Each of those agents ends up making "decisions" that lead it to look at some things over others. Given infinite the same agent could eventually fully explore all those options, but each one explores things a bit differently due to different forks in the road due to randomness in token generation. Thus sharing information is useful.
An LLM is not me. It is capable of doing more "work" but it also doesn't have my mental models and experience. My colleagues know roughly my areas of expertise. If I say "I think that X is happening" they'll assume that is the result of my expertise being applied and react accordingly. If I say "[LLM] thinks X is happening" they'll react differently and I want them to. Just like I might say "I suspect X is happening because Y" or "X is definitely happening" when I want to modulate their trust in what I'm saying.
the phrase "[LLM] thinks X is happening" is good chunk of the problem
it didn't. it produced an output, and you hopefully invoked some human judgement before believing it was worth sharing, and you are hopefully expecting to be accountable if that judgement was incorrect (whatever the failure mode)
Fine. "The LLM produced a chain of reasoning consistent with a belief that X is happening."
I've found it fairly helpful to anthropomorphize systems to communicate things about their behavior, despite them obviously not being human ("The load balancer thought the West US region was chilling", or "that cluster was unhappy because of bad packet loss"). I don't really have many problems with people taking those statements that I think computers have emotions.
So the theory is that the page is faulted in, then somehow evicted within 10 instructions, then re-faulted in somewhere else, resulting in writes not making it to the page? That would need a context switch and another page fault to happen in short succession. But the context switch to evict the page back out would have necessitated pending writes to have finished. Nothing actually makes sense with that explanation.
The way I read it was that it wrote to a page and then did a read on that page within 10 instructions and that was not enough time for the memory system to have gone through the process of creating the page backed with real memory and that the timing bug widened in Linux 7.0 such that more things like this exposed the bug. Or there was flapping in the TLB for some reason, or the locking wasn’t correct, or whatever else that the kernel devs will figure out was the cause.
I’m guessing it wrote to a page and read back zero page as the memory was in the mist of being allocated.
But the write should have triggered a page fault immediately in that case, and should have been blocked on the page being actually allocated before resuming.
yes, i went back and re-read the report. an immediate read 1 instruction later works, but 10 does not. the page marked dirty but pointing to the zero page. some context flapping i assume
Sounded like a race condition causing a correct new mapping in the TLB to be cleared away by mistake, reverting back to the zero page mapping it was before.
The tech's also been there to put cameras everywhere, and to wiretap every phone, etc. We put guardrails in place to control how that tech is deployed.
Not sarcastic, but I probably didn't convey the subtlety of what I was trying to say in a one line comment. I was objecting to the defeatist "oh the tech is there, so we can't do anything about it" attitude. I tried to choose the examples I chose that the tech being there definitely has some consequences and significant privacy implications, but some controls exist too (like, wiretaps are still applied very selectively, there's been a growing movement against Flock cameras and scaling back of their deployments in some places recently).
To pull the example of Discord since the ExHashRing was mentioned in the OP: Needing to hash a few hundred things instead of one thing adds up when you do it a lot of times. They went with consistent hash ring over rendezvous hash because of that; every message needs to do one of these hash ring lookups, also whenever someone connects, they need to do a lot of these hash ring lookups to find all of their servers and friends.
There's plenty of scale below FAANG where efficiency matters.
The hierarchical (log(n)) approach to bucketing here is fine for an "I just want to shard this N ways, N will never change" but is extremely intolerant of bucket mutations.
Part of the point of rendezvous hash and consistent hashing is that adding and removing elements minimizes the amount of things reassigned. That is, if you add nodes, the only items being reassigned are those that are moving to the new nodes. If you remove nodes, the only items being reassigned are those leaving the departing nodes.
If you know your set of nodes never changes, or you don't care about the cost of reassignment, you don't need a rendezvous hash or consistent hash, you just need a plain old hash function.
> Now, if someone searches for an espresso machine, Gemini will pull up your most relevant products and instantly write a custom explainer highlighting why your product may be the right choice for them.
This is like the essence of the evil of AI ads distilled down to one sentence. For an advertiser this is a dream. For a user this reads like getting bombarded with ads tailor made just for you based on the context of what would be most effective.
It's just plain fraud. LLMs are hallucinatory, this is a basic fact of their basic design. You can't have them write product descriptions especially in advertisements, without any human supervision.
If the LLM invents a product feature that doesn't exist, you have advertising fraud done fraud. And if the LLM un-invents a feature that does exist, you have done fraud and pissed off the advertiser.
To not have these risks, you need to play it incredibly safe. E.g.: The bottom half of the Vertuo Up's blurb is just off the website.
<meta name="description" content="Vertuo Up is our new fast coffee machine, ready to brew in just 3 seconds. Enjoy 6 cup sizes and app connectivity for effortless control. Shop Pearl White.">
This would've been on old-Google. If you're an advertiser, Google is going to charge you their premium rates for a sloppy first paragraph you could've put there yourself if you wanted.
Note how the search query in that example asks for a "compact machine" but the explainer doesn't say anything about the size of the machine. The dimensions are right on the product's webpage. This advertising product doesn't want to risk the LLM fucking up something like the dimensions, so it just does nothing at all.
And the kicker is that none of this has to be a problem. It's Google, they can just ask the advertisers to hand them over a standard-format datasheet, and put the LLM to work figuring out what parts of the data the user wants and include those verbatim. If the LLM hallucinates, it creates a perfectly truthful but slightly less effective ad. If the LLM doesn't hallucinate, you've created an ad product that is better than most product comparison sites, something users want to use.
Well. If you were a for profit company and were offered to get the most effective ads ever at the cost of 0.1% of advertisement containing falsehoods about your product, would you take it?
What about if you know your competitors are taking the offer?
Silver lining: Is there a chance the ads will be relevant? Enterprise targeted advertising sophistication level is currently at "He just bought shoes; wow he must like shoes" or "He's 34-45M; make the reels only boobs.
I listened to a song a few weeks ago... now that song is in almost every page on YouTube for me. Homepage, sidebar, search results. It's just everywhere.
I've already listened to it. I don't want to listen to it again every single day for the rest of time.
Can’t wait for the AI rated™ product reviews. “This coffee maker doubles as a coin bank for your collection of rare pennies. Why brew a boring cup of coffee? Start your day with the taste of copper.”
I do wonder if that sort of system would be responsive to feedback. I would have told it something like "because you advertised this espresso machine to me, I will explicitly never purchase it. It's effectively banned from my household. Never recommend a product again."
Before it was:
"Shop for balding pills",
Now it's going to be:
"Your calendar indicates your mother's birthday is tomorrow, your album photos of her living room seem to indicate she likes elephants. In a past LLM conversation you mentioned she has trouble with technology. A good gift idea is the new Google Geminibook with this animal laptop skin. Add to cart?"
More like, "Your mom's birthday is tomorrow and you forgot again, didn't you? I've gone ahead and added this perfect gift to your cart that will be delivered in time. I'll purchase and ship it now if you'd like?"
But intelligent beings are fundamentally fallible? That's kind of the nature of doing leaps of reasoning: sometimes those leaps are amazing, sometimes they're wrong. It's what's advertised.
