,
13 tweets,
5 min read
Read on Twitter
Today's meta-Twitter summary for machine learning:
None of us have any consensus on what we're doing when it comes to responsible disclosure, dual use, or how to interact with the media.
This should be concerning for us all, in and out of the field.
None of us have any consensus on what we're doing when it comes to responsible disclosure, dual use, or how to interact with the media.
This should be concerning for us all, in and out of the field.
I know each person and organization and they're all well intentioned. Some are getting aggravated at each other but I believe that's as they all see this as important and are reacting to a live-fire exercise split across the already contentious areas of {ML, PR, AI hype, ...}.
The work that caused the kerfuffle was a large scale language model from @OpenAI. Think of it as a super powered version of the predictive text on your phone that has read more data and can generate fairly coherent text.
The @OpenAI team released the paper and code for training the LM but not the largest of the pretrained models. The reasoning for not releasing the underlying model was concerns around dual use
(dual use tl;dr good tech that can easily be used for bad)
(dual use tl;dr good tech that can easily be used for bad)
Whilst @OpenAI have no requirement to open source the code or the pretrained models many in the community are troubled. Some are focused on reproducibility of science, others on dual use, others on the PR narrative, ...
As there's no consensus anywhere there's no agreement.
As there's no consensus anywhere there's no agreement.
The code for this work was released today and the technology hasn't changed substantially since last year. Larger groups are likely able to replicate the results - or at least an appreciably similar result - quickly. This impacts reproducibility for many smaller groups though.
No release aligns with @OpenAI's charter but leaves many in the community unclear as to considered risks. The ML community can't necessarily discuss the threat model as in the worst case you may be saying "Oh, that house has a key under the doormat".
Releasing a "restricted" model in this way has other (intentional or not) consequences - primarily that "AI is scary" is cat nip to reporters. Hence by acting like a good dual use citizen you can accidentally provoke the AI hype beast.
This also runs in to the issue that research from big labs often ends up Press Release'd rather than released to the research community. As noted, few in the community knew it was coming out. I was a researcher that journalists reached out to early but...
...even if a journalist has good underlying sources, they're in a zero sum game. Multiple journalists are publishing on the research and if they cat nip their story with "AI is scary" they can likely win more audience. This hype escalation is not pretty 😶
This isn't the first time that @OpenAI ended up in this predicament. Their 1v1 DotA bot caused a great deal of angst in the ML community as it was released with nearly no details and a grand PR reception. This time they're explicitly better but the modern era has new issues.
As everyone has a different point of view, it's just collisions everywhere :S
- When does a model go from "safe" to "dual use"?
- How much of a "dual use" delay do we need to add?
- Should we release to journalists first or researchers?
- How can small labs participate in PR?
- When does a model go from "safe" to "dual use"?
- How much of a "dual use" delay do we need to add?
- Should we release to journalists first or researchers?
- How can small labs participate in PR?
None of these questions have black and white answers and we don't even know what the shades of gray look like.
