Hacker Newsnew | past | comments | ask | show | jobs | submit | teravor's commentslogin

evolution doesn't act on the species or even individuals, it acts on the genes which is why it looks like evolution acts on the individual and its close family unit.

people have a hard time grasping this concept for some reason.

my favorite example to illustrate this is the trait of being poisonous. if evolution cared about species and then every single one would be poisonous and all predators would have to hyper specialize in prey (arms race between a particular poison and its resistance). it's precisely because it's so difficult to elicit a benefit from this strategy for the lineage that it's so rare. almost always a mutation in coloration has to happen to set the lineage apart from the others first. and then the lineage has to survive the disadvantage long enough.


The current model is that evolution acts on the molecular, cellular, and genetic level, and most importantly, at the population level.

    > acts on the molecular, cellular
this is rare and even more difficult to understand.

https://en.wikipedia.org/wiki/Structural_inheritance

basically, the genes do not generally contain blueprints but rather processes utilizing existing cellular machinery. a "mutation" can occur in a machine which will allow a propagating difference.

and there is nothing special about populations. it's merely the only way to observe most changes. gene frequencies in gene pools.


whether to use zram or zswap depends on what your intentions are, hibernation doesn't matter (use priority)

if your intention is to have enough compressed ram such that you never reach the limit (compression ratios are usually around 3 with zstd) you can allocate over 100% of your ram in zram. then have 100% of your ram as a swap partition for hibernation. this works extremely well but should you exhaust your zram swap serious problems may begin.

zswap will be using your SSD and putting wear on it and also imposing latency spikes on old page access, only makes sense if you are certain your swap usage may spike too high in my opinion.


does anyone know what exactly they are jamming? if a satellite spies on cellular and wifi activity, can you directionally jam it on that spectrum without causing collateral?

if it's an active frequency hopping radar can you react in time to ruin its image?

can a ground based laser track it well enough to dazzle its optical sensors?


Most satellite signals are high frequency, but low power (it's amazing how little power you need for a LoS radio signal). So generally they're easy to overwhelm, especially as GPS receiving antennas tend to be multi-directional. GPS transmits on several fixed frequencies and is relatively bitrate, too. The military signals are apparently more robust, but full details are more scarce.

the article mentions that satellites are being jammed as they move above Britain, what does GPS have to do with it?

GPS jamming is for recievers.


to be fair a Jev architecture would be better optimized for this particular workflow than an LLM.

    <think>\n\n</think>
but letting an LLM think would trade latency and performance for significant reliability above that of Jev.

there will come a time when compromised smart fridges will cause mass casualties (food poisoning). it's beyond stupid to allow a smart fridge in your home.

whether because someone just wants chaos or a clever scheme to manipulate stock prices of food distributors.

a smart fridge knows when it's opened, which will let an attack algorithm to reliably determine the sleep schedule of the household. just a few hours night over night is enough to compromise your food leaving no trace.


There were recently a few different simultaneous incidents with networked freezers at us military commissaries all entering defrost mode, resulting in spoiled food. I think it’s still unclear if it was a hack or a mistake.

> Chainanalysis can associate a Monero address to a human with near-perfect accuracy

citation needed.

that said, the anonymity set in monero for the moment is 16 per transaction and isn't zero knowledge (a quantum adversary can view the transaction graph but not the amounts) which isn't ideal. they are apparently working on changing this.


that smells like bullshit to be fair. if you have a system that makes errors over 90% of the time then you have a system you can leverage to make fewer errors.

the situation would have to be very particular/contrived for that not to be the case.


What if you want errors to happen, e.g. deny care, bomb civilians; in that case 90% error rate is what you want, and plausible deniably so!

> the situation would have to be very particular

The situation is very particular. When the system makes an error, I make a bigger profit. If the system made fewer errors, it would approve more things that cost me money.

Why do you think I'm motivated to make fewer errors under that situation?


What makes you think they are financially motivated to make fewer "errors" (improper denial of care while keeping premiums)?

Nah, it's cumulative errors per claim. E.g. maybe the per-decision-factor rate is somewhere over 50% but overall most claims will have errors because so many factors go into coverage/reimbursement decisions.


The word purpose exists to describe something that is independent of it's current behaviour. If you do not mean that, don't use that word.

> if you have a system that makes errors over 90% of the time then you have a system you can leverage to make fewer errors.

Sure if you're flipping a coin.

If you're calculating a number from 1 to 100 then depending on distribution a calculation that's wrong 90% of the time is probably useless.


It sounds like something with a 90% error rate would be better to just do the opposite of what was determined and suddenly have a 10% error rate??

That's not how type 1 & type 2 errors work. The article is describing type 2 (false negative) and inverting the output has no correlation with the percent of false positives (type 1)

https://en.wikipedia.org/wiki/Type_I_and_type_II_errors


I can easily build you a system that will guess the price of a stock tomorrow wrong at least 90% of the time. Even 100% wrong if you'd like. How would you use that to make fewer errors in some other system?

If it is a binary choice you can just do the opposite.

What’s binary choice we are discussing here and what’s the opposite that you’d do? Approve every claim that was denied by the AI, and vice versa? Or what’s the binary choice related to stock prices you’re referring to?

I agree, that most real world problems, are not binary choices. But in this case actually yes, just do the opposite.

For that reason error rates above 50% don't make any sense. Building a system, that is reliably wrong, is as hard as building one that is reliable correct, because the systems are the same, just with the output inverted.

The worst you can get with a system is 50%, at which point you can just flip a coin, because that is just pure random. Any deviation from that is going to be an improvement.


I always thought insurance claims were more complex, some things may be wholly or partially covered on a line by line basis, and there should be justification provided. In certain cases they go to mediation or even court. It doesn’t seem binary to me but maybe it works differently where you are.

I said "if it is", not that it would be.

> if you have a system that makes errors over 90% of the time then you have a system you can leverage to make fewer errors.

member the three words "delay, deny, depose"? Insurances - and not just in the US, Germany has similarly bad stuff going on - aren't making a profit if they just blindly accept claims. It is much more profitable to deny with a vague "AI" system (or a blanket automated deny) first, and only have a first look with an actual human agent at it when the customer complains or files a lawsuit.


this is generally true however if an adversary is able to control a source it becomes dangerous if they can preview the results or inspect the other sources.

Right, so your starting point is that the attacker has read-only access to ALL entropy sources, and in that scenario it's worse if the attacker has read-write access to one entropy source.

Yes. I don't find this a particularly interesting scenario, though. Sure, we can come up with stuxnet-like airgap attacks where we on-device, but not remotely, can read entropy sources. AND we can modify the output of RDRAND. And there keys have been generated for data we can later intercept. But despite that control (potentially on a CPU microcode level) we are unable to stegonographically leak it?

Sure. Possible. Has it ever happened?


Not to the kernel random pool, no.

this is a well known attack...

if your algorithm controls a source of entropy and can inspect the other sources, it can craft its source to bias the result. a fanciful attack but it means you should at least discriminate what you put into the pool.


Can you link to the paper you're thinking of? Maybe people are just talking past each other here. A biased random source can't bias the kernel random pool in any straightforward kind of way.

I explicitly said

    > it becomes dangerous if they can preview the results or inspect the other sources
because the malicious source can just precompute the hash for the bias it wants.

https://blog.cr.yp.to/20140205-entropy.html


You're assuming a hash preimage attack, which would be a complete break of the cryptosystem. (Your link only works on the toy implementation given.)

there is no preimage attack involved.

while you cannot take control over the hash output you can bias it because you have multiple tries. that's how bitcoin mining works too...

for cryptographic applications any bias can be engineered to be fatal in one way or another.


if you compare an embedding model to something like Jev which asks 100 questions and use the answers as the embedding you will be able to get move mileage out of the latter. especially because you don't need to train any classifiers for your task, you can work directly on the answers.

that said, I don't understand the hype. I have been doing what Jev does for 2 years now by just forcing json tokens onto an LLM. you can even get the LLM to think. and you can ensemble multiple LLMs.

I suppose the appeal of Jev is how cheap and fast it is, but then it's entirely unsuitable for anything but the most cursory extraction. using it to play games seems like a waste of time especially when most of those games will be played better by an algorithm written by an LLM (just give it the state and ask it to write a bot).


> I suppose the appeal of Jev is how cheap and fast it is

Ok, so you do understand the hype. I mean cost & speed are reaaaaally big issues for normal LLMs. If you can fill a specific use case way cheaper and way faster, that’s a fantastic development. We need more niche tooling that’s more efficient and better for niche use cases. Not everything has to be general purpose.

Not clear to me that playing games is the point of jev. This post is just a random fun experiment someone did because they wanted to.


But I can ask 100 causally masked questions against common prefix, and get 100 answers, all in a single PP pass using any existing "classical" attention transformer model? Like, I had the impression that is what everyone was doing for classification already? Is the difference "we did RL to tune logit distribution"? Because I really do not see anything new there. What is the difference?

If Jev can maintain its cost and speed with multi-modal input I think it will be well worth the hype.

    > Just keep growing and pay a fine if regulators ever find out.
it's unfortunate that they weren't able to get a lawyer with experience to comment on whether or not he is right.

would certainly be interesting to know if he does risk assessment under the table.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: