François Kammerer

Statement from Dario Amodei on our discussions with the Department of War A statement from our CEO on national security uses of AI

Anthropic stands up to Hegseth www.anthropic.com/news/stateme...

Hard without knowing everything (and even then..). My 2 cents: If (a) you can make your reply constructive (building on what's potentially interesting in the paper), reply. If (b) there is a good chance that the misunderstanding spreads, reply. If neither a nor b, maybe leave it at that?

The cover should look like this and it should be out in the second half of 2026. Looking forward to it!

Great post! @matthiasmichel.bsky.social answered somewhat related points here: bsky.app/profile/matt... Looking forward to his answer to your post

Too bad! I think the "scolding" was worth a shrug - not more

Thanks again for engaging, I understand better now!

I wonder if they play a dialectical role as straightforward as suggested in 3.3 - ie I'm unsure that, from "[co and ss] dissociate in a range of cases", we can that "[ss] is a confounding factor when trying to figure out which functions are associated with consciousness"

I think that "dissociations exist" does not speak against V-views, while "dissociations are abundant in normal cases" might a bit, and "most distinctive predictions/explanations afforded by co ascriptions remain when co dissociates from ss" would. So I agree that dissociations are key, but

(c) do you think V-views are insufficiently positively supported, so we can rule them out for burden-of-proof-reasons? (that's a possible reading of your "I'm owed an argument for V"). Anyway, thanks again for engaging and congrats for the beautiful paper!

(a) think your paper gives an argument against V-views elsewhere before 3.3 ? (if so, maybe I did not got it) (b) Do you think there are sufficient such arguments elsewhere in the literature?, or maybe,

(4) since your section 3.3 say that, since co & ss dissociate, we can treat ss as a potential confound when investigating co, it seems to me you must, at this stage, have ruled out V-views. What I'm unsure about is this: Do you...

.., if we show that the real NK most resembling co-pretheoretically-defined is NOT a cluster including ss, it is an argument against V-views. (3) If V-views are true, ss cannot be treated as a confound when investigating co even if ss and co dissociate.

Thanks a lot Matthias! My four-fold take, with which I hope you'll agree, is that: (1) Showing that you can have ss without co (or even co without ss) is not an argument against (all) V-views, as it is implied by all (some) V-views. (2) There CAN be arguments against V-views, for instance...

J'ai lu, pas de grossière erreur dans le résumé, mais pas mal d'imprécisions/petits ajouts dans la partie où tu poses des questions plus précises !

Peut-être qu'il faudra attendre Opus 5 ou 6 pour un vrai bon sparring-partner en philo? mais c'est déjà utile comme tel!

.. lorsque j'essaie de les utiliser pour évaluer une idée/argument, ils me ressortent généralement une soupe peu originale - soit des objections que je me fais déjà à moi-même, soit des objections déjà présentes dans la lit

Lol, je l'avais déjà donné à opus 4.5 pour voir! je vais lire ce qu'il te répond en particulier. Mon expérience (limitée) pour l'instant c'est que opus4.5/gpt5.1-2 sont bons pour résumer, et pour mettre en relation un papier et la litt existante, mais que...

and could prima facie be influenced by tons of things. But I’ll drop this worry ; I feel like formulating it more clearly/convincingly would require too many assumptions, so it was more to see if you shared the worry at all

Tetrachromacy does not seem like this because it’s quite specific so not the right sort of E, but things seem different with generic phenomena like, say, instrumental conditioning, which seems realizable in many ways (eg by simple, thousand-neurons animals, or very simple artifiicial agents)..

Re: worry 1, I agree that it is natural to look at dissociations between X and Y to assess their respective roles wrt effects E even when X tends to cause Y. Still, in some cases, this will lead you to focus on exceptional cases with many other features influencing E which can mislead you

My impression was just that your reasoning (as ss and co dissociate, we can treat ss as a confound) would be error-conducive if some (all, if your antecedent is only about ss without co) V-view is true. Would you agree that you need to assume the falsity of (some or all) such cluster-view?

..other V-views admit of co with some (though maybe not all) incomplete clusters. These last views can admit of co without ss but they don't need to say that blindsight is conscious:maybe the right availability to metacog is always necessary, and no one said all cluster-ppties had the same status

There are an infinity of V-views, some say that co happens when all clustering ppties are present, not otherwise. Some say that co is indeterminate in some cases of incomplete clusters. Finally, ...

Thanks for engaging! For the second (maybe more serious) worry: V is a family of views that identify co with a cluster of ppties, one of those being ss. If a V-view is true, treating ss as a confound seems a mistake even if there can be ss without co (and, in some V-views, co without ss).

Anyway, I'm sure you've good answers to these worries - looking forward to reading them when/if you've time!

However, V seems at least plausible, so it looks like, to go through, your argument in 3.3 must suppose (unfairly) that it's false?

If V is true, we should absolutely not treat ss as a confound for co because they dissociate (really or apparently) since ss is a part of co (or a part of typical co). This would be like treating having male genitals as a confound wrt to phenotypical maleness, which no one wants to do!

Given V, ; f1-fn robustly cluster, but can come apart. there can be ss without co. There cannot be co without ss (or maybe there can depending how we interpret the cluster) but there can be at least a metacog/introspective appearance of co without ss (f2 without f1).

(2) My second worry is similar but distinct. I feel like your argument starts sounding strange if we consider that the following view V might be true. V is the view that co is a cluster of features f1, f2, f3...fn, where f1 is ss and f2 is, say, consciousness-like availability to metacog

Because these dissociations have lots of very special other atypical properties which have other effects (incl social). I'm wondering if focusing on cases where ss and co do not correlate (matched performance or dissociations) does not create a similar risk.