adtac

That explains why there's a premium but not why the premium is ~30%.

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 investment. What if that’s the average period of investment (i.e. time between buy and sell) for a typical retail investor for any given stock?

If that’s the case, 30% is the minimum premium at which not only are you speedrunning returns for existing investors, you’re also doing it for the average person who was going to invest in Squarespace today.

> 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.

No, it just tells me inflation isn't a single scalar multiplier for all kinds of items.

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.

As always, code is the best documentation: https://github.com/ggerganov/llama.cpp/blob/8dd1ec8b3ffbfa2d...

If you're on Safari, open devtools, click the monitor icon on the top left, disable javascript, reload. The content reads fine without JS.

FWIW RSA_public_decrypt is an 90s way of saying RSA_signature_validate

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

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.

It's actually pretty sophisticated. You don't accidentally write a in-memory ELF program header parser.

And now for my next trick: smuggling a backdoor into iptables.

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 lzma | awk '{ print $3 }')

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.com/meta-llama/llama/blob/main/LICENSE#L65 [2] https://opensource.stackexchange.com/q/4543

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 (maybe not that bad since there are infinitely many programs that produce the same output). GSM-8K has numerical answers, so an average 3 digit answer implemented with coin flips would have a O(2^-9) chance of being correct randomly.

Moreover, using the same axis and scale across unrelated evals makes no sense. 0-100 is the only scale that's meaningful because 0 and 100 being the min/max is the only shared property across all evals. The reason for choosing 30 is that it's the minimum across all (model, eval) pairs, which is a completely arbitrary choice. A good rule of thumb to test this is to ask if the graph would still be relevant 5 years later.

What security vulnerability becomes possible with native CSS-in-canvas support that's not already possible today? Or becomes easier?

>vi is the only sensible option

Of course, that's a tautology.

What? It's flamewardnesday.

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 switches and the current set of best switches changes faster. With more switches, it's harder to know which are the best switches because the future is less predictable. And even if we figure out the best ones, they might change before we flick them. And even if we get around to it in time, we might fat finger it and accidentally flick an adjacent switch. And our opponent never gets tired or injured.

This is why I suspect AIs have a much higher ceiling even if we limit them to half the APM pros have. Better strategy matters less, but I admit it's our only chance lol.

FWIW, I've never played Dota but I've played a lot of AoE2 and from what I know they're similar enough (but maybe someone can correct me).

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.

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 :)

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 large media file, but JWTs are most commonly used in the kind of API requests where the body is often empty or something tiny like {"success": true}. It'd be a shame to undo all the wins from things like smaller TLS handshakes using ECDSA (more of a pleasant side effect, but still).

Is YC accepting applications for interstellar body rental stations like Hertz is for cars? I'd bootstrap it, but I think this requires venture scale funding.

No, you are your consciousness. The self exists only in the story the mind tells itself, so both versions would think they are the original you.

Besides, the serialisation process is a form of quantum measurement. Depending on how coarse-grained it is, there might be no way to take a snapshot without modifying you (maybe the measurement process turns the original brain matter into soup).

Of course, everything is. Doubling the speed of light means your network packets get there twice as fast, but accelerating matter to relativistic speeds, which too is limited by the speed of light, has less marginal utility from the doubling when it comes to energy needed for acceleration/deceleration and time dilation.

With sufficiently advanced technology, travelling between stars will be more of a transfer of your consciousness using interstellar WiFi (remember to set TCP_NODELAY!) rather than transporting slow and heavy atoms. You just get transferred from one biological substrate to another. None of the time dilation, all of the space exploration.

The opposite is true too. Things that used to be expensive are abundant.

An encyclopaedia would have cost you thousands of dollars, but you can use the greatest and most up to date compilation of human knowledge for free today (if you can deal with a few donation banners from the Wikimedia foundation around December every year).

Douglas Hofstader has a talk on this topic called "Albert Einstein on Light; Light on Albert Einstein" that I often revisit: https://www.youtube.com/watch?v=ePA1zq56J1I (watch 10:30 to 12:00 if you're pressed for time, but I recommend the whole thing)

Even a SHA-256 hash would leave 10^31 margin of error. The "percent of a percent of a percent of a percent" part is sweeping a lot under the proverbial rug here ;)

Purely probabilistically, trying things off the beaten path is just a matter of higher LLM temperature. Turning up GPT-4's temperature is basically an expensive /dev/urandom pipe today, but I don't see any fundamental reason why LLMs can't catch up. Maybe all it takes is tinkering with how temperature is calculated.

I must be getting old because I almost wish they didn't add this as it's slightly ambiguous whether it's ranging from [0, 10) or [1, 10] and anything ambiguous is probably going to haunt me at 3am some day

> build/test speeds weren't significantly impacted

it must be getting a little repetitive to test every version and find that Go continues to build crazy fast every time :)