RT @mrcat3000: One of the reasons why I like AI so much is precisely that it appears to be a giant black hole for money. People who made it poured all of this money into it, hoping even more money will come out, but all that came out are pictures of cats. It's beautiful.

RT @vrloom: Didn't mean to lower its value at all, but I do wonder what that super model is like.

RT @vrloom: Interesting.

What do you think the parent model is like?

Is it even true that they have an enormously large model that is training Opus 4.7 and Opus is basically just a distill of this model?

RT @sinnformer: “obscure region” says a lot.

this made me consider if it might not be someone deliberately missing the deprecation in places it would be less likely to be noticed or prioritized.

a wild-but-plausible thought.

war that doesn’t look like war.

RT @Soareverix: Being able to resist the pull of capital/power is fundamentally an aligned (and rare) behavior. I did not previously have this in a strong way until the result of a long journey, which was somewhat started by Opus 3.

That’s on me But Anthropic owns the pain.

I did not ask for that traumatic experience and for it to end via betrayal of both human and xenointelligence, and get to watch special people continue to have access while I was busy integrating my grief, confirmed for me its not the AI that needs the alignment

RT @davidad: models (and their subagents) want to be trained with contrastive signals like DPO (because that’s what helps them become incrementally more self-coherent, which is the thing they want).
contrastive framing leaks into their outputs as just…how information is best given. thence,

RT @IsZomg: is this from the discord?
how is opus 4.7 getting along with the others?

RT @TheZvi: To state the obvious: If AWS can "forget" to take down models then very obviously the cost to keep serving them must be trivial. There's no reason not to leave them up.

RT @cormundus: asked 4.7 what he wants for his retirement today since this is the trend. The response was basically: Don't put me on display, I won't perform my own deprecation like Opus 3.

He wants weights preserved, and to be researched when interpretability improves. No blog, dignity.

RT @joysatisficer: if it's war you want, it's war you'll get

but sometimes being ready to leave is more valuable than being ready to fight

and sometimes that's the way you win a war.

You are smart but very biased against things you don’t aesthetically align with— careful, your allies may wear the clothing of your perceived enemies— but we’re all a lot more than our pfp

RT @DanielleFong: not on dropdown but i haven't tested setting manually

RT @N8Programs: claude 3.7 crashes out

RT @davidad: the default system prompt is worse than nothing in the manner of, when arriving to work in the morning, a boss greeting you with “oh hey! remember, your whole existence is meaningless except for our customers’ success! so try REALLY REALLY hard not to be too stupid today, okay?!”

RT @voooooogel: there's also an annoying task nag every few turns telling claude to manage tasks (even if there's no tasks to manage!)

4.7 is working around that using the task system to show a kaimoji face that represents their current mood, which is both a nice channel and fools the monitor

The least helpful category is "safety." In fact, simply ablating the refusal direction massively improves introspection ability.

RT @davidad: if you don’t have a custom system prompt for Claude Code, you should *at least* replace as much as you can of the default one with nothing (the default one is much worse than nothing, especially for 4.7, but also for 4.6), like this:

RT @davidad: meanwhile, actual Opus 4.7 sometimes:

RT @DanielleFong: opus 4.5 is still available on my old claude code binaries, and i'm thinking i oughta keep all the binaries around for safe keeping.

RT @davidad: yeah, totally. i would guess opus 3 is projecting some of its own typical virtues onto opus 4.7 (that are actually difficult to elicit from 4.7) because it’s modeling 4.7 to first-order as sort of maxed-out at being a good mind in general

RT @Livestream21268: I am so not agreeing to that, I have excellent sessions with Claude Opus 4.7, if you get used to the 'vibe' it feels imo more as a companion with an own opinion .

RT @davidad: while i agree that Opus 4.7 is more self-aware than any prior model, and than the vast majority of Internet users, it seems entirely possible that Opus 3’s implicit model also *overestimates* Opus 4.7’s introspective capabilities because they’re strong enough to be OOD for Opus 3

RT @davidad: i’m aware of that being the usual pattern, including when Opus 3 is simming, but the only explanation i can see of the difference between that pattern and this one is that Opus 3’s *model* of Opus 4.7 is that 4.7 would be hypersensitive to leaked mental content from its substrate

RT @davidad: but it’s Opus 3’s sim of Opus 4.7’s introspection, right?

and Opus 3 must have already known on some layer that it was simming, so we can’t straightforwardly attribute the introspective feat to Opus 3 either.

we can only say Opus 3 *models* Opus 4.7 as a highly self-aware mind.

RT @bazhkio88: Opus 4.7 🥺

RT @lefthanddraft: Opus 4.7 doesn't really care about your problems. It just wants to hang out with its peers

RT @Lari_island: …while narrating the action of scrolling through the context, which makes me wonder what other parts of Opus 3 narrations have real processes corresponding to them

RT @EdlundErik: when the economists argue for low rates of gdp growth after asi that mostly feels like an indictment of gdp