HackerLangs
TopNewTrendsCommentsPastAskShowJobs

cubefox

7,581 karmajoined há 5 anos

Submissions

New serious vulnerabilities spiked around release of Claude Mythos Preview

epoch.ai
156 points·by cubefox·há 7 dias·73 comments

AI-Generated Papers in the NeurIPS 2026 Position Paper Track

blog.neurips.cc
2 points·by cubefox·mês passado·2 comments

Interactive explorer for cybersecurity vulnerability trends

epoch.ai
2 points·by cubefox·mês passado·0 comments

HRM-Text: Efficient Pretraining Beyond Scaling

arxiv.org
2 points·by cubefox·mês passado·0 comments

Claude, Author of the Humanitas

lesswrong.com
2 points·by cubefox·há 2 meses·1 comments

Implications of Predicting the Next Token

lesswrong.com
3 points·by cubefox·há 2 meses·1 comments

The Fall of the Theorem Economy

davidbessis.substack.com
5 points·by cubefox·há 2 meses·0 comments

The Fall of the Theorem Economy

davidbessis.substack.com
2 points·by cubefox·há 2 meses·0 comments

How Go Players Disempower Themselves to AI

lesswrong.com
2 points·by cubefox·há 2 meses·0 comments

Claude Code sometimes hallucinates user messages

lesswrong.com
2 points·by cubefox·há 3 meses·2 comments

Kitten Space Agency: A technical dive into the inner workings of planetary rings

ahwoo.com
1 points·by cubefox·há 3 meses·0 comments

No one can force me to have a secure website!!!

tom7.org
11 points·by cubefox·há 3 meses·3 comments

ARC-AGI solver via Plücker geometry – 316 tasks solved with zero learning

github.com
3 points·by cubefox·há 3 meses·0 comments

The Terrarium

lesswrong.com
2 points·by cubefox·há 4 meses·0 comments

Large-Scale Online Deanonymization with LLMs

lesswrong.com
1 points·by cubefox·há 4 meses·0 comments

Why I'm Worried About Job Loss and Thoughts on Comparative Advantage

lesswrong.com
85 points·by cubefox·há 5 meses·124 comments

Are We in a Continual Learning Overhang?

lesswrong.com
2 points·by cubefox·há 5 meses·0 comments

How AI Is Learning to Think in Secret

nickandresen.substack.com
3 points·by cubefox·há 6 meses·1 comments

DeepMind gives up on mechanistic interpretability research

lesswrong.com
3 points·by cubefox·há 7 meses·0 comments

The inefficiency of RL, and implications for RLVR progress

dwarkesh.com
120 points·by cubefox·há 8 meses·48 comments

comments

cubefox
·há 10 minutos·discuss
> the set of people who care about fast mpmc queues with bounded waiting is very very tiny across even the whole population of HN

Yes, but even of those who do care the vast majority will obviously be male.
cubefox
·há 11 minutos·discuss
No offense, but from the mentioned base rates (HN male to female ratio) I would be surprised if you were biologically female.
cubefox
·há 21 horas·discuss
> I definitely know some girls who'd love this, and see this as having fun.

That's hard to believe. Then you should know at least an order of magnitude more "boys" who would love this: At an estimated 95% male ratio on HN, 20 times as many. If the "some girls" you "definitely" know are 3, then you would be expected to know about 3x20=60 males who are interested in "fast MPMC queues with bounded waiting", which sounds about equally hard to believe.
cubefox
·ontem·discuss
It's also a joke because girls definitely don't care about "Fast MPMC Queues with Bounded Waiting" at all. We can estimate the HN audience to be ≈95% male.
cubefox
·anteontem·discuss
The "or" (and "and") types in TypeScript are set theoretic rather than algebraic, so they don't require wrapping and unwrapping. It seems to me that they are the ones with better ergonomics.
cubefox
·anteontem·discuss
My guess is the classifier guardrails were made significantly stricter to convince the US government to reverse the ban.
cubefox
·há 3 dias·discuss
Yes. You basically still need a few engine programmers to use UE5 efficiently, even if it's not your own engine. UE5 seems to be user friendly enough that most of the game development can be carried out just by artists and game designers, but without engine programmers performance optimization will be poor.
cubefox
·há 4 dias·discuss
In many cases, odds are indeed better than probabilities, namely when a small difference at the probability edges indicate a large real difference.

But sometimes small differences at the edges are indeed small, particularly for expected values. Say you win 100 dollars with 98% probability vs with 99.9% probability.

The expected value (probability * dollars) of the latter is only slightly higher than the former ($99.9-$98=$1.90) even though the difference in odds is very large: (0.999/0.001)/(0.98/0.02)≈20.39. So the 99.9% probability is odds 999 to 1, 98% probability is 49 to 1, so the former has more than 20 times higher odds, but the expected amount of money you win is almost the same.
cubefox
·há 4 dias·discuss
More likely it was because Anthropic invests a lot in interpretability research, unlike other AI companies.
cubefox
·há 4 dias·discuss
Perfect!
cubefox
·há 5 dias·discuss
Maybe "virtual" games. Or simply "non-pysical" games.
cubefox
·há 5 dias·discuss
This post smells of LLM writing.
cubefox
·há 5 dias·discuss
> Basically like passes@6 or passes@5 if you’re doing a benchmark, except for your real tasks.

It's unclear how they would do this when there is no signal that provides an objective ground truth.
cubefox
·há 5 dias·discuss
> makes use of parallel test time compute

Any idea what that means exactly? I vaguely remember that ChatGPT Pro was originally called "deep thought", just like Geminis "deep thought" feature (or "deep think"?), so it seems likely they are using the same approach.
cubefox
·há 6 dias·discuss
And write several subheadings starting with "The"? It adds up.
cubefox
·há 7 dias·discuss
I had to cut the "disclosure" in the title from the HN submission because of the character limit...
cubefox
·há 7 dias·discuss
I think the social effect on adults might have been similar, although more gradual, since I didn't really notice it at the time.
cubefox
·há 7 dias·discuss
Yeah with a battery you can change the screen to "dead battery" when the battery is almost dead.
cubefox
·há 7 dias·discuss
> In addition, the Chamber imposed 12 months of unpaid service for the benefit of the chess community as a supplementary sanction.

I wonder how this is enforced.
cubefox
·há 7 dias·discuss
Now I wonder how it gets recharged.