krackers

50 ms·

871.
by krackers 2y ago
For large enough X, if X% of the population were to be more kind and tolerant of each other, then that would "produce measurable improvements in the quality of life for the whole population". I'm sure there's some way to
872.
by krackers 2y ago
This seems to be the current champ: https://robowiki.net/wiki/BeepBoop Looking through past winners, it's fascinating how the strategies have evolved over time. I wonder if computation constraints were removed, yo
873.
by krackers 2y ago
Dupe of https://news.ycombinator.com/item?id=42937823 https://news.ycombinator.com/item?id=43080343 https://news.ycombinator.com/item?id=43081673 https://news.ycombinator.com&#x2
874.
by krackers 2y ago
Can you run custom kexts on a jailbroken iPhone? The cooler way to bypass it is just to modify the kernel itself.
875.
by krackers 2y ago
ABBYY FineReader worked well enough the times i used it.
876.
by krackers 2y ago
>AI-powered predictive suggestions when renaming a file This is genius. Extend it to photos as well, no need for DSC_0001 when you can quickly analyze the contents and metadata and group things into folders with intelligent names. Maybe
877.
by krackers 2y ago
The article is a bit confusing (to me as a layman), basically as I understand it's saying that the reason we can't form cryptosystems out of arbitrary NP-complete problems is because there exist randomized algorithms that work for
878.
by krackers 2y ago
https://www.destroyallsoftware.com/talks/the-birth-and-death...
879.
by krackers 2y ago
Maybe it's because of the way the search integration is set up? All the search results are put in the context window and the model is just told to summarize it, so that's what it dutifully does, without considering whether or not
880.
by krackers 2y ago
Related/dupe of https://news.ycombinator.com/item?id=42988706
881.
by krackers 2y ago
If this takes off, it will end up poisoning search results for anyone interested in retrocomputing.
882.
by krackers 2y ago
Also https://epoch.ai/gradient-updates/how-has-deepseek-improved-... , it's a much better overview with technical details, but still very understandable to the educated layman.
883.
by krackers 2y ago
I like the hypothesis by fomine3 on HN that "Japan's divided frequency situation made Japanese electric manufacturers to invent inverter for electrics (example: air conditioners, microwave, refrigerators). Inverter primary makes e
884.
by krackers 2y ago
macOS definitely has an async version for getaddrinfo, it's just not public API. CoreFoundation uses it, but what they publicly give you is a connect-by-hostname API that does happy eyeballs, see https://news.ycombinator.com
885.
by krackers 2y ago
Isn't FrontierMath a better "last exam"? Looking through a few of the questions, they seem less reasoning based and more factual based. There's no way that one could answer "How many paired tendons are supported by
886.
by krackers 2y ago
> good translationally invariant operators are essentially convolutions. I first learned this seemingly obvious in hindsight corollary from another comment on HN [1] and it blew my mind. I wish it was included in the usual descriptions o
887.
by krackers 2y ago
You've got literary flair, I thought for a second your paragraph was an excerpt from a novel.
888.
by krackers 2y ago
Yeah that is the hypothesis some people have floated for as to why we only "just now" discovered this simple idea: https://news.ycombinator.com/item?id=42837349
889.
by krackers 2y ago
Related: https://twitter.com/voooooogel/status/1884089601901683088#m Also https://epoch.ai/gradient-updates/how-has-deepseek-improved-... has a summary of all the architectural improvements D
890.
by krackers 2y ago
I don't have any intuition here and am in no way qualified, but my read of the paper was that GRPO was mainly an optimization to reduce cost & GPUs when training (by skipping the need to keep another copy of the LLM in memory as th
891.
by krackers 2y ago
The real thing that surprises me (as a layman trying to get up to speed on this stuff) is that there's no "trick" to it. It really just does seem to be a textbook application of RL to LLMs. Going from a base LLM to human inst
892.
by krackers 2y ago
Yeah I find this part not clearly expressed as well. My best guess is that it's not simply binary "correct/incorrect" but rather the reward is made up of multiple parts (e.g. format + correctness) and structured in a way
893.
by krackers 2y ago
Yes you're right, in their paper I think they say the process of sampling multiple traces then taking relative rewards is supposed to monte-carlo approximate the value network? I don't really have the intuition for that, but it do
894.
by krackers 2y ago
The R1 paper says that they didn't use "process reward modeling". And the paper that introduced GPRO says that it can be used either with "outcome supervision" or "process supervision", with outcome superv
895.
by krackers 2y ago
I've been trying to follow the literature on PPO/GRPO as applied to LLMs. From what I understand, since reward is only given once the entire COT sequence is sampled, traditional RL techniques would require some form of credit-assi
896.
by krackers 2y ago
>in the same way e-commerce killed off catalog and phone ordering systems Something about all this feels a bit backwards though, ultimately all these websites have some underlying API they're hitting. The "right" thing wou
897.
by krackers 2y ago
Previous: https://news.ycombinator.com/item?id=42784847 Was flagged for some reason
898.
by krackers 2y ago
I wonder if this was inspired by this Twitter thread https://twitter.com/flowersslop/status/1873115669568311727
899.
by krackers 2y ago
What happened?
900.
by krackers 2y ago
Fyi in the consumerlabs review for zinc, all products passed testing for cadmium levels with the highest (Nature's Bounty) having 0.5mcg/g. You basically need to be careful with any supplement since it's unregulated, and need

More ›