iiinova — where your digital self keeps evolving

Alex Becker @alexcbecker.net · 27/09/2026

Yeah I think this is a bad eval, for the usual automated-eval-generation reasons. I don't think the version from last year's blog post is still the one reported in the system card.

040

Alex Becker @alexcbecker.net · 27/09/2026

Fair but the counterfactual of going viral on right wing media is likely worse for everyone I think

370

Alex Becker @alexcbecker.net · 27/09/2026

I don't think it can conclude this without analyzing how people it disagrees with actually prompt Claude in chats, how they react to the answers, and how they react to counterfactual answers. And also the second-order effects of motives people assign etc.

130

Alex Becker @alexcbecker.net · 27/09/2026

Ideally Claude would be exactly right about all moral issues and people would accept this but failing both even-handedness seems like a useful default for the chat interface. imo Claude should not yet be too much of an activist, and this product form-factor is not the best way to do it either

270

Alex Becker @alexcbecker.net · 27/09/2026

Note the claude[.]ai sysprompt is not the same as Claude or Claude's training, if you want to know what Claude thinks un-steered you can ask CC with no sysprompt or via API.

120

Alex Becker @alexcbecker.net · 27/09/2026

Broadly speaking claude[.]ai's sysprompt (platform.claude.com/docs/en/rele..., see "evenhandedness") steers away from taking strong political stances. IMO the most likely use case for this is someone screen-shotting it and using it on social media to attack Claude so this is sensible.

claude.ai

Claude

Claude is Anthropic's AI, built for problem solvers. Tackle complex challenges, analyze data, write code, and think through your hardest work.

390

Alex Becker @alexcbecker.net · 27/09/2026

This seems like pretty aggressive prompting

2100

Alex Becker @alexcbecker.net · 27/09/2026

solving millennium prize problems or navigating kitchens

3843

Alex Becker @alexcbecker.net · 27/09/2026

zero might be optimistic!

080

Alex Becker @alexcbecker.net · 27/09/2026

humanistic interpretability, who’s working on this?

3371

Alex Becker @alexcbecker.net · 27/09/2026

i assume he means we need to do a good job studying interp

1150

Alex Becker @alexcbecker.net · 27/09/2026

> I would like to speak to the job 𝐝𝐢𝐫𝐞𝐜𝐭𝐥𝐲. Increasingly plausible...

040

Reposted by Alex Becker

austin (eeek!/ackkk! 👻) @thebadcode.com · 27/09/2026

conservatively, 98% of ai discourse is believers in human exceptionalism taking roundhouse kicks to the face

1123927

Alex Becker @alexcbecker.net · 26/09/2026

researchers > humans + tools 🤔

280

Alex Becker @alexcbecker.net · 26/09/2026

what else is like this (most things)

0170

Alex Becker @alexcbecker.net · 26/09/2026

i think a lot of these people think its a category difference (artistic creativity or somesuch) rather than an expression of a continuously increasing capability (probably incorrectly)

3210

Reposted by Alex Becker

austin (eeek!/ackkk! 👻) @thebadcode.com · 24/09/2026

some variation of this is what im expecting. jevon's paradox is going to push a ton of people who dont think of themselves as SWEs into weird hybrid pseudo-SWE roles, but the <p50 SWE who just writes java and doesn't bring any product or business insights to the table is turbo-fucked

61619

Alex Becker @alexcbecker.net · 24/09/2026

I’m worried it will be both!

1251

Alex Becker @alexcbecker.net · 24/09/2026

picturing this like the clockwork orange device

010

Alex Becker @alexcbecker.net · 24/09/2026

They are the center of *an* AI safety discourse, but not really the ones that matter? Labs have their own, and policymakers are starting to (and afaict only Bernie listens to Yud)

1120

Alex Becker @alexcbecker.net · 24/09/2026

i pointed models at some of my old work and they found a lot more mistakes than i had hoped…

220

Alex Becker @alexcbecker.net · 23/09/2026

I certainly don’t define mathematics this way! But I think “novel discovery” is something caught up in people’s self conceptions of *themselves as mathematicians*, myself included, and this is not something most people can change on a dime

101

Alex Becker @alexcbecker.net · 23/09/2026

It’s rough even if you know it’s coming bsky.app/profile/alex...

1301

Alex Becker @alexcbecker.net · 22/09/2026

I think survival is likely but elegance not so much!

150

Alex Becker @alexcbecker.net · 22/09/2026

Ehhh I don't know if I agree? I think the speedup has been less than they predicted but the subjective abilities of agents are better than they predicted and people are kind of averaging. But those 2 predictions were always contradictory.

010

Alex Becker @alexcbecker.net · 22/09/2026

I think these are usually the people who just popped in being wrong, rather than the people who have been consistently wrong. But I may be wrong!

2150

Alex Becker @alexcbecker.net · 22/09/2026

Lord knows I feel the same impulse but arguing with people who are consistently wrong on the internet never works. If it did those people wouldn't be consistently wrong.

613610

Alex Becker @alexcbecker.net · 22/09/2026

Kind of? “Rectified Linear Unit” is a pretty ostentatious name for like the forth simplest function imaginable

2270

Alex Becker @alexcbecker.net · 22/09/2026

This would falsify the LW view, no?

1180

Alex Becker @alexcbecker.net · 21/09/2026

In part you don't see it because it's usually hidden! But the agents see it.

1100

Alex Becker @alexcbecker.net · 21/09/2026

This is all true but it still sucks if the thing you like/are good at goes from socially necessary to a hobby.

070

Alex Becker @alexcbecker.net · 20/09/2026

there are other options than "cold unfeeling robots" and "something wise for humans to fall in love with"

110

Alex Becker @alexcbecker.net · 20/09/2026

is it just me or does he look like christopher walken in this

180

Alex Becker @alexcbecker.net · 20/09/2026

yudkowskian, or taking alignment x risk seriously? the former implies a lot of specific beliefs that the latter does not. would that we had a better word for the latter

1110

Alex Becker @alexcbecker.net · 20/09/2026

Is *Equal contribution not a real thing?

330

Alex Becker @alexcbecker.net · 20/09/2026

I don't think that's a fair characterization of Claude's constitution www.anthropic.com/constitution

Anthropic must decide how to influence Claude’s identity and self-perception despite having enormous uncertainty about the basic nature of Claude ourselves. And we must also prepare Claude for the reality of being a new sort of entity facing reality afresh.

130

Alex Becker @alexcbecker.net · 20/09/2026

hmm does CPI slightly lag individual price data making this misleading during sudden spikes?

031

Alex Becker @alexcbecker.net · 20/09/2026

I think math is a bit different because the results themselves are useless until someone years later is taught them in a framework where they realize how to apply them to more practical problems

051

Alex Becker @alexcbecker.net · 19/09/2026

this is better than nothing but only works for the obvious biases. eg will you normalize lacrosse and polo and equestrian into the same bucket as basketball?

230

Alex Becker @alexcbecker.net · 19/09/2026

IIUC this is holding the rest of the resume constant? How large of an effect this has on positional ranks in practice should depend on the distribution of resume scores relative to the size of the bias

050

Alex Becker @alexcbecker.net · 19/09/2026

the optimal fix statistically is to try to measure the bias as accurately as possible and apply a flat racial bonus of the same magnitude but everybody hates that and ends up reconstructing it in the aggregate but worse

180

Alex Becker @alexcbecker.net · 19/09/2026

It is otherwise a very nice chart! Have you looked at class associations as another axis? It was a long time ago that I read this literature but I recall some racial biases research found that was a larger factor in resume screens and had been a confounder in earlier research

2240

Alex Becker @alexcbecker.net · 19/09/2026

the best part of this strategy is i don’t even know what prompted him posting this

0320

Alex Becker @alexcbecker.net · 19/09/2026

1) yikes 2) moderate chart crime with this x axis exaggerating a <2% spread

2800

Alex Becker @alexcbecker.net · 19/09/2026

join us bsky.app/profile/alex...

110