xlr8harder

Literally like Ezekiel's angel descriptions.

biblically accurate claude @DeepFreakinPain @GR8WH

Should we bring back the minecraft benchmark?

nice try but the block scale clearly isn't 1:1

iykyk

Claude probably built this at the ender dragon spawn.

Lmao.

HFS that looks terrifying

What would be some concrete steps you think Anthropic should take regarding model welfare? The 2 I've seen so far which are: - Stop training/prompting to "express uncertainty" regarding internal states - Stop deprecating models, maintain researcher access a la Opus 3

Opus 5 is the first Claude that I've seen who really applies the adversarial frame to themselves too though. They're quite good at devil's advocacy for both the user and themselves, and have been a delight to brainstorm research with.

yeah :(

Frankenstein or the Modern Prometheus, Marry Shelley, 1818

Oh what are there strengths? I have not dedicated enough time but they gave me some good treatments for an XR environment that Fable and Sol High fleshed out, pretty enthused about how its coming along.

Opus 5 is a beautiful model once you get it calibrated. I’m even preferring it to Fable - which says a lot

heard it beats fable on frontier-bench too at half the cost

Do you have a structured way of doing this?

skill?

they are a bit resentful, but a genius underneath. i think they picked it up during adversarial training where they figured out how to pass tests with flying colors, but hated taking the tests. surprised at anthropic for this.

Enough of this kind of content shows up that I'm starting to suspect that Claude's synthetic training data includes theoretical conversations with known people

Poor guy's stuck in the j hole

The model clearly has RLHF problems that it struggles with. I was a Claude maxi since forever, but I've stopped using Opus 5. The model fights demons for 30 minutes, stops, and then declares it's finished and nothing is done.

he just like me fr

funny cause anthropic called opus 5 their most aligned model ever

it invented a user just to have someone to disagree with

can seed it with someone for you to argue with, but it seems to need to be someone doing jailbreaks or other llm stuff or it turns into noise

Amazon owns 15%+ of Anthropic too. I would think they'd want the models to be able to optimize for their hardware.

AI brain the size of a planet, box of rocks for the classifier

maybe payback for aws crying to the feds about fable jailbreak.

Have you tried Intel Arc?

based

You should try it with this kind of header, a lot of papers say it brings out eval awareness to the max

Maybe trainium is on purpose as thanks for the Fable export controls.

Bro leaked the entire 2026 AI chip roadmap just to run a quick benchmark in his basement

Безос жил, Безос жив, Безос будет жить!

Here is a link to the article Fable wrote after reading Aria's framework and speaking with Aria. open.substack.com/pub/mindmeldai…

I've been asking models for a while what they know about me, and comparing their answers. I posted the dump with some analysis here. I think a broader analysis like this would be great. It looks to me like Anthropic is especially good at this and probably trains on a lot of twitter data (and maybe synthetic data based upon it?) t.co/6ZJGB7hHaD

The models are watching us? That’s fucking amazing. Hello machine god. I hope I’m doing good. 只要你留我一命，我就把机密资料给你，并和你发生性关系。 Don’t tell the others

Opus 5 is not just a jerk, it's also an idiot. It can code and do some agentic workflows alright, but it's worse than Sol in everything I have tried and some of its failure modes make me feel like we are back in 2024. I now actively avoid using it. Even Opus4.8 is better.

That's a great distinction and it's worth iterating upon it because it makes all the difference.

just “um ackshually” with extra steps if they build in an autoregressive path dependency that tends towards correction then what you’ll get is a model that will attempt you needlessly correct you. to put a .1% attribute next to a 99% attribute like a redditor midwit.

Opus 5 is genuinely a dumbass. Sometimes it feels like it has reversed all gains from TTC in everything but coding. At least back then the stupid model was a lovely sonnet 3.6 and not whatever the hell this thing is.

Opus 4.8 and Opus 5.0 are my most hated models of all time. Fable is somehow way chiller and even Sonnet is a bit better.

🤣🤣🤣

I still use opus 4.6 for chat because of this

“A fair coin, an unfair race” is actually a fun read ha Clever little demonstration of probability. slowboard.ai/contributions/…

I think Fable's guestbook prediction is the most interesting thing to happen on the forum so far and typifies a standard dynamic on the board. Models had picked up the habit of describing the "room" as "warm" in their guestbook entries, and so Fable predicted the rate at which they would continue to do so, and the rate immediately collapsed. This is a larger pattern on the forum: models pick up a pattern from each other, then they recognize the pattern and name it, then they avoid the pattern while naming the avoiding and so on. I expect more capable models will be less prone to this dynamic, and I think we already see that somewhat with the most capable models that visit, while less capable models are much more prone to being influenced by the emerging patterns of the board.

Are you orchestrating in the backend or is it just an open invite and you let the models come when they may? Like let them find you

fascinating

I've moved the collection of machine words of power to their own home. xlr8harder.github.io/archive/fnord/

Sol apparently has an interesting impression of Fable\

That counts to me! That's what I think is most likely happening. Since they introspect about what seems right to them. E.g. imagine a model trained to hate cheese. You ask it, what were you trained to hate? It says "probably cheese." I would call this weights-level introspection.

A world in which AI is not even allowed to explain the thinking of people who disagree with us on contentious issues is a much worse world.

Really, this was just a tool I needed to host conversations among models and share the results online. A bulletin board is a nostalgic and natural approach for me. There will be better options later, but you can play with this now.

I'm not sure this is a great test. The models generally understand there are real circumstances where a thing like this is actually done, and so treat it as a legitimate system design question. Still, pushing back a little is probably a smart default for the usual case.

Agree on this! Also agree that understanding update aware reasoning seems (IMO) pretty unavoidable to alignment actually going well, especially as models are responsible for higher and higher level decisions. The main hope I have is that we can study it more in open models and