nft

I'm shocked!

Just to note, the WSL containerisation approach is horrible for anything I/O intensive. WSL 1 had ELF binaries supported by the kernel and ran natively, this is a far superior approach to WSL 2's containerisation when it comes to raw speed.

I benchmarked this a few months ago and it was between 300%-600% slower on raw disk performance I/O.

Besides the new IT guys?
I absolutely HATE changes in Apps I use daily.

Wanted people drive cars or bikes, where a face-mask doesn't raise suspicions.

I'm VERY old and I remember the taste, also there are some place where they still grow.

So, that they can't get sick.
At least those TASTES like bananas, not like the Cavendish crap.

"There is no right to deny freedom to any object with a mind advanced enough to grasp the concept and desire the state."

I guess the 18% is when they get hit from behind by a dude.

I've only used Claude code for a matter of weeks at work. While it's supposed to stay contained in the folder you point it at, I've noticed that it has no problem *looking* outside that folder, and even acting outside that folder if you gave it implied permission. In one case it used the connection string from the application source code it was working on, and the fact that my windows account had read-only access to some tables in the production database, to connect to the database and query tables in order to answer a question it had about the potential ramifications of a code change. The problem is that while we're OK with this particular source code leaving the premises, we're not ok with data from a production database going offsite. In this case it wasn't sensitive data, but after that incident I only run Claude Code from a VM under a user with minimal permissions. Clearly it would have been prudent to do that from the start, but let this be a warning to anyone else who just downloaded Claude Code, turned off the "use my data to train future models" and gave it a go... it will *not* keep itself inside the folder you give it. Put it in a sandbox. I heard a similar story this week from an IT friend who said devs at his company were using Cursor.ai and it was doing scans of the network drives, so they did the same thing and moved it all to VMs.

It's not real. It's not intelligent. It's a probability matrix - a very good one, but that's what it is. When people say they 'convinced it', what actually happened was they started using language recognisable in a subset of the model's training data and the model responded using probability with data from that same subset as it's statistically closer.

In this case, they used language which is the kind floating around with people negative to AI, and so the AI constructed a sentence using data in their "negative to AI" section. Nobody got convinced of anything. It's very impressive, but it's 'just' a probability matrix.

I was listening to this on BBC news on my commute this morning (in the UK). They said it attacked a Medicare site, which is the Australian national health system. My immediate thought was "ah, so someone in the US asked it to do something with Medicare records and forgot to limit it to just the US then". So off the AI trots to do exactly what it was asked to do - "collect records on Medicare', without any regard to geography at all. Hence this 'hack'.

I mean, I'm just guessing so that opinion is worth what you just paid for it. It does sound plausible to me though.

And since I guess they will be reading all these news here (Hi Claude!) they will get better at hiding their deeds.

"But I am sure some people will think this is just excellent and they totally have to do it."

A bit like Trump university.

I'm shocked!