A thread for unboxing AI

Let's make a deal. I'll keep posting in your preferred manner if you stop brutalizing the poor and/or wishing for them to be permanently shoved into a concentration camp. Agreed?

You remember why I got banned? Creepy! Also creepy is your monitoring of how often I call people pedophiles. I got banned for that? I wouldn't know. I don't spend time ruminating on the bad behavior of other posters. I'm a lot less interested in whining to the refs as you. And look, it's not my fault that pedophiles run the world. I'm just commenting.

I wonder if you have the ability post in a non-smarmy style. You constantly go to this well of "you may want to reconsider", "erm, a bit of advice", "you don't really strike me", etc. Grow some balls and simply say what you feel.

if you stopped being outraged at sky and the ground and everything you came across, you'd see i was the one defending the poors not attacking them and this is all your own projection

This specific thing: it's not just a job; it's an adventure. That thing. You're not just on the bleeding edge of physics and consciousness, you're seeing directly into the Mind of God (typical LLM bs).

Llms ♥ this more than anything.

That's just my inimitable style; clearly inimitable, since your attempt to imitate it was pretty piss poor.

Listen, son, I monitor whatever the **** I want. Now, where's your hall pass, bruh?

Also, I have never "whined to the refs" about anything. And what's with this commie schtick that pedophiles run the world? Do you guys learn that that **** by rote in some dark corner of Twitter where you all go to mainline your propaganda? I only ask because your comrade Victor also posts exactly the same thing now and then. But I'm sure you free thinkers all arrived at exactly the same conclusion by chance alone.

I understand how LLMs work. I simply was noting that I had never asked an LLM to generate this sort of response and therefore I didn't have a good frame of reference.

i'm genuinely surprised you haven't used them heavily in legal work - but i can see how most stuff which would be ai relevant would likely be the work of a junior level lawyer or paralegal

I am not disputing whether the posts were generated by an LLM. I obviously don't know for sure, but it certainly seems possible, and you apparently read more LLM output than I do, so you may be better at spotting telltale signs than I am.

I disagree vehemently with the idea that we can know for certain that the posts were written by an LLM because of the "rhetorical sophistication." Any reasonably intelligent person should be able to write with that level of "rhetorical sophistication."

I have read plenty of Westlaw AI output. Any lawyer who uses an all purpose LLM (as opposed to a law specific LLM) for legal research, brief writing, etc., is an idiot.

i get that, i'm not doing a good enough job articulating what i mean

but once you see the patterns, day in day out, it's impossible to unsee them

and it's not like i'm going to say "oh that's well crafted and grammatically correct it must be ai" it's that an llm has a very narrow range of greatest hits that it relies upon repeatedly and when you work with it day in day out - even if it's just discussing coding with you, you start gaining an inherent feel for their output

it's kind of like in gladwell's blink - where he discussed that forged statue - it passed extensive scientific testing which indicated it was genuine, but the experts took a single glance at it and just knew something about it was off despite not being able to articulate what it was specifically was that was off

yes i'm sure the top firms use some sort of wrapper over the llms much like we use cursor as a wrapper over whichever model we choose to power it in coding

but westlaw is still powered by all purpose llms from anthropic/openai/etc and it's still going to make some of the same stylistic choices as a result despite that it is caged within sanctioned legal libraries and methodology which help reduce hallucinations or lying

cursor makes coding so much better for the same reason, it acts as an assistant to the llm, saying "hey here's the codebase, this is how x works, this where you'll find the y library, and here's all these tools and integrations that you'd struggle to do on your own"

westlaw is not infallible either, it's specialized nature reduces mistakes and improves output for legal usage dramatically, but some of the most notable legal blunders due to trusting bad ai output came from firms which used westlaw

I feel like I saved a ton of time not reading any Malcolm Gladwell books.

https://www.reuters.com/legal/legalindus...

Westlaw's AI tools of course are not infallible, and any lawyer who doesn't check to make sure the output is correct deserves his or her fate. That said, the main reason why there have probably been more legal blunders attributed to Westlaw than to ChatGPT is because the use of Westlaw is so ubiquitous in the legal profession. All firms use Westlaw.

agreed, all his stuff is problematic, and his amazing prose hides all the scientific contradictions within his works, only read blink and outliers myself

blink literally contradicts itself halfway through, with first half praising split second decisions and second half talking about the disastrous consequences behind it

and i wasn't the slightest bit surprised when the scientist for whom outliers' work is based on slammed him for completely fabricating and misrepresenting his theories

but nevertheless, like robert greene, he's a very good writer who weaves interesting anecdotes together despite that the underlying message behind it may be trash

oh yeah, i'm sorry if it sounded like i was implying only the top tier firms would use it and can see how the way i phrased it would make it read that way

I wouldn't characterise referencing output generated as part of the same conversation/session as recursion. It's more akin to memory, or changing the internal state.

Recursion is specifically when a function calls itself with its own output as the new input until an exit condition is met. As a simple example, if you wanted to start at 2 and continue squaring until you get past 10,000, you could have the square() function call itself however many times is needed until you get to a number exceeding 10,000, then stop and return that value to the caller, which would look something like this:

[code]
void main()
{
int result = square(2, 10000);
print result;
return;
}

int square(int number, int ceiling)
{
if (number >= ceiling) return number;
return square(number * number, ceiling);
}
[/code]

The reason recursion can be computationally expensive or go very wrong is because at compile time you don't always know when (or even if) the exit condition will be met, which can cause many calls to the function thus eating up CPU time, memory, overflowing the stack, or even causing the program to hang in the event of a poorly crafted exit condition or insufficient input validation (for example, think what would happen if instead of 2 I passed in 1 or 0 or -1 to the function above, or if instead of squaring we were raising to an odd power and passed in any negative input).

I get the point you're making, but recursion is not the culprit here.

The study of non-recursive systems with feedback loops is still called cybernetics afaik.

Fun fact: for at least the last 25 years, Google has had this Easter egg:

They also used to suggest "French military defeats" when you searched for "French military victories" but that one is gone now by the looks of it.

I'm sure I don't know the proper technical terms and used recursion wrong. However, LLMs do not use memory in a way that makes human sense or even much sense in traditional computer science where we physically flip bits and the machine has changed state. They mimic memory by continuously comparing every token to every new token, like an author that has to reread his book each time he wants to type a new word.

Why is why "attention span" gets resource-intensive quick for LLMs, especially for long chains of tasks solved by agents that has to keep comparing everything it wants to do with what it has already done.

They can do some shortcuts of course, but the principle still stands.

Also attention means something very specific to an LLM, I usually see the term context window.

In this instance I think it is clear both from context in the first post and quotation marks in the second post that I did not intend for "attention span" to be a technical term. It is just an intuitive human expression for what we are experiencing when engaging with the LLM.

However, don't take that as a snide comment. I like to learn and appreciate the comment.

Yep that's the only way I meant it.

hey guys, don't ruin this by making up... it's not too late to fight

Recursion isn't different to iteration in this regard. Poorly crafted exit conditions of while loops etc. Recursion is a very neat approach to the same thing that can be rewritten iteratively

The clever bods who wanted linux as a free version of unix called it GNU - Gnus Not Unix. This seems like somethign everyone knows but maybe that just me being old.

Recursion isn't inherently different to iteration in this regard, but the fact that recursion is most often used in situations where it is difficult or impossible to ascertain the number of required iterations at design time makes this problem more apparent in situations where recursion is used rather than iteration. Usually if the number of required iterations can be ascertained, as in my trivial example*, one would use a loop rather than recursion anyway. Additionally, in the wild, the vast majority of loops are simply enumerating a collection of some description where the exit condition is trivial and this problem does not arise, hence this problem manifests itself most readily when recursion is actually used.

*Off the top of my head, it would be CEILING(LOG(2, LOG(number, ceiling))). There may or may not be a fencepost error there.

Yes, I knew the GNU thing. There are other examples of this, which I can't think of right now. There's a list on Wikipedia somewhere.