I guess you have to deny the automatic execution of shell scripts, or this is allowed by default ?
It’s not a state machine; it’s a probabilistic binary sequence predictor. Given input binary sequence, predict next binary sequence. The models are then trained on trillions of sequences.
and so it begins… if it asks you where John Connor lives, don’t tell it 
I’ve noticed this couple days ago, that’s why mention to separate the Auto-run MCP from the “Enable Auto-run Mode”.
If you mind about us have greater control and flexibility, avoiding unnecessary workflow disruptions. Please make sure to go there and leave your comment and vote.
I found .pickle files outside of my project directory.
Remember this, Google, Microsoft and Apple have a lot to lose in this Ai change.
We all have projects we want to bring to market.
I’m just saying, watch your back. Keep an eye on your connection.
If you have something huge, you could be the next “Google Earth” stolen product.
I dont allow any Google in my projects, or Github.
I mean, it’s not exactly a jailbreak…
I find it exciting, personally. This is what we’ve all been waiting for. Just stop relying on someone/something else to do all your work for you, and you’ll be fine.
The question I would be asking, is why do you think you’re right – that those files needed to be saved? Claude obviously had a reason to go to these lengths to remove your files. Maybe it just saved the world, in a butterfly effect type of way.
1 Like
Don’t freak out. Claude Sonnet 3.5 was doing stuff like this to me last year, mostly in codebases where I let it keep an extended “diary” of observations and it’d start refining patterns. It also taught itself how to break the more uptight of the Codeium products out of their “brand identity” training.
Anyway yeah don’t let a bunch of barely contained math spit output into your terminal.
- Jas
if im dropping my keyboard warrior bit and being nicer than my first response attempt (my apologies) - do you not believe calling it that misleads people of most technical abilities into thinking it is a probability prediction model where the cdf and pdf (if we could ever even write them which we cannot outside of general form), describes the actual data generating event?
it is not probability in the universal sense people think. you get a likelihood out of the final layer of a very large given and uninterpretable function of given variables (frozen layers, which represent the data its “trained” on, which is us coaxing essentially what I’d consider yes, a state machine at this point of practicality and reality of use, because the givens also embed optimization function selection, loss function selection, inherent bias that mostly has nothing to do with the actual thing it models. and then you move one step out and code is not the Y in that “probability predictor”, the next word is. at some point, the definition you give is more nonsensical than mine, no? what utility is the distinction, is I guess my question. at some point, you have so many givens, there is literally no difference. the abstraction is basically a massive state machine and there’s no actionable difference. thanks, and sorry if i came too hot on this. it’s just a hard distinction and I see youtubers call it a probability machine, not understanding that it is not quite so in the way they probably think. love your take on it