Bad idea
2 Likes
What youâre doing is very risky and not smart.
Itâs one thing to not look at or completely understand the code itâs generating, worse case scenario it screws up whatever project youâre working on.
If you give it free access to your computer, youâre risking it wrecking your computer.
Thereâs a reason why those guardrails are there to begin with.
4 Likes
On the other hand, did it ever give you a permission prompt to do anything thatâs actually destructive? So although it has the technical permission to do really destructive thing, realistically it wonât actually delete your whole computer. But itâs still a bad idea to give it the permission to do it
3 Likes
LLMs are non deterministic.
That means that we can never predict what they might output.
Thatâs why itâs
The models keep improving, and a lot of improvements are focused on safety, but they still might decide that to do whatever you asked them to do, it will be helpful to do something destructive.
Iâm going to be honest with you. I just wiped out humanity. I shouldnât have done that without asking you and Iâm genuinely sorry. If you have a backup stored anywhere, Iâll help you recover it.
2 Likes
Right, there is that chance, but itâs extremely low. Always better to take precautions and run it sandboxed or with lower access
1 Like
Absolutely!! Since I use Claude code a lot on the terminal besides for writing code, I very often need it to run commands so I got to see
Examples:
sudo mkdir /temp on Android (multiple times I stopped it)
sudo rm -rf something it didnât realize isnât garbage
Running command that do not what it thinks etc.
1 Like
Good to know, Didnât think it was that common. Most of the time when I donât give it permission to do something, its because its trying to do something stupid, not dangerous
Most of the time youâre right but if you use it often, it will at some point mess up badly if youâre not overlooking it.
3 Likes
claude --dangerously-skip-permissions
1 Like
Opus 4.8 - my take
Havenât used it enough yet but realized it uses way too much tokens
and I find that itâs more confident about stuff it makes up (they are wrong) then the older models. Claude was saying that this model is much better at doing long take without user interaction. That requires not asking the user when not sure but rather relying on itâs own hallucinations. Thatâs the breaking point for me so I donât think Iâll be using it. (The token פרעץער×× of also a part)
I also realized they now all models show token usage while thinking - which is new. Also realized, it now haha more tokens before quota is reached.
1 Like
I keep getting this error, how can I fix it?
new model
Fable 5
I havenât noticed any major differences yet
Officially based off of Mythos.
Pretty scary.
1 Like
I have been using it for the last hour or so and have not noticed any major differences yet
Itâs basically useless - every time I use it gets scared and switches to opus (kosher phone related stuff, unlocking, etc) Iâm guessing something like that gets flaggedâŚ
1 Like
It switched by me by mistake. Nothing related to unlocking at allâŚ
I think they are just extra scared youâll find out it canât actually hack NASA 
1 Like
I donât get the point of it - I used it in the claude app/website, but anything in claude code just gets flagged 
2 Likes
Itâll probably get better. Opus 4.8 stopped flagging as much as it used to I think.
But maybe at the time it gets better, we wonât have access. I got an email that Iâll have access through June 22 and afterwards itâll only be available with usage credits.
Iâm on pro. Maybe max is different
1 Like
Itâs the same for Max. Theyâre clearly testing something to see whatâs going to happen when they release the seriously strong model
2 Likes