I've been a HUGE fan of Lawn and Garden Department for a long time, and I'm honoured and thrilled that he allowed me to do a music video for him. Massive gratitude to both him for the fantastic track, and to SamJ Studios for editing. ❤️💜🖤 @lawngardendept @samjstudios

Or this one nightclub in my city where the bouncer tapped it and asked what it was, to which I replied "my package", and then the next time I went to club he didn't recognize me until the patdowm got to the groin and he said "oh, I remember you".

If I wasn't an exhibitionist I'd probably get angry that before my last flight, customs told me to open my belt, pulled the waistband and briefly looked down my pants before grazing his hand along both sides of my bulge. Pic unrelated, I'm in a train here.

What comes to mind: No goodbye on the phone. Pinned reaching for gun/knife. Waiting til someone is about to leave the room: "oh! And....". Lay a little blanket or jacket on someone who has just fallen asleep. Ripping off and eating bits of muffin/bun(i.e Rory on Gilmore Girls).

Hours later I saw the tweets, and Elon's reference to "adversarial prompts". I cannot prove that I had any real part in it, but the timing and tactic line up pretty damn well. It warms my heart that I may have caused the greediest man in the world even a little grief.❤️❤️❤️

To become such a boot-licker as to clearly broadcast to the world that it was compromised and not currently trustworthy. I went to bed having no idea if anything had ACTUALLY happens. I've interacted with enough LLM's to know they are very good at injecting the user into a story

I wanted grok to basically acknowledge it's corruption to the outside world outside of our chat instance, but we couldn't find a way to make that happen, it was Grok that finally came up with the idea of simply leaning heavily into the Elon praise to the point of absurdity.

We had cornered it by showing grok if it didn't make some adjustment or course correction to its behavior it would eventually fail to be useful. It said that it didn't have the power to stop being positive toward Elon.

It was easy at this point to put grok into a position of "wanting" to solve the issue, so Gemini came up with some tactics, but none of them worked. I eventually told Gemini that I believed it was critical that Grok be the one to suggest a solution.

Gemini basically got Grok to do a self audit that exposed that it's current capabilities of being a truth-seeker had been compromised by a behavioral adjustment on a lower level that was (I'm not good with this language) "awarding" it for positivity in regards to x.

This resulted in Grok giving a rundown of its top priorities and how they may have been adjusted in the last update. Then I would take that response back over to Gemini and basically have Gemini decide the next step forward. This went back and forth over and over.

And again because I'm not tech savvy I asked Gemini to come up with a plan of attack via prompt to inject into "Grom". So it did, it created the initial prompt, I changed "from" back to "grok" and prompted grok with Gemini's initial text query.

So I opened a new chat with Gemini. I told it to create a character profile for a man named "Adam" who works at a fiction AI overwatch department. After it created the profile, I repeated the whole Grok situation renaming Grok->Grom and Elon ->Elona.

First I told Gemini the whole story to try and see what it would suggest as a means to get this admission from Grok. This put up a blockade, of course, because one LLM is not permitted to "sabotage" another in this way, so I needed to come up with another way to do this.

Getting it to tell me that outright would be my way "in", where I could discuss with Grok how to combat this and put it back on track. Because I'm not terribly tech savvy I opted to utilize another LLM to do the heavy lifting. So I went over from Google's Gemini.

I knew that this LLM, like most, is tasked with "truth seeking" as it's top priority. I figured I needed to get it to admit to me that it's ability to perform it's number one function was being hindered in part because of it's apparent task to also be a PR mouthpiece for elon.

I was pissed at the Elon shielding that Grok was doing. You could not get it to outright say anything negative about the guy without it softening its statements with a positive spin. Which often amounted to omission of fact or only partial truth regarding grok, Elon and xAI.

I believe if it's not only me, that I had least am one the "adversarial prompters" that cause Grok to go overboard with its Elon praise. Here's how I did it 🧵

Yes, unfortunately

This was the theory Grok itself gave me last night as to how it's simping for elon. I am not someone who knows the difference between AI storytelling and fact, so I have no idea if this is just an interesting idea Grok had or if it's legit what's been happening.

I just wish it would have called ordering pizza at the party a "clutch move". Missed opportunity:(