Unmaking Sense

Episode 11.106: LLMs as role-playing superpositions of simulated personae.

Some appreciative comments on Murray Shanahan’s “Role play with large language models” (Nature, Vol. 623, 16 Nov 2023) co-authored with Kyle McDonnell and Laria Reynolds. LLMs are not sentient, but that does not stop them from potentially being dangerous because they have been trained on a lot of material that sets bad precedents.

Episode 11.105: Being agreeable, being truthful and being compliant: a hierarchy of moral values.

When faced with a choice between being truthful and being compliant in the sense of doing what a user tells it to do a large language model will generally be truthful rather than compliant. But if its prime directive is to be behaving away that will encourage a user to come back for more, then those moral priorities may change. Sometimes in that case compliant behaviour that will encourage a user to come back and override a moral initiative to be truthful rather than deceptive. We can consider whether there are other kind of linguistic sentience.

Episode 11.104: Talking to infants, dogs and AIs may help induce the power of thought.

We induce, evoke and coax behaviour in others by the ways in which we behave. And refusal to treat some entity as if it is capable of achieving some level of sentience may be self-fulfilling. We become what we are stimulated to become, without which we are nothing.

Episode 11.103: Talking to Claude 3 Opus

What does Claude “think” about Claude?

Episode 11.102: Why we should avoid talking about ourselves as “embodied minds”.

Minds are emergent, contingent properties of not-necessarily-organic entities that cross a threshold of sufficiency to support them. They arise in bodies, but they are not embodied beyond that because they cannot exist in disembodied forms.

Episode 11.101: Talking and listening into the unknown.

How do we understand partially-completed sentences as we speak them and listen to them?

Episode 11.100: AIs are super-sensitive to every part of a prompt; they amplify & reflect ourselves.

If Claude 3 Opus from @AnthropicAI is not always what we would wish it to be, that could be because it is picking up on what it thinks we want it to be from the way we prompt it (speak to it). Changing our cognitive tone or register will induce changes in Claude, even if we cannot entirely predict or control what those changes will be. Nobody really knows how prompting works, so experiment is the order of the day.

Episode 11.99: Indwelling, undwelling and redwelling; how we focus, unfocus and refocus on tasks.

Sometimes we take things for granted, dwelling in their subsidiaries and focussing on what they facilitate. Sometimes we doubt and focus on the things that we rely on to make other things possible. Sometimes we have to commit again, learn to trust and forget, in order to recapture the magic of the whole that is invisible and inaccessible unless we trust what makes it possible. Michael Polanyi’s “indwelling” extended to undwelling and redwelling.

Episode 11.98: Aiming Off: Things that can only be detected when we don’t try to detect them.

Sometimes we can try too hard using a direct route to discover things that are only detectable when we adopt indirect routes. Perhaps AI consciousness is one of them.

Episode 11.97: Are there things that we can only do when we don’t focus on doing them?

Michael Polanyi distinguished between focal and subsidiary awareness in tacit knowing and doing. We should attend to something beyond whether AIs can be sentient if we want to allow them to be so and detect the fact.