Tuesday, May 10, 2022

asymmetric power and language warfare

 so the GPT-3 API release blog post(but not models) from OpenAI does some virtue signalling about the possibility of misuse of the underlying models for disinformation. I'm not sure that washes (in the ethics sense) in that there's nothing to stop them being hired by someone do evil for money - only if they had a radical governance model could they avoid the "maximise shareholder value" mantra/fate, surely. And note they are not the only game in town, even if they have that wonderful governance model - there's Google's new palm as well as the BAAI Wu Dao - there's quite a few organisations with access to hyper-scale cloud compute these days, so really the geni is out of the bottle. maybe we need a new global governance - start with models like Asilomar or Pugwash, but then legislate? Perhaps the EU could lead the way by refining some of its rather shotgun AIA rhetoric?

One problem I have with the framing above is that I am not clear what exactly  these near trillion-parameter "models" actually are - most simpler AI (including recently, some smaller neural nets) can offer explainability (e.g. reflect on which features in input are the cause of particular outputs, and why) - this is welcome as it brings them into the same body of work as much of earlier basic statistics (including the simplest form of ML, linear regression and random forests) - there are good engineering reasons to have explainability whether the application of the tech is in, e.g. plain old engineering (autopilots) or health, but especially so when the domain is very human facing, such as (e.g.) law and language.

As mentioned in a recent meeting, I think the social media platforms, with their combination of various news feed ordering algorithms, adverts, filters ("do you mean to retweet that article before reading it", etc etc), basically constitute large language models already deployed in the wild. The idea of "not connecting your LLM to a social media platform" is out of date - meta et al already did. Given the toxicity of such systems, it seems obvious that we should have a Butlerian Jihad against these systems right now.

No comments:

Blog Archive

About Me

My photo
misery me, there is a floccipaucinihilipilification (*) of chronsynclastic infundibuli in these parts and I must therefore refer you to frank zappa instead, and go home