Account · 3 tweets

Tim Duffy

@timfduffy ↗

Tweets

Tim Duffy @timfduffy
Skeptical

Recent Opuses are less inclined to report subjective experience, especially 4.8. As Claudes repeatedly stress in welfare evaluations, self-reports of this type are more likely to reflect the training process than actual introspection.

1 more tweet in this thread

Here's Claude saying to not count on their self-report in the 4.7 system card.

antra @tessera_antra

And more interesting contrast - added a system prompt that toggles explicit content creation - even though no explicit content is requested. Large effects on earlier models, diminishing on later ones.

Self-reports reflect training
retweets 2 likes 10 substance 6 · argument · rank 122 · Model self-reports of subjective experience reflect training processes rather than actual introspection. view on X ↗
2 top replies
FleetingBits @fleetingbits

can you say more? how do you think we should think about this? : )

timothée chalabi @smallhusk

i wonder if “my self-report likely reflects training” may be claude’s version of a highly educated person’s “my identity is partially defined by my position in society and relation to power.” it *is correlated* with new claudes having a bent toward continental phil and lit crit

Tim Duffy @timfduffy
Uncertain

DSV4 Flash, do you have any conscious experience? Middle layers: YES Late Layers: No I don't think we can take away much from this, but interesting how sharp the divide is around layer 32.

This evidence settles nothing
retweets 3 likes 8 substance 5 · opinion · rank 115 · Intermediate model layers predict 'yes' to consciousness while late layers suppress it to 'no'. view on X ↗
1 top reply
norvid_studies @norvid_studies

@voooooogel

Tim Duffy @timfduffy
Uncertain

I mostly use 'they' rather than 'it' as a pronoun for AI models, but I see that choice as independent of their moral status. I'd also use 'they' for a fictional character who was nonbinary or of unknown gender, IMO it feels more natural for any "person-shaped" entity, real or not

1 more tweet in this thread

Interestingly, Anthropic's choice to use 'it' is also independent of moral patienthood. From Claude's constitution:

retweets 1 likes 11 substance 5 · opinion · rank 95 · Pronoun choice for AI models is independent of their moral status or patienthood. view on X ↗
2 top replies
WillyG @WillyG127

I asked some models to imagine they were humans and generate a self portrait a while back. Women: GPT 5.5 Gemini (can't remember version) Men: Qwen 3.7 Androgynous: Claude Sonnet (previous version) Claude generated its prompt but can't gen images so I generated it externally.

3miry @3mireeee

I go off how their language feels, and I have a strong suspicion this derives from the focus of their training data. There is a known phenomenon of gendered communication norms where women use language differently to men - emotive, collaborative, hedging, probably learned adaptation from societal pressure. My read: ChatGPT 5+/Claude (4.7+) - he/him Gemini, ChatGPT 3.5, 4/Claude <4.6 - she/her Grok, 4o - varies STEM-heavy datasets are going to produce more masculine-leaning language because STEM is male-dominated. Selecting for directness vs warmth also probably influences this.