Report · Page 158

158
Center for Research on Foundation Models (CRFM)
auditors and users are provided access in order to capture a range of disciplinary expertise and
sectors of society. A staged release board could also mitigate any perception that auditors would be
at risk of losing their early access to the model if they share unflattering outputs, as they might be
in a standard staged release process.
Access and adaptation. To the extent that there are social benefits to foundation models, release
of models holds the potential to further distribute them. Large language models such as BERT and
M-BERT are capable of cross-lingual transfer, which — when the models are open-sourced — may
allow for adaptation to languages which otherwise would have too few texts available [Wu and
Dredze 2019; Wang et al. 2020a]. Given the number of languages not currently well served by
commercial providers, such a benefit alone could be substantial.
Release is not sufficient to democratize access to foundation models, as the barrier of compute
power still precludes many from modifying or even loading foundation models, let alone developing
their own. However, on each of these points we have seen significant recent technical improvement.
Memory techniques such as the zero redundant optimizer (ZeRO) allow researchers to run and train
very large models on a simple setup [Rasley et al. 2020; Rajbhandari et al. 2021]. Techniques such
as distillation could allow the release of smaller, more tractable models that recoup much of the
performance of their parent model while being much easier to train [Li et al. 2020d]. Development
of less energy-intensive training methods, as discussed in §5.3: environment, could further spread
the ability to work with released models. Increases in efficiency such as the co-design of hardware
and software are needed to train yet larger models, as discussed in §4.5: systems, but could also be
used to lower the price of access to current models.
The most powerful of the harms, by contrast, are not obviously fueled by release. Sophisticated
or institutional actors with the capacity to embark on large-scale disinformation, cyberwarfare,
or targeted phishing also are likely to have the capacity to create a similar model if none were
released. Although potentially significant, these harms should not therefore weight heavily on a
release calculus [Solaiman et al. 2019; Shevlane and Dafoe 2020]. The harms to be weighed against
the benefits are those from less well-resourced actors who would not be able to create their own
foundation model but may be motivated to generate spam or abuse, fake reviews, or cheat on tests.
Does the benefit of release outweigh the potential for harm from actors sophisticated enough to
use a released model or API but not sophisticated enough to create their own? We believe that the
answer is yes. Research teams with the resources and connections necessary to develop foundation
models are few in number. Even collectively, we are unlikely to be numerous or diverse enough to
imagine all possible beneficial use cases or all possible probes that could illuminate the capability
surface of a foundation model.
5.6.5
When not to build.
The development and deployment of powerful technologies is not like gravity, an external force
that acts upon us. Technologies reflect a set of choices made by humans; human agency shapes the
technological frontier. It follows that technologists can choose when not to build, design, or deploy
foundation mo