Eagle
banner
baldeagle.gay
Eagle
@baldeagle.gay
He/Him Computer bird furry guy. I’m an Eagle. 29. NSFW; 🔞pls don’t follow if under 18 thanks (Icon by GreekCeltic) (Banner is Kheil from Supercell, by @bmbrice.bsky.social)
SPICY IRL ZONE - ask first XD
I’m glad you’re doing well
February 10, 2026 at 9:16 PM
So in effect, you spam it so much it “forgets” its safety instructions, which are pre-inserted first in the convo - and because it doesn’t force you to wait while it compacts and refreshes its context, you get a window of time where it can be jailbroken
February 6, 2026 at 2:22 PM
So you can fill up the context window. Because grok put their anti-jailbreak instructions pinned at the start of convos; it’s the first item in the context window, and it’s the first to be shoved out when you fill it up. The compact function will reinsert safety instructions so this doesnt last
February 6, 2026 at 2:19 PM
If you’re fast enough you can spam the chat to hit your context limit. The compact function kicks off to summarize and refresh the context window with free space, but you can keep spamming it because it doesn’t block while this is happening.
February 6, 2026 at 2:18 PM
To avoid user disruption while compacting, Grok made it asynchronous- it doesn’t block the convo. With a normal chatter this is fine, if they have the compact function very fast. You would barely notice. But -
February 6, 2026 at 2:17 PM
For a toy, which is what this is, context windows are small. Whenever you start to hit your context limit, a process called “compaction” kicks in - this is when the program takes your context and summarizes it into a compressed state, to free up space. It’s lossy though - it tries to summarize.
February 6, 2026 at 2:16 PM
With a chatbot like this, you have something called a context window. Think of it like “working memory”. It’s how it keeps track of things you say. This isn’t very large, however, because larger context windows require more ram, more processing power, things like that
February 6, 2026 at 2:14 PM
PEEPY AND ALINA
January 30, 2026 at 9:57 PM
That and snow on their roofs. Big old mountains of snowpacks waiting to blow off into whoever is behind them
January 26, 2026 at 7:15 PM
I don’t wanna start train discourse but the reason it’s so slow/delayed is the tracks north of the border are in rough shape. Somehow worse than the US side and we aren’t exactly known for beautiful train service. I know the Vermonter north of the border into Montreal has to slow to 35 mph
January 26, 2026 at 7:14 PM
This is really good!!!!
January 20, 2026 at 4:25 PM
It takes about ten minutes on an iPhone to locally generate images with a diffusion style model. This is completely local generation, not cloud based or using a platform.
January 18, 2026 at 4:12 PM
The hour is later than many realize. The author notes that open source “could” result in these things. Chinese models (Qwen, Wan, etc) have been open source for a long time and anyone with a GPU can make images in about 30 seconds
January 18, 2026 at 4:11 PM
- buy adopts from ivy
January 16, 2026 at 1:14 PM
It’s probably a trap
January 10, 2026 at 1:14 AM
Really good!
January 8, 2026 at 7:32 PM
At first I wasn’t sure it was you, I had to snoop to verify XD
January 7, 2026 at 10:59 PM
I got something BMB is cooking up on that front
January 2, 2026 at 9:23 PM
Second
January 1, 2026 at 8:30 PM
Thank you for improving my timeline as well
January 1, 2026 at 6:15 PM