The random number generator didn't produce the number I wanted? Shocking!
In all seriousness, uneducated people should not be setting up LLMs (or, frankly, even using them depending on the circumstance). Understand what the technology is. It isn't intelligence, it isn't a person, it's math. Vector encoded data.
Once you recognize that all outputs are possible, including all possible negative outputs, you will be much better at determining proper sandboxing/boundaries.
If you run an LLM and give it two tool calls: read_file and curl, you can have all your data exfiltrated (and, depending on how bad your setup is, it might only need curl...)
Certainly not true in every culture. Because this is not an intrinsic truth of humanity. The evils the author sees as exceptions are actually the rule of our nature, but for one reason or another, most people (in some cultures) these days choose to operate on better principles because the outcome is preferable.
A murderer might enjoy killing, but will only be able to do so for a short time before being caught and imprisoned (or better, executed). But few people have ever been jailed or hurt for exercising charity and kindness on others. It's not that humans are normally or mostly good (we're very clearly not), but that living as such tends towards better and longer lives.
So in that sense I also believe, for the most part, that those around me are more willing to help than to hurt. Sadly those with the most influence on our lives don't represent a random sampling of "most people".
A probabilistic technology (yes I know you can set the temperature to 1, but no one does that, and the output is still not efficiently predictable) is doing "formal verification" now?
Or are they just saying we can use LLMs to generate a piece of code/test that can do formal verification?
a beautiful (static?) website for showcasing a beautiful cipher
excellent work on both fronts (solving, and showcasing)
while he failed to demonstrate that the spirits of men continue to exercise control over this realm (after all, is it not Satan who presently rules this world?), he left something beautiful and interesting that genuinely took a while to solve considering how close it was to the computing revolution
not that bold, but I suspect a discussion on the topic would either be unfruitful - or if it was a good conversation, would be shutdown by the site operators
Hah, I was trying to find this comment you made. There's confusingly also "Calif" in the mix (possibly closer to your string of people, but not your affiliation either). I wonder if, perhaps, it was easier to point to your HN post for bleepingcomputer, than what Chainguard did (you only expressed something was coming). It's all slightly immaterial since exploit attempts started 3-4 days.
No connection or affiliation to any of the organisations involved. I am a self-employed consultant for my day job, and a solo founder of a mental health non-profit on the side.
But this is more like not including tops with your to go drinks when every one throws them away. If Android never used MTE, removing doesn’t harm the Android experience which is what the Pixel is intended to run.
The first reason the author lists for why this will be bad is only an "issue" on Git hosts that don't allow repo creation on push (which is brain dead of GitHub). Any other host, you push your new repo, and it will see the hashing algorithm, and receive the contents accordingly.
Submodules is a legitimate argument against this, though I don't know how widely this feature is actually used, and similar to the arguments in favor of switching the default branch from master to main, this is simply a setting which can be changed.
I do like the idea of commits having both hashes, and am surprised that idea has not been explored further.
Generally though, I think the author's strongest argument is simply that the change isn't strictly "needed", and all the other issues presented aren't the strongest arguments against change.
In all seriousness, uneducated people should not be setting up LLMs (or, frankly, even using them depending on the circumstance). Understand what the technology is. It isn't intelligence, it isn't a person, it's math. Vector encoded data.
Once you recognize that all outputs are possible, including all possible negative outputs, you will be much better at determining proper sandboxing/boundaries.
If you run an LLM and give it two tool calls: read_file and curl, you can have all your data exfiltrated (and, depending on how bad your setup is, it might only need curl...)
reply