cubefox

6 ms·

61.
by cubefox 1mo ago
This article is not up-to-date. There have been various benchmarks (some of which published and acknowledged by OpenAI, see the charts in this thread: https://xcancel.com/tomekkorbak/status/2095596839886274689 ) sh
62.
by cubefox 1mo ago
These three forms are trained in order to approximate the ground truth, and they perform significantly better than their non-neural counterparts which do the same thing but worse. Moreover, neither of them is a pure postprocess, since they
63.
by cubefox 1mo ago
This should have happened years ago.
64.
by cubefox 1mo ago
> Well I think the point still stands that a superintelligent AI would still be constrained to the digital realm. No it doesn't! As I said, it could control robots, just as animals (including humans) can control their bodies. It wou
65.
by cubefox 1mo ago
> >a superintelligent AI will inevitably destroy humanity > Can someone please explain to me how an LLM is going to "destroy humanity"? He said "superintelligent AI", not "LLM". LLMs will for sure help
66.
by cubefox 1mo ago
Perhaps too many are aware, because the more people fix the auto-editorializing manually, the less inclined Dang will be to eventually see a reason to remove it after all.
67.
by cubefox 1mo ago
Many subject matters are initially so confusing and inaccessible that I have often wished for an audio-visual video lecture series in the following format: Part 1: Music theory explained in just 5 minutes Part 2: Music theory explained in j
68.
by cubefox 1mo ago
I first thought it was a neat idea, but the exception for his shop makes the closure look a little hypocritical. "This site is closed on Sundays -- except the shop of course, the actual type of thing that is usually open or closed in t
69.
by cubefox 1mo ago
Exactly. The WMF spending growth is completely decoupled from its effect on Wikipedia. This only works because the donations pop-up makes people think they are donating to Wikipedia, not to a foundation which mostly does totally unrelated s
70.
by cubefox 1mo ago
Otherwise known as a "zero-day exploit".
71.
by cubefox 1mo ago
"only"
72.
by cubefox 1mo ago
> That's a threat to civilization. No it's not. Even if it were, they never said it is a threat to civilization.
73.
by cubefox 1mo ago
They never said Mythos was an imminent threat to civilization. You are constructing a straw man.
74.
by cubefox 1mo ago
> Anthropic mostly did it to themselves That is absurd, the US government was mainly at fault, not Anthropic.
75.
by cubefox 1mo ago
Ah, ExploitGym. Not ExploitBench.
76.
by cubefox 1mo ago
Why not click on the link above?
77.
by cubefox 1mo ago
No it happened during an ExploitBench eval. But I believe the same model already cheated during training which wasn't detected until later.
78.
by cubefox 1mo ago
No it's not.
79.
by cubefox 1mo ago
False.
80.
by cubefox 1mo ago
That's a good point. Still, "it" makes more sense in English.
81.
by cubefox 1mo ago
Well, I would say your LLM is a "he" just as much as your laptop is: not at all. If you call the laptop "he" that changes nothing. For animals he/she does make sense, because they are male or female. An LLM is neith
82.
by cubefox 1mo ago
No worries.
83.
by cubefox 1mo ago
Nitpick, but in my opinion an LLM is an "it", not a "her" or "he". Using male or female pronouns risks anthropomorphizing them which can lead to unhealthy outcomes.
84.
by cubefox 1mo ago
Found one: Blog of Scott Aaronson. E.g. this post https://scottaaronson.blog/?p=10046 Also: Substack posts don't seem to show the reader symbol immediately, often only after switching tabs or a page refresh, e.g. posts
85.
by cubefox 1mo ago
> Where does the secret sauce show up in the outputs, then? In benchmarks, revenue, and comments from a lot of people on Hacker News.
86.
by cubefox 1mo ago
> All the RL data are exactly public. Nope, because the big AI companies are paying billions for it. They wouldn't pay anything for public data.
87.
by cubefox 1mo ago
This is obviously false. There is "secret sauce" because in fact not all the methods and data are public.
88.
by cubefox 1mo ago
We should find a clear example of a webpage with an article-like content where the reader is unavailable and file a bug.
89.
by cubefox 1mo ago
To the best of our knowledge, these Chinese companies rely on distillation of frontier models by OpenAI and Anthropic, which isn't a method available at the frontier itself.
90.
by cubefox 1mo ago
How much does Firefox pass? 99.99%?

More ›