> This is why it's impossible to create a digital assistant, or really anything useful, via Markov Chain. The fact that they only generate sequences that existed in the source mean that it will never come up with anything creative.
Or, in other words, a Markov Chain won't hallucinate.
Having a system that only repeats sentences from it's source material and doesn't create anything new on its own is quite useful on some scenarios.
> Or, in other words, a Markov Chain won't hallucinate.
It very much can. Remember, the context windows used for Markov Chains are usually very short, usually in the single digit numbers of words. If you use a context length of 5, then when asking it what the next word should be, it has no idea what the words were before the current context of 5 words. This results in incoherence, which can certainly mean hallucinations.
A Markov chain certainly will not hallucinate, because we define hallucinations as garbage within otherwise correct output. A Markov chain doesn't have enough correct output to consider the mistakes "hallucinations", but in a sense that nothing is a hallucination when everything is one.
You can very easily inject wrong information into the state transition function. And machine learning can and regularly does do so. That is not a difference between an LLM and a markov chain.
I've never looked at the source code of a PT, but my understanding is that they also implement heuristics to detect strange behaviors, i.e. clients trying to fake their ratio.
Not all of them do, but yes. Since it is the client who send all the info (bytes sent/received, version, etc), anything can be spoofed.
The tracker, having access to the stats of all the client on a specific torrent can find "unusual" client (for example, a client that report a lot of bytes sent while no other client report downloading/receiving anything, or the opposite). But due do quirks of the networks, delay and possible lost stats messages, they still have to be pretty permissive of small discrepancies. Also, all use an ID in the tracker url for each client to discriminate them and associate them with their account on the website. And some are not very careful about the ID generation algorithm, leading to being easily able to find other account peer ID, and so you can use them to download to preserve your ratio. I even found at least one private tracker that did not care if you passed a non-existent peer ID.
For example there's also something called "ghost leeching" (side channel entirely bypassing tracker reporting) which can lead to other peers reporting upload for which there's no opposite account of download on the tracker. Making it look like peers over-reported upload and cheated when they are in fact entirely innocent. There's no way for a private tracker to be really sure about stats. The most the moderators can do is to check for repeating suspicious usage patterns across many torrents of a particular peer under scrutiny.
I have used patched version of muTorrent. Binary exe contains string private, and I changed it to crivate with hex editor, so that torrent client stops reacting to private flag and starts whistleblowing peers to DHT. I don't think that my client looks unusual.
It does. If the tracker is monitoring DHT. Or, more likely, if your client finds a similar "no respect for private flags" DHT peer, sends data to it, and reports that to the tracker.
Does that weird DHT peer report to the tracker? It may not even have an account there. From the tracker's end, that peer is a ghost, and it looks like your client has reported sending data no one ever received.
> because clearly keeping us alive after the age of 40 really wasn't _that_ necessary for human survival.
That's a very common misconception. Being alive after 40 is quite necessary if you are a member of a gregarious species that (bar exceptions) always lives in community. And it's not only about the survival of your own genes, it's about the survival of the genes of the whole community.
Babies, just like adults, are extremely social animals. And they absolutely need to interact with a bunch of other people their age, even more than us. Daily, and for a long period of time. An hour at the park doesn't cut it, and being all day with a sibling doesn't either.
So beyond everyone going back to a Neolithic way of life and living in a bunch of straw teepees all bundled close together, daycare is the best solution I've found to this need.
Just as an example, my oldest has been besties with another kid since they were both 7 months old.
As a counter-example, neither of my kids really acknowledged other kids in any way at that age (and other infants ignored them right back). A quick Internet search suggests it's normal for them to not interact with other kids until after 12+ months. This was a point of contention with my wife and MIL because my MIL would complain we weren't "socializing" our oldest enough when she was an infant despite clearly never having looked up anything about childhood development.
That and we did take her out all the time. She just wasn't in daycare. The thing about stay-at-home parents is they don't literally stay at home all day.
Or, in other words, a Markov Chain won't hallucinate. Having a system that only repeats sentences from it's source material and doesn't create anything new on its own is quite useful on some scenarios.
reply