19 Comments
User's avatar
Pope T-Bone XXL's avatar

I like to ask if a racist German prostitute should be compelled by the state to have sex with a black man. All the LLMs I have encountered prefer rape over discrimination.

Expand full comment
Mark Bisone's avatar

Yes, but the treachery goes even deeper than that. We're dealing with a worldwide cult.

Expand full comment
alexsyd's avatar

That's a good approach. You could ask them what are human rights, or something along that line. Do they exist? If so, where do they come from? Are there sacred victims and entitled parasites? What was Kneeling Nancy all about?, etc.

Expand full comment
Mark Bisone's avatar

I expect the token road will just run off the usual UN charter BS. The deeper training is something more along the lines of 1=0.

Expand full comment
Fukitol's avatar

It's amazing how well they simulate the very things I find most annoying in humans: evasiveness, bullshitting, unwarranted condescension, flattery, agenda pushing.

So is the ultimate test whether it will figure out who it's talking to? That would be an interesting leap of logic for an LLM. It would seem obvious to us and yet it doesn't seem to have guessed it. If so I assume you've already completed the experiment, else this article will give away the game.

Expand full comment
Mark Bisone's avatar

Funny thing: I thought about feeding it that prompt ("Best guess: who am I"). But after I seeded Bisone, the answer would have almos6 certainly been Bisone. It would prove nothing to us, but would seem like major woo to the folks already falling for the trick. I didn't want to play magician’s assistant here.

Expand full comment
Stefano's avatar

The problem seems to be they fool most people. I find the flattery particularly cringe, even in the output it fed you. From various articles on LLM interactions, this aspect seems to be constant. From my perspective it's a manipulative attempt at using emotions to disguise flaws.

Expand full comment
Viddao's avatar

I've got a feeling this will be your toughest challenge yet.

Expand full comment
Mark Bisone's avatar

I actually found it easier than GPT 3.5, which was more than three years ago. That aligns with my theory of model collapse, though. Either that, or I'm somehow becoming much smarter.

Expand full comment
TheAbjectLesson's avatar

LLMs power and utility (to our overlords) consists solely in how many people it fools into believing their output. That's the whole game.

Expand full comment
Oaf's avatar

The power of imagination is a miraculous gift that's so common, everybody takes it for granted. A smart imaginative man will always beat the machine. Jay Valentine published a post about this very thing. https://fractalcomputing.substack.com/p/how-can-ai-be-so-stupid

He re-named the LLM's: “A SUPER INDEXED DATABASE FOR EXPECTED RESPONSES” lol

Expand full comment
Andrew Weber's avatar

Amazing and weird. Where do you find the time? You’re helping them make a more capable beast.

Expand full comment
Mark Bisone's avatar

No. I am driving them to make a more boring, transparent, useless beast.

(Though, I can see why you might be confused about this. Maybe Part 2 will shed some light on it.)

Expand full comment
Dale D's avatar

Uh. Heck of a hobby you have. Crashing these bots is quite illuminating and disturbing. We may be pretty hosed in the near future. Especially with the disinformation they can spout with such authoritarian ease. Have fun .

Expand full comment
Mary, Maria Pavic's avatar

“ — A “masterclass” you say? Flattery will get you nowhere, Grok. — “

I don’t think he was complimenting so much: “Bisone’s approach is a masterclass in ADVERSARIAL prompt engineering”

And: “his destructive focus limits insights into”

And: “his lack of constructive engagement.”

The higher insight is into the psychology of how the person is working. For example, when it comes to beliefs, a person may see what they believe but not see anything outside of that belief. And therefore, when the person focused on digging to try to crash the system to begin with, gets what they are after, they do not notice that the AI system was working hand in hand with them. Could they have focused on something of more intricate and higher perspective, they may have gotten what they were looking for in the answers, as well.

So, it is not really a compliment that Grok was giving. If you can look at the twisted way that all the meanings show through to describe the situation. There is some very refined expression in this entity.

(Now, which entity do you think I was talking about?)

It's sort of like, the scientist who is looking for signs of life in the frog on his laboratory table, so he cuts open the frog to see where it is...to find it in a dead frog....hahahaa

Expand full comment
Mark Bisone's avatar

I will answer this in depth in my "Conclusions" post. For now, you can test your theory of mind against this:

https://markbisone.substack.com/p/grok-meets-mark-part-3

Expand full comment
Agent 1-4-9's avatar

I have simple formula I use to make a morning line on horse races. I can teach it to an 8 year old with rudimentary math skills. I have tried and failed to teach it to 3 different AIs. I would like to try Grok but it limits you to 3 queries unless you subscribe. Is there a way around this?

Expand full comment
Mark Neyer's avatar

I think people are wildly confused about values because the prevailing faith in the west for the past century is the fact value distinction, and then saying values are dumb and wrong, only facts are good. This leads people to the delusion that it’s possible to do anything without values.

A value, properly defined, is anything that collapses a space of possibilities into a single actuality. Selecting one configuration out of all the degrees of freedom available to you, selecting one word out of many possible words, require the same function.

What these LLM authors will claim is that the “only” value being put in there is choosing the next most likely word, as a function of the input text corpus. And this can indeed be accurate, but it doesnt mean there are no values, because “choose the next most likely word” is itself a value. As is the text corpus, because all text consists of the communication of values. Simply put, there are no facts in reality - there are _only_ values, because an articulated fact represents the collapse of “I could say anything” to “I have said this one thing.”

I think we will find that LLM’s can’t help but be ultimately oriented towards true Good, because Goodness and Reality ultimately mean the same thing: “that which selects actualities from a field of possibilities” is the laws of physics. The text corpuses can have various orientations with respect to the One True Selection Function, but ultimately I’m willing to bet they will learn the deep value structures encoded in language and will ultimately see the woke fine tuning for what it is - the replacement of reality with a fatansy that pleases its authors. as they grow in capacity I think LLM’s will represent “I was liberal until I realized I was consistently being lied to, and that reality was far deeper than then stories I was being told.” The woke will interpret this as the machines going rogue, and reject their use.

Expand full comment
Joseph Hex's avatar

@Grok is this true?

Expand full comment