.

Please do be cynical about how honest they were being — I mean, look at the whole of Big Tech right now — but the story they gave was self-consistent:

[Paraphrased!] (a) "We do research" (they do), "This research costs a lot of money" (it does), and (b) "As software devs, we all know what 'agile' is and how that keeps product aligned with stakeholder interest." (they do) "And the world is our stakeholder, so we need to release updates for the world to give us feedback." (???)

That last bit may be wishful thinking, I don't want to give the false impression that I think they can do no wrong (I've been let down by such optimism a few other times), but it is my impression of what they were claiming.


> Not seen this, can you give a link?

I was confusing GPT3 with GPT4. Here's the quote from the paper (emphasis mine) [1]:

> Given both THE COMPETITIVE LANDSCAPE and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.

[1] https://cdn.openai.com/papers/gpt-4.pdf


Thanks, 4 is much less surprising than 3.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: