ozozozd

Solid explanation. A+.

That’s EA entering the chat, arguing that all currently available ways of helping are ineffective and until there is an effective way to help is found, they are absolved from helping.

And since they are one day planning to help, they must maximize wealth in the interim. Obviously this wealth maximization can take unorthodox paths because ends justify the means.

It’s very distinctly anti-EA, as in EA is known to advise not to do these things, because it’s not the most “effective” altruism.

And people don’t like them for that exactly, quite deservedly.

Thrift stores? Really?

Super cool!

I wish the controls were a little easier on mobile.

They said AGI is here. What you are describing is RSI.

/s

Always. Folks even believe all of us including themselves are holding it wrong.

Kind of like how we all fail to see the bearded guy in the sky.

It is now! Motion denied!

Next!

It worries me that so few people realize that this is a known phenomena: https://en.wikipedia.org/wiki/Messiah_complex

I am astounded you thought about this but didn’t buy it!

Yeah I never understand this idea that “if you avoid getting caught long enough, you deserve to enjoy the spoils of your crime.”

It’s cool and all but this communication style where the lede is buried, and then reburied, is really really really annoying.

You may get paid by the tokens printed, I don’t get paid by the tokens I read.

A lot of them are definitely flirting with Messiah complex. Not sure if that’s mental illness or just extreme arrogance.

“Beliefs” are reasonable in a technical field?

Hypotheses would be reasonable. Beliefs, almost by definition, don’t have much to do reasoning, and thus, definitely not _reasonable_.

I appreciate that you have correctly designated them as beliefs though.

If you are a fast typer, maybe the longer query might just come out naturally.

If you are properly lazy, you must expect or at least attempt to get results for the shorter query. Especially on a phone.

You are incorrect. The article explicitly mentions the crawler/screenshotter service and httpbun.

Fully sandboxed means no Internet access. You can also specify which packages are accessible and put it in the sandbox. Or you can be lazy and give them access to a package manager that had Internet access, but you don’t get to say “we intended to fully sandbox it.”

Not sure why the reproducibility is a requirement that would contribute to the security. Not that fully sandboxing is harder with reproducibility, but that is a moot point when reproducibility isn’t a requirement.

OP pointed out clusters being hijacked specifically being a bigger concern than rogue clusters, your comment hijacks their comment to talk about “rogue clusters.” Or perhaps this is a promotion for Dwarkesh?

Well, the heat death of the universe hasn’t happened yet, but the monkeys became homo sapiens.

Completely agree.

And I further dislike Anthropic for creating 2 situations in 2 weeks that are giving this admin an obvious opportunity to be right. The other one is David Sacks responding to the “Pace the Frontier” crap. These were the table tennis equivalent of lobs for the admin.

Also worth noting that no one knows what kind of task would get a rejection with Anthropic’s models. The documentation is near-non-existent with the air of “security,” rules change daily, the harness is buggy, the clients are buggy, and very likely the backend gating and safety mechanisms are buggy. Even if the DOD assumed the best of intentions to comply, the reliability isn’t there.

What military would buy guns that sometimes stop working, and not necessarily for reasons anyone understands, and still be taken seriously?

Also Claude clearly had no issues identifying a school that has been a school for over 10 years as a military target. Why would the military buy tech that would make them more likely to commit war crimes?

Thank you. Super impressive.

I rarely notice usernames on HN, but after reading this comment it was so easy to guess that the classy contribution on another comment about legal liability was also yours.

And you are right again! The Australian government must compensate OpenAI for having shitty security.

It must be a bliss to transact with you in real life!

Can you elaborate a bit on the semlabel? And/or share a repo?

Did I read this right?

This repo is really outperforming the OG Jev in the public benchmarks?

There was no time to benchmaxx. How is this possible?

I can’t tolerate anything other than Pi anymore.

Why? - session management - accurate (unlike OpenCode last time I tried), tree structured (allows me to jump back and forth to manage context growth), compaction is just another message and pi manages which subtree to send to the model - fast - super customizable - no hidden stuff - sane defaults - it is a shame that this is a differentiator.

Man, it shows that someone was getting paid by the word.

Thank you!

I suppose GPT models throw shade when your input is less favorable to Codex:

“On the surface, Unreal Agent achieves the same outcomes with fewer model turns and fewer input tokens.”

(Emphasis mine)

The frontier labs have dominant market share and distribution. They also have the majority of the compute.

They can make a lot of money by cutting down on R&D and inference costs until competitors can catch up.

They can also buy competitors that catch up.

And if they can get the government to regulate in the meantime, it’s party time for the monopolists!

30kb model is super impressive.

What’s the model architecture?

The title is the conditional that underpins the claim.

Do you read our AI output end-to-end and decide that it’s the accurate content communicated concisely? If yes, this is not an issue.

If you are sending people generated text you haven’t read, how do you know what you sent doesn’t fall in the category of writing the author describes?

I also avoid signing in to iCloud to prevent Apple from scanning my files and photos.

I think the App Store sign in is separate from iCloud btw.