In layman's terms, this seems to mean that given a certain unedited LLM output, plus complete information about the LLM, they can determine what prompt was used to create the output. Except that in practice this works almost never. Am I understanding correctly?
No, it's about the distribution being injective, not a single sampled response. So you need a lot of outputs of the same prompt, and know the LLM, and then you should in theory be able to reconstruct the original prompt.
Ad-hoc blocking of bad actors is bound to be an endless futile game of wack a mole. The way I see things going, the internet is continuing to move away from an open web and into walled gardens. Those with resources will create large walled gardens like the gardens of Meta, OpenAI and Alphabet, each with their own issues and serving the interests of their owners. Smaller walled gardens will exist, but any time they grow anywhere near the scale of the global web of old, they'll face increasing challenges from bad actors anywhere from spam to scams to ai to propaganda and only those with resources will be able to maintain those walled gardens, and they'll only spend their resources on that if it suits their interests.
All IPs are allocated to CIDR blocks and Autonomous Systems, the latter identified by their Autonomous System Number (ASN). It's reasonably straightforward and tractable to track good/bad behaviour by either, and (thanks to the Law of Large Numbers and Power Laws), there's virtually always a very small number of absolutely horribly-misbehaved blocks from which a large fraction of abuse originates. Moreover, at sufficiently fine detail, it's possible to identify both friendly and hostile address spaces, permitting carve-outs for the former and scaled response against the latter.
The second part of this approach is that defences need not be all-or-nothing, universal, and/or unscaled. A netblock with a few bad actors might be subject to a slight performance penalty. A netblock with no non-hostile traffic could be blocked entirely (or tarpitted or otherwise subject to negative performance impacts). And of course, reputation data can be shared, as a broader view (one which, say, a large CDN or monitoring service might have) is going to provide both earlier warning and greater detail of where hostile activity originates. And individual instances of good behaviour could be excepted from broader blocks.
Ultimately, connectivity providers, whether of data centres or residential / organisational / mobile Internet services, should be encouraged to police their own outbound traffic and take actions themselves in the event of identified abusive behaviour. (That's been a long-standing dream of mine, it's ... stubbornly refused realisation.)
I read The Body Keeps the Score about 10 years ago after experiencing a particularly traumatic event. I had been searching for answers at the time for how to heal from the event, and someone recommended the book to me.
I had distilled my memory of the book into the intuitive idea represented by the title that the body remembers what happens to it which of course there is some truth to. So my first reaction to the headline was a bit defensive, "of course it's not bullshit!"
But I had forgotten how much emphasis was put on there being significant lasting changes from events that we couldn't even remember.
> The idea that trauma causes long-lasting damage to the brain and or body is central to van der Kolk’s thesis.
> his narrative paints this hopeless picture of trauma victims as being people who most aspects of their lives are “dictated by the imprint of the past.”
And now I remember that when I read the book originally looking for answers to my own traumas, it left me feeling hopeless, overwhelmed and permanently damaged from what I had gone through. I remember thinking that the high stress I was going through at the time was going to leave me permanently struggling with issues like high cortisol and inability to function normally. Absolutely NOT the message I needed at the time.
Ten years on, my brain is normal and healthy, and I don't have any perceptible problems with cortisol or PTSD-like symptoms. I'm living a healthy, normal life, and not walking around with a heavy trauma score in my body.
That's not to say that trauma from my past didn't play a role in making me who I am today, or that I don't still carry some memories of the difficult events. But I've found my brain to be very plastic and to heal and rewrite itself quite well. And all of the measurable markers of brain health, stress hormones, etc are fully back to normal, healthy ranges.
That's just my anecdote, and I also appreciated the thorough scientific analysis in this article!
> You take 100 such cases, ask AI "how does it work?" and in 99 of those, the answer is somewhere on the spectrum between "total nonsense" and "clever formulation but wrong". One turns out to be right.
They're still using the scientific method, the only thing they're getting from AI is hypotheses to test. And AI is great at brainstorming plausible hypotheses.
Based on the marketing page and App Store page, I can't really tell what sets this apart from ChatGPT. It looks like essentially the same thing, with a slightly different UI. What features does it have that add value over ChatGPT?
That would be a great idea if I were on vacation in a cabin in the woods. But realistically, I need my phone for just about everything I do on a daily basis, from payments, to navigation, to communicating with friends and family, and logging into accounts for work.
At least a few of these, like payments and basic communications, can be done from a watch.
Work accounts, camera, and maps are the big blockers for me. I know I can buy a camera but 90% of the times when I take a photo it's to instantly send it via a messaging app, mostly for work.
reply