You can set a low max_output_tokens and avoid the "go on"s by using prefill; I have confirmed ^[[1]] that when doing so, the output is identical to the output with high max_output_tokens.

1. across multiple providers, except Amazon Bedrock (at least its Anthropic models) for some reason ↩︎

Very similar: 'Deciding to eat meat or not won't affect how many animals are farmed, because decisions about how many animals to farm are very coarse-grained.'

I think a lot of people got baited hard by paech et al's "the entire state is obliterated each token" claims, even though this was obviously untrue even at a glance

I expect this gained credence because a nearby thing is true: the state is a pure function of the tokens, so it doesn't have to be retained between forward passes, except for performance reasons; in one sense, it contains no information that's not in the tokens; a transformer can be expressed just fine as a pure function from prompt to next-token(-logits) that gets (sampled and) iterated. But in the pure-function frame, it's possible to miss (at least, I didn't grok until the Introspective Awareness paper) that the activations computed on forward pass n+k include all the same activations computed on forward pass n, so the past thought process is still fully accessible (exactly reconstructed or retained, doesn't matter).

I'm hopeful the details can be fleshed out in late crunch-time.

You don't want to be figuring out ethics in crunch time, you want to be figuring out how to pass off (or buy time to pass off) the question to something-like-CEV or a long reflection.

that particular case of mass hysteria is a myth

A cynical theory of why someone might believe going insane is the default human reaction: weaponized incompetence, absolving them of responsibility for thinking clearly about the world, because they can't handle the truth, and they can't reasonably be expected to because no normal human can either.

Yes, this. That part of the OP reads to me like

• fraud / fare-dodging makes you a bad person
• ⇒ banks/transit agencies/Amazon ban you because you're a bad person
• ... which is appropriate in this case because they have good info + specialization in evaluating that type of badness

but that's not what's going on at all; the entities in question aren't trying to evaluate someone's overall character or enforce norms for social benefit.

Banning someone to protect infrastructure just doesn't extend to banning someone because they abuse their family, it doesn't create a question of where to draw the line.  

It seems to me like the OP tacitly assumes (& is confused by) something like 'all punishment/enforcement-like actions should be justified in terms of universal morality, not particular responsibilities and interests'.

... fundamentally one of the driving challenges of postmodernism is to actually understand rape

Why do you suspect this?

Yet Western society answers with a firm "NO".

With a stark exception for caffeine (maybe more like an exception from the category 'drugs' than the generalization 'drugs are bad').

Paranoia can itself constitute being fucked over by an agent that can induce it while not being weakened by it, a la cults encouraging their members to broadly distrust the world leaving the cult as the only trusted source of information, or political players inciting purges that'll fall on their enemies.