adtac

9 ms·

91.
by adtac 2y ago
everything works fine with just tmux + mosh, but I like to bind sshd on the wireguard interface so that there aren't any listening ports accessible from the public internet
92.
by adtac 2y ago
tmux over mosh over wireguard is all you need
93.
by adtac 2y ago
The aphorism typically says cache invalidation is hard. Not because you don't know what index to invalidate but because it's hard to invalidate the thing at the right time. Caching itself is quite easy, just ask the designers of
94.
by adtac 2y ago
> triple the training From what I understand, not quite. It looks like the cost of training might be similar, but less parallelisable within a specific token sequence. This is because they have to compute the KV of token T before they ca
95.
by adtac 2y ago
The KV cache is just another tensor to be used with matmuls. Unlike the model weights which are fixed, the KV cache is uniquely constructed for every input. Think of it as the model growing new weights to represent the new knowledge it lear
96.
by adtac 2y ago
1. LLMs want to mimic conversations on the internet 2. Disagreements on the internet usually end up toxic 3. Anything perceived to be toxic has been hammered out with RLHF Play sycophantic games, win sycophantic prizes
97.
by adtac 2y ago
>Token volume needs to double just for revenue to stand still Profits are the real metric. Token volume doesn't need to double for profits to stand still if operational costs go down.
98.
by adtac 2y ago
Maybe I'm missing something basic but that still doesn't explain why it's 30% and not 50%. I don't think purely qualitative arguments work here.
99.
by adtac 2y ago
DCF and other tools to value companies make sense when the valuation is somewhat stable, but the real world often isn't. MSFT was worth $1T a few years ago because the world presumably expected $1T in profits over the lifetime of the c
100.
by adtac 2y ago
That explains why there's a premium but not why the premium is ~30%.
101.
by adtac 2y ago
The other replies explain the 30% with circular reasoning and I don’t find them convincing, so here’s a more absolute and testable hypothesis: at the average rate of S&P500 return adjusted for inflation, 30% is about 3-5 years of inves
102.
by adtac 2y ago
> properly signed packet How would an adversary do this without the private key? Wireguard uses ChaCha20Poly1305, an AEAD scheme. The first A stands for Authenticated.
103.
by adtac 2y ago
No, it just tells me inflation isn't a single scalar multiplier for all kinds of items.
104.
by adtac 2y ago
You would've probably sold it when it doubled. Or 4x. If you still hadn't sold it at 17x, you'd probably never sell it, so it's not liquid anyway.
105.
by adtac 2y ago
As always, code is the best documentation: https://github.com/ggerganov/llama.cpp/blob/8dd1ec8b3ffbfa2d...
106.
by adtac 2y ago
If you're on Safari, open devtools, click the monitor icon on the top left, disable javascript, reload. The content reads fine without JS.
107.
by adtac 3y ago
FWIW RSA_public_decrypt is an 90s way of saying RSA_signature_validate
108.
by adtac 3y ago
Lol the irony of publicly announcing the addition of end-to-end encryption in one app (Whatsapp) while secretly breaking TLS in another, all in the same year #Tethics
109.
by adtac 3y ago
It was a few added characters in a header file to make it possible to deliver the actual payload: 80+ kilobytes of machine code. There's no way to actually tell, but I'd estimate the malware source code to be O(10000) lines in C.
110.
by adtac 3y ago
And now for my next trick: smuggling a backdoor into iptables.
111.
by adtac 3y ago
It will still work if the connecting client offers a RSA key. The only real way to be sure it's not on your system is if your liblzma version is strictly less than 5.6.0 (first infected version): ls -al $(ldd $(which sshd) | grep
112.
by adtac 3y ago
Ironically, the LLaMA license text [1] this is lifted verbatim from is itself probably copyrighted [2] and doesn't grant you the permission to copy it or make changes like s/meta/dbrx/g lol. [1] https://github
113.
by adtac 3y ago
None of the evals are binary choice. MMLU questions have four options, so two coin flips would have a 25% baseline. HumanEval evaluates code with a test, so a 100 byte program implemented with coin flips would have a O(2^-800) baseline (may
114.
by adtac 3y ago
What security vulnerability becomes possible with native CSS-in-canvas support that's not already possible today? Or becomes easier?
115.
Show HN: Greasemonkey script to get Perplexity answers in Google(github.com)
2 points by adtac 3y ago|0 comments
116.
by adtac 3y ago
>vi is the only sensible option Of course, that's a tautology. What? It's flamewardnesday.
117.
by adtac 3y ago
If we model the game as someone flicking switches, strategy is ability to know which switches to flick when whereas technical skill is the ability to quickly and precisely flick the chosen switches. In more complex games, there are more swi
118.
by adtac 3y ago
If Dota was twice as complex, do you think an AI would be more than 2x better or less compared to your scenario? I suspect the more complex the game, the bigger the advantage over humans.
119.
by adtac 3y ago
For anyone deeply familiar with building one: what are the biggest problems you run into 3 months in that you didn't foresee? Just curious because IME that's the point where the fun problems surface :)
120.
by adtac 3y ago
Even if the cookie size limit was 8192 bytes instead of 4096, I think it's insane to include a several KB signature on every request over the network (literally worth multiple packets). It's not a big deal if the payload is a larg

More ›