Why is interpretability of AI models important?

We're excited to be able to continue our work (with @begusgasper) on rethinking language, thanks to @schmidtsciences HAVI. We will test what properties of human language (sound change, dialect, and more) appear in nonhuman communication. This research lays the groundwork for reconceptualizing language as an evolutionary information technology applicable across human, animal, and machine communication.

UC Berkeley researchers are developing AI models that learn language by imitating speech, much like babies do. Their work could make AI more efficient while revealing new insights into how humans and other species communicate. bit.ly/4BeESMn

Midas has played a big role in robotics and technology ethics. We propose following Midas's example not through his story as a warning or as a frictionless success, but through his liminal abilities, guiding the field of robotics to consider machines on their own potential rather than human imitation.

Thank you for the feature @HarvardMagazine! We got so many Harvard alumni in this volume we failed to mention Gerardo Con Diaz (08) and Joseph Dumit (Prof) :)

Old conceptual frameworks need an update. Very excited to be awarded the Schmidt Sciences HAVI Award for "Rethinking Language" in collaboration with @ninabegus. Humanities have a bright future! @schmidtsciences

Imagitation it itself imagitation. Read this great book! @mthvn

What should cultural analytics do in the age of AI? We propose a focus on narratives around AI, about AI, and within AI, and a methodological turn towards constitutive rather than decorative humanism. dx.doi.org/10.2139/ssrn.7586…

it’s here!! so proud to be part of this important book. @ninabegus

It is not smart to bet against intelligence that never dies, learns magnitudes faster than us, and is perhaps unlimited in how large it can get.

The most transformative education happens when students learn uniquely human skills over meals and tea time

Read here: begus.substack.com/p/pascals…

If we believe in AI’s ability to control us, and AI is unable to do so, we lose little. If we don’t believe in singularity, and AI can in fact control us, we stand to lose everything. Wager, then, without hesitation, that AI is (able to control us)!

Doomers unite.

Liminality is the most productive way of AI exploration (IMO), and Pierre Hugyhe is the king of liminality. Thanks to @ranjodhd for snapping these at Fondation Beyeler!

"At a time when #AI is starting to outperform us in many cognitive domains, building alliances and identities together with other biological intelligences might be more important than ever" --@begusgasper escholarship.org/uc/item/37t…

“Reframing how we think of language…” read more here…

Read here: begus.substack.com/p/the-nex…

We form our identities always in relation to the other. The other used to be the animals, now the other is artificial, silicon intelligence. This provides an opportunity to identify as a biological intelligence that will unite the natural world.

Perhaps the next Theory of Evolution is the realization that we are just one of many possible and diverse intelligences and that our legal, societal, and value systems should be adjusted accordingly.

Read full essay here: begus.substack.com/p/the-nex… Or: zenodo.org/records/23071867