Thread by @emilymbender

Okay, taking a few moments to reat (some of) the #gpt4 paper. It's laughable the extent to which the authors are writing from deep down inside their xrisk/longtermist/"AI safety" rabbit hole.

>>

Things they aren't telling us:
1) What data it's trained on
2) What the carbon footprint was
3) Architecture
4) Training method

>>

But they do make sure to spend a page and half talking about how they vewwy carefuwwy tested to make sure that it doesn't have "emergent properties" that would let is "create and act on long-term plans" (sec 2.9).

>>

I also lol'ed at "GPT-4 was evaluated on a variety of exams originally designed for humans": They seem to think this is a point of pride, but it's actually a scientific failure. No one has established the construct validity of these "exams" vis a vis language models.

>>

For more on missing construct validity and how it undermines claims of 'general' 'AI' capabilities, see:

…ets-benchmarks-proceedings.neurips.cc/paper/2021/has…

>>

Also LOL-worthy, against the backdrop of utter lack of transparency was "We believe that accurately predicting future capabilities is important for safety.

>>

... Going forward we
plan to refine these methods and register performance predictions across various capabilities before
large model training begins, and we hope this becomes a common goal in the field."

Trying to position themselves as champions of the science here & failing.

A cynical take is that they realize that without info about data, model architecture & training set up, we aren't positioned to reason about how the model produces the results that it does ... and thus more likely to believe claims of `AGI' and thus buy what they're selling.

>>

But given all the xrisk rhetorical (and @sama 's blogpost from Feb) it may also be possible that at least some of the authors on this thing actually believe their own hype and really think they are making choices about "safety".