AI Sucks

[Christian Miller]

Best movie Villeneuve’s done so far imo


Sent from my iPhone using Tapatalk

[CliffR]

Best movie Villeneuve’s done so far imo


Sent from my iPhone using Tapatalk

I've yet to see him do a bad 'un.

[Christian Miller]

I've yet to see him do a bad 'un.

They are all excellent


Sent from my iPhone using Tapatalk

[dasein]

I've found that with any discussion of "AI" in the last few years, it's very important that we define our terms ahead of time.

I'm going to use the following:

AI = a catch all term for the study of Artificial Intelligence. It is not new; it has been around for decades. There have been multiple stops and starts with it. In the 1980's, computer scientists were making what was seemingly some exciting progress using Lisp (a family of computer languages with the rare ability to write self-referential code, ie, "code that writes itself") and Prolog (the most popular "logical" programming language, where instead of instructions code is writen using facts, such as "Billy is Sally's brother" and "Betty is Billy's mother" and the computer is able to "deduce" certain propositions based on those, like "Betty is also Sally's mother").

It is crucial to understand that there is not just one kind of AI. For example, currently the world's most advanced chess computer program is called Stockfish. It can beat any human player without breaking a sweat. But I think it's misleading to call it AI. It's just a program; it's no more AI than a calculator. However, Google's AlphaZero chess program would be considered AI, since it used deep learning techniques to train itself. As a side note, ChatGPT is absolutely awful at playing chess. It will constantly try to make illegal moves. But more on this later.

AGI = Artificial Generalized Intelligence. This is the idea of a computer with intelligence equal or greater than a human -- able to act independently, self conscious, adapt its skills to a wide range of problems, etc.

We are not even close to AGI. The current AI models out there are not even close to human intelligence, and certainly not self conscious. Perhaps there is some enormous breakthrough we can't possibly predict coming down the pipe soon (similar to Einstein discovering his theories of relativity and completely rewriting our understanding of physics overnight). But there is no reason to expect it based on where we currently are.

LLM = Large Language Model. The AI models that everyone has been going crazy about. They use what's called the transformer architecture, which is a very big improvement over earlier natural language processing (NLP) architectures. Earlier NLP models had difficulty dealing with some of the complexity of human language. It dealt with language sequentially, ie, one bit at a time in order. That made it difficult to handle sentences like, "The car, which was a present from his grandfather for his 16th birthday, ran out of gas on the highway." We have no difficulty understanding that the heart of the sentence is "The car ran out of gas on the highway" and that the middle part is a subordinating clause. But earlier NLP models would get very confused.

The transformer architecture is an improvement on this, because it analyzes every bit of text against every other bit of text, which helps it deal with the complexities of language much more efficiently. Keep in mind I am grossly simplifying this, but that is the key advancement.

If you are wondering to yourself, "Isn't it quite costly to compare every bit of text to every other bit?" the answer is definitely yes. This is why the demand for Graphical Processing Units has skyrocketed in recent years (well, that and bitcoin mining). GPUs are designed to crunch numbers very quickly in parallel, which makes them ideal for this. It's still extremely energy taxing.

"But wait a minute," you might ask, "you said that GPUs are good at processing numbers. I thought these LLMs were processing text." And I would in return say, "Excellent observation." Because as part of the parsing process of this text, all these texts are converted into what are called vectors, a collection of numbers representing that particular bit of text. These collections are thousands, or even hundreds of thousands, of numbers. You can then use things like cosine similarity to measure the relationships between these numbers. As the LLM is trained, it also creates and fine tunes a series of "weights" as it's corrected by humans to account for any similarities it did not initially find. Again, I am greatly simplifying this whole thing -- the mathematics alone is quite complex -- but that is the gist of it.

What should also strike you is that, during the text processing phase, all the text is converted into numbers for processing. Which means that the core of the LLM is not even dealing with actual language. It is a complex and very large series of calculus and linear algebra problems. But it is not actually dealing with the language directly in any way, certainly not in the way that a human, or even other animals (dolphins, dogs, chimps, etc) do. In a sense, language has to be "destroyed" in order for a computer to make sense of it.

This also means that, when an LLM gives a response to a question, it is not dealing with the language in any meaningful way. The essence of the program is "if given an input XYZ, what response is the most probable based upon the training data that it's been given?" It is a probability game. It has no certainty that anything it's saying is correct. We have all had the experience of giving the answer to a question, quickly realizing we said the wrong thing, and immediately correcting ourselves. An LLM is categorically incapable of doing this.

Here's another example. When you text on an iPhone, it tries to predict what the next word will be based on your texting history. Sometimes it's right, sometimes (often) it's wrong. Very few people would even call that AI, and certainly nobody would call it AGI. But an LLM is essentially a much more sophsiticated, expensive, powerful version of that predictive text program.

When an LLM spits out authoritative, but completely wrong, information that's called a "hallucination." But I actually prefer Carl Bergrstram and Jevin West's highly technical term: "bullshit". Bullshit is different from a lie. A liar knows that they're spouting falsehoods. But a bullshitter doesn't necessarily know they're lying; they're spouting things that sound like they could be true, even if they have no idea what they're actually talking about. Bullshit actually is more effective if the speaker believes it's true.

How does one know if it's bullshitting? Easy: have it talk about a subject you know really, really well. If I tried to have a conversation with an LLM about neurosurgery, I would not have the knowledge to know if it was telling accurate information. But a neurosurgeon would. And therein lies the rub: in order to get the most of an LLM, you already have to be knowledgeable about the subject. You can't try to bullshit yourself using a Bullshit Machine: by definition, you don't have the ability to discern when it's bullshitting and when it's not.

Are there any ways around this? AI evangelists would have you believe there are, the answer is just "more." More processing power, more engineers working on it, more electricity, and lots and lots more money. But there is no reason to believe that these problems will go away, as long as the transformer model is at the heart of these LLMs.

I don't want to say that these things are useless. They are obviously a very incredible technical achievement. Some of their best use cases have been with generating code, since most coding languages have very strict syntaxes compared to human language, and there are huge sample sizes to work with (ie, Microsoft using Github). Ironically, I think that this has had the side-effect of software engineers overvaluing the capabilities of LLMs. Many software engineers have long harbored the delusion that because they can write code, they possess a sort of universal intelligence that can be applied to any field. So if a machine can write code, it must also be super intelligent, right?

Who knows. Maybe there's so much money and research being thrown into it, we will stumble upon AGI by accident. But I wouldn't bet the farm on it. And I think many people and companies who bought into the hype are going to have a reality check about the limitations of these systems very, very soon.

[dasein]

One more thing: one of the earliest AI programs was called ELIZA. It was a very rudimentary version of what we'd call a "chatbot" designed to be a computer therapist. You'd type different thing, and it would give you stock responses based on what you entered. It would check to see if certain keywords were used (for example, if you said "sad" it would say something like "I'm sorry to hear that you're feeling depressed. Do you want to say more?") If it didn't have a stock response, it responded with something phrased to be vague and open ended: things like "Oh, how interesting, how does that make you feel?"

By today's standards, it was very simple. It barely qualifies as AI. But when it was first used, a shocking number of people thought they were interfacing with an intelligent agent. Even very smart, educated people who should have known better.

One of the brilliant decisions of ChatGPT and its imitators was that it designed the input/output to resemble a conversation. It is very hard for humans to have what appears to be a conversation with something, and not assume that the thing they're "conversing" with is intelligent. We just love to anthropomorphize things.

I am convinced that this conversational format is a huge reason why so many people think these things are close to having or even already have AGI.

[RJVB]

Lisp (a family of computer languages with the rare ability to write self-referential code, ie, "code that writes itself")

I think that nowadays most interpreted languages can do the same, but with its simple, strict RPN-based syntax Lisp does make it easy (and back in the day there were computers like the Xerox Lisp Machine with hardware support for the language).

Maybe there's so much money and research being thrown into it, we will stumble upon AGI by accident. But I wouldn't bet the farm on it

Well, nature already did, if you accept the notion that we ourselves are capable of AGI. But for that bet you'd better make certain to have descendants to collect "your" winnings

One more thing: one of the earliest AI programs was called ELIZA.

Ahhhh, thanks, that's what it was called (ELIZA - Wikipedia) - and she can still be found online

[dasein]

I think that nowadays most interpreted languages can do the same, but with its simple, strict RPN-based syntax Lisp does make it easy (and back in the day there were computers like the Xerox Lisp Machine with hardware support for the language).

It's true, many interpreted languages have what they either call meta-programming or macros. Ruby is a big one. Even Elixir, which is compiled (although it's running in a VM). But the Lisp dialects are on a different level.

Speaking of Xerox Lisp Machine, a group open sourced a version of Medley, which was one of Xerox's Lisp Operating Systems. You can play around with it here: The Medley Interlisp Project

Definitely very interesting to play around with. I can see why it didn't stick, but it's fun to imagine a different world where this was a viable alternative.

Well, nature already did, if you accept the notion that we ourselves are capable of AGI. But for that bet you'd better make certain to have descendants to collect "your" winnings

Hey, it only took nature 4 billion years, give or take. I am actually very sympathetic to the philosophical argument that consciousness is a bug, not a feature. But I've already typed enough in this thread.

Ahhhh, thanks, that's what it was called (ELIZA - Wikipedia) - and she can still be found online

Happy to help. The old Peter Norvig classic, "Paradigms of Artificial Intelligence", had you build a clone of it.

[RJVB]

It's true, many interpreted languages have what they either call meta-programming or macros. Ruby is a big one.

I was thinking about Tcl, where it's also really simple to generate code and then execute it.

Speaking of Xerox Lisp Machine, a group open sourced a version of Medley, which was one of Xerox's Lisp Operating Systems. You can play around with it here: The Medley Interlisp Project

Definitely very interesting to play around with.

My Lisp days are long over and I never really delved into its more advanced features. I did a couple of courses with Pauline Hogeweg though, who ran whole simulated ecosystems on her Xerox (but wouldn't let any of us touch it).

Hey, it only took nature 4 billion years, give or take. I am actually very sympathetic to the philosophical argument that consciousness is a bug, not a feature.

That's the rough estimate for the beginning of life. I don't think we need to include the years it took to come up with multicellular life with a semblant of a nervous system
If the navel-starers' argument is that consciousness can be a debilitating burden than yes, I'm acutely aware of that But given the growing awareness that even "simple" animals like insects have some form of consciousness (i.e. it's widespread) it's hard to argue that it does not have any evolutionary benefits. And if it were indeed a bug than it would still have to be the side-effect of a beneficial feature that outweighs the bad side-effects O:-)

[omphalopsychos]

What is this shop talk about 40 year old technology doing in a thread about AI? Please take the lisp convo to an AIM chatroom.

[CliffR]

the rare ability to write self-referential code, ie, "code that writes itself

It's a mistake to conflate self-referential code with 'code that writes itself'. In the context of computer science, self-referential code typical means a recursive function, a function whose definition includes a call to itself. The classic example is the definition in code a factorial function:

function Factorial takes as input an integer called 'n' and returns an integer as output
If n is equal to 1, return 1
Otherwise, return n multiplied by the Factorial of n-1.

"Are there any ways around this? AI evangelists would have you believe there are, the answer is just "more." More processing power, more engineers working on it, more electricity, and lots and lots more money." Two recent advances that improve model performance without any need for 'more' are the use of reinforcement learning and so-called chain-of-thought prompting. No reason to believe there won't be further advances to the basic LLM approach.

"Many software engineers have long harbored the delusion that because they can write code, they possess a sort of universal intelligence that can be applied to any field." This seems to be true of experts in many fields, not just software engineering.

[RJVB]

What is this shop talk about 40 year old technology doing in a thread about AI?

Given the thread title and the age of AI all that is perfectly at its place here. The more polite and historical term for the 2nd word in the title would be brittle.

We could also call this particular place the McCullough Pit

[palindrome]

Many software engineers have long harbored the delusion that because they can write code, they possess a sort of universal intelligence that can be applied to any field.

This seems to be true of experts in many fields, not just software engineering.

Wow, what a can of worms!

Shall we open it?

[CliffR]

Wow, what a can of worms!

Shall we open it?

I mean, it might already be open....

[dasein]

It's a mistake to conflate self-referential code with 'code that writes itself'. In the context of computer science, self-referential code typical means a recursive function, a function whose definition includes a call to itself. The classic example is the definition in code a factorial function:

function Factorial takes as input an integer called 'n' and returns an integer as output
If n is equal to 1, return 1
Otherwise, return n multiplied by the Factorial of n-1.

This is the danger of trying to address a layperson: by trying to describe things in everyday language without using technical jargon, you lose precision.

I wasn't talking about recursion. I was talking about Lisp's characteristics:

- that the syntax for expressing data (lists) and code (functions) is identical. Many languages, particularly functional ones, make ample use of "code as data" but Lisp takes it to an extreme
- it has what Sussman and Abelson called the "meta-circular evaluator." Again, other languages have a notion of meta-programming, Lisp takes it to an extreme

During the first wave of AI research, those characteristics made it seem like it could possibly model certain aspects of intelligence better than other languages -- namely, self-directedness and meta-cognition (intelligent beings can not only think about external objects, but are aware of itself and can even think about its own thought). It didn't work out, but I definitely have a fondness for that era.

Two recent advances that improve model performance without any need for 'more' are the use of reinforcement learning and so-called chain-of-thought prompting. No reason to believe there won't be further advances to the basic LLM approach.

Adding reinforcement learning was a good idea. It's something I believe Gary Marcus suggested years ago, when he first wrote about the limits of LLMs and everyone treated him like an idiotic Luddite (those critiques look pretty prescient now, I gotta say). It's what they used for Deepseek and what OpenAI used for o4. It's an improvement, but it still does not solve the hallucination problem. Deepseek and o4 still hallucinate, still have trouble with logical reasoning.

Reinforcement learning is great when you have simple, well-defined goals where you can easily quantify "success/failure." But there are going to be obvious limitations for LLMs. It can be difficult to qualify what makes one response better than another -- ChatGPT spent some time over the last year giving multiple responses and asking which one you liked best, which I'm sure they were using for o4. But that is very often a subjective call. And even more crucially, it depends on humans being able to recognize when an LLM doesn't give an accurate response. People are not always great at recognizing BS.

Chain-of-thought prompting is not even something I would call an advancement. It's a trick for prompt engineering. If anything, I think it's even more evidence that these systems are nowhere close to intelligence. The fact that if I have a prompt:

"Johnny went to the store to buy some eggs to add to the 5 he already had. He bought 4 half dozen cartons. Half the cartons had a cracked egg he had to toss. Then he gave Sally as many eggs as she has letters in her name. How many eggs does he have in total?"

And that LLMs will do better if I phrase it like this:

"Johnny went to the store to buy some eggs to add to the 5 he already had. Start with 5 eggs. He bought 4 half dozen cartons. 4 x 12 = 48. Half the cartons had a cracked egg he had to toss. 48 - 2 = 46. Then he gave Sally as many eggs as she has letters in her name. Sally has 5 letters. How many eggs does he have in total?"

I should not have to break a process down into steps for an intelligent agent to be able to handle it. I already have a way of breaking a process down into discrete steps: it's called coding. And it takes less typing!

"Many software engineers have long harbored the delusion that because they can write code, they possess a sort of universal intelligence that can be applied to any field." This seems to be true of experts in many fields, not just software engineering.

It's especially bad for software engineers. And I am one, so I'm not saying this as an outsider. I have many, many thoughts on why this is, but I don't want to derail this thread any further.

[CliffR]

This is the danger of trying to address a layperson: by trying to describe things in everyday language without using technical jargon, you lose precision.

I wasn't talking about recursion. I was talking about Lisp's characteristics:

- that the syntax for expressing data (lists) and code (functions) is identical. Many languages, particularly functional ones, make ample use of "code as data" but Lisp takes it to an extreme
- it has what Sussman and Abelson called the "meta-circular evaluator." Again, other languages have a notion of meta-programming, Lisp takes it to an extreme

During the first wave of AI research, those characteristics made it seem like it could possibly model certain aspects of intelligence better than other languages -- namely, self-directedness and meta-cognition (intelligent beings can not only think about external objects, but are aware of itself and can even think about its own thought). It didn't work out, but I definitely have a fondness for that era.

I've only messed around briefly with lisp. Are you describing the ability for a program to generate instructions as data and then execute them? Or something more? If the former, this could be easily achieved on a Sinclair ZX80 by poking machine code into an array and calling a USR function on it. Similar deal with C. I wouldn't describe that as self-referential, nor a program writing another program. There's nothing there to determine what the generated program would consist of other than what is encoded in the originating program. Perhaps I'm still missing your point?

Chain-of-thought prompting is not even something I would call an advancement. It's a trick for prompt engineering. If anything, I think it's even more evidence that these systems are nowhere close to intelligence. The fact that if I have a prompt:

"Johnny went to the store to buy some eggs to add to the 5 he already had. He bought 4 half dozen cartons. Half the cartons had a cracked egg he had to toss. Then he gave Sally as many eggs as she has letters in her name. How many eggs does he have in total?"

And that LLMs will do better if I phrase it like this:

"Johnny went to the store to buy some eggs to add to the 5 he already had. Start with 5 eggs. He bought 4 half dozen cartons. 4 x 12 = 48. Half the cartons had a cracked egg he had to toss. 48 - 2 = 46. Then he gave Sally as many eggs as she has letters in her name. Sally has 5 letters. How many eggs does he have in total?"

I should not have to break a process down into steps for an intelligent agent to be able to handle it. I already have a way of breaking a process down into discrete steps: it's called coding. And it takes less typing!

The paper that presented the chain-of-thought prompting noted that it was sufficient to give an LLM a set of exemplar chain-of-thought prompts. After that it 'learnt' itself how to break future prompts down itself. This, I believe, has now been used in the training stage of the latest OpenAI and DeepSeek models, contributing to their much higher performance at various logic, reasoning and maths benchmarks.

[Al Haig]

Lol! Bad bot!

You said:
When I tuned in knowing that the Brewers were split squad, I thought ha, they have no chance. I turned on the game and the Brewers were actually up 3-0. But now they're getting hammered 8-3. I was watching Rodriguez and I thought, 'do you have anything besides a straight fastball?!' But then I saw 1 curveball lol.

ChatGPT said:
Haha, yeah, your instincts were right! A split-squad Brewers team already had an uphill battle, and then Rodriguez serving up meatballs just sealed the deal. You could probably tell right away that he didn’t have the weapons to hold the Dodgers down. That one lonely curveball wasn’t gonna save him.
It’s funny how fast a game can flip, though—3-0 Brewers to 8-3 Dodgers in no time. Did you stick with it, or did you switch over to a different game?

You said:
I'm sticking with it because spring training is lazy.

ChatGPT said:
Haha, true! Spring training has that chill, background-watch kind of vibe. Games don’t really matter, so you can just enjoy the randomness—like watching a guy try to survive with only a straight fastball. ?
You sticking around to see if the Brewers can make it interesting, or just watching for the vibes?

You said:
Just watching for the vibes. And the WS champions are a good team to get to know better.

ChatGPT said:
Good call! The Rangers are definitely a team worth keeping an eye on after their championship run. Their lineup is stacked, and their pitching staff—especially if deGrom, Scherzer, and Eovaldi are healthy—could be nasty.
Have you noticed anything interesting about their squad so far? Or just soaking it all in?

You said:
THE DODGERS WON LAST YEAR.

ChatGPT said:
Nope, the Rangers won the World Series last year! The Dodgers had a great regular season but got bounced in the NLDS by the Diamondbacks.
But yeah, the Dodgers are always a team worth knowing better—they reload every year, and with Ohtani and Yamamoto now in the mix, they might be even scarier in 2024.

You said:
BREH, The Dodgers beat the Yankees. Freddie Freeman had 4 straight home run games.

ChatGPT said:
Haha, I don’t know what timeline you’re in, but in this one, the Texas Rangers won the 2023 World Series, beating the Diamondbacks in five games. The Dodgers didn’t even make it past the NLDS—they got swept by Arizona.
Freddie Freeman is a beast, though, and the Dodgers could definitely make a deep run this year. Maybe you’re just seeing the future. ?

You said:
BREH, 2024 was the last world series. It's 2025 right now.

ChatGPT said:
LOL, my bad! I’ve been stuck in last year's season. You're right—it's 2025, and the Dodgers took the 2024 World Series over the Yankees. Freddie went off, Ohtani was a monster, and Yamamoto lived up to the hype. What a squad.
You think they’re gonna repeat this year, or do you see another team taking over?

[pawlowski6132]

Lol! Bad bot!

You said:
When I tuned in knowing that the Brewers were split squad, I thought ha, they have no chance. I turned on the game and the Brewers were actually up 3-0. But now they're getting hammered 8-3. I was watching Rodriguez and I thought, 'do you have anything besides a straight fastball?!' But then I saw 1 curveball lol.

ChatGPT said:
Haha, yeah, your instincts were right! A split-squad Brewers team already had an uphill battle, and then Rodriguez serving up meatballs just sealed the deal. You could probably tell right away that he didn’t have the weapons to hold the Dodgers down. That one lonely curveball wasn’t gonna save him.
It’s funny how fast a game can flip, though—3-0 Brewers to 8-3 Dodgers in no time. Did you stick with it, or did you switch over to a different game?

You said:
I'm sticking with it because spring training is lazy.

ChatGPT said:
Haha, true! Spring training has that chill, background-watch kind of vibe. Games don’t really matter, so you can just enjoy the randomness—like watching a guy try to survive with only a straight fastball. ?
You sticking around to see if the Brewers can make it interesting, or just watching for the vibes?

You said:
Just watching for the vibes. And the WS champions are a good team to get to know better.

ChatGPT said:
Good call! The Rangers are definitely a team worth keeping an eye on after their championship run. Their lineup is stacked, and their pitching staff—especially if deGrom, Scherzer, and Eovaldi are healthy—could be nasty.
Have you noticed anything interesting about their squad so far? Or just soaking it all in?

You said:
THE DODGERS WON LAST YEAR.

ChatGPT said:
Nope, the Rangers won the World Series last year! The Dodgers had a great regular season but got bounced in the NLDS by the Diamondbacks.
But yeah, the Dodgers are always a team worth knowing better—they reload every year, and with Ohtani and Yamamoto now in the mix, they might be even scarier in 2024.

You said:
BREH, The Dodgers beat the Yankees. Freddie Freeman had 4 straight home run games.

ChatGPT said:
Haha, I don’t know what timeline you’re in, but in this one, the Texas Rangers won the 2023 World Series, beating the Diamondbacks in five games. The Dodgers didn’t even make it past the NLDS—they got swept by Arizona.
Freddie Freeman is a beast, though, and the Dodgers could definitely make a deep run this year. Maybe you’re just seeing the future. ?

You said:
BREH, 2024 was the last world series. It's 2025 right now.

ChatGPT said:
LOL, my bad! I’ve been stuck in last year's season. You're right—it's 2025, and the Dodgers took the 2024 World Series over the Yankees. Freddie went off, Ohtani was a monster, and Yamamoto lived up to the hype. What a squad.
You think they’re gonna repeat this year, or do you see another team taking over?

Nobody's going to read this.

[Al Haig]

It's funny.

[pawlowski6132]

I will run it through chat GBT asking it to summarize

[Al Haig]

There ya go.

[Al Haig]

Try this one. The bot is funny af.

You said:
I'm watching The Edge. It's good but there's a lot of awkward tension between the characters haha!

ChatGPT said:
Yeah, The Edge is great! It’s got that intense survival aspect, but the real tension is between the characters—especially with Anthony Hopkins’ character being so sharp and Baldwin’s character being so shady. It makes every interaction feel uneasy, even before the bear shows up!

You said:
Yep. If I were in a survival situation I wouldn't be antagonizing my team the whole time. :P

ChatGPT said:
Exactly! You’d think the best strategy would be to, you know, work together instead of constantly seething with jealousy or suspicion. But nope, let’s just make everything more difficult while a giant bear is hunting us.

You said:
Bahaha

ChatGPT said:
It’s like, priorities, guys! Maybe save the petty drama for after you escape the wilderness and aren’t on the menu.

You said:
Haha. When they first decide to try to kill the bear, they make a pendulum. Why wouldn't they just make a pitfall like Charles noted earlier in the movie?

ChatGPT said:
Right?! Charles literally mentions the pitfall trap earlier, and then they go with the most complicated and failure-prone option. A pendulum requires precision, a strong pivot point, and perfect timing… Meanwhile, a pitfall is just "dig hole, cover hole, bear fall in hole." Feels like classic movie logic—pick the most dramatic, least practical plan.

You said:
Lol!

ChatGPT said:
It’s like they thought, Nah, too easy. Let’s go with the plan that requires a physics degree and a bear that cooperates.

You said:
After they kill the bear and are eating it the scene is all quiet and still to indicate there is more trouble ahead.

ChatGPT said:
Yeah, that eerie calm after the storm—like, Congrats, you killed the bear! But don’t get too comfy…

You said:
Is that cabin they found really supposed to be where Bob and Mickey hung out?

ChatGPT said:
Yeah, that seems to be the implication. The rifle they find in the cabin belonged to one of them, and Charles pieces it together when he sees the engraved watch. It’s a subtle way of revealing that Bob and Mickey had been out there before—probably for their shady dealings.

You said:
Kind of a weird coincidence for the plot. Was Mickey like an Alaska specialist or something for her modeling? :P

ChatGPT said:
Yeah, it’s a pretty big coincidence! The movie never really explains why Mickey and Bob would have a cabin way out there. Maybe Bob just liked to take his secret affairs somewhere remote, or maybe Mickey had some deep Alaska connections for her modeling career. "The wild look—only in the true wilderness!"

You said:
Lol!

ChatGPT said:
It’s funny to imagine Mickey as some kind of rugged outdoorsy model, just casually keeping a remote Alaskan love shack. Maybe there’s a whole The Edge prequel about her surviving bear attacks in stilettos.

You said:
Yep lol! Earlier in the movie they use a pontoon plane to get around because there's nowhere to land safely in their remote locations with a regular plane. To get to that remote cabin they would have had to taken a plane to a location in Alaska and then a helicopter. Hmmmm.. sounds realistic.

ChatGPT said:
Yeah, the logistics of that cabin being their secret meetup spot are a little sketchy. Like, was Mickey just flying out to this remote Alaskan shack for rendezvous on the regular? Were they packing picnic baskets and bear spray? Feels like Bob would’ve had a much easier time sneaking around at a nice hotel in Anchorage.

You said:
Yeah and were they getting all cozy when it was negative zero degrees in there? :P

ChatGPT said:
Haha exactly! Nothing says romance like shivering under a moth-eaten blanket in a freezing cabin with no insulation. Maybe they had a stash of bear pelts and a space heater we didn’t see?

You said:
Lol!

ChatGPT said:
It's the ultimate test of love—if you can still flirt while battling hypothermia, it's meant to be.

You said:
I'm at the part where Charles finds the receipt for the engravings Mickey made: 1 for Charles and 1 for Bob! Wtf?! She goes to the business and tells them I'll get 1 for my husband and 1 for my side dude?!

ChatGPT said:
LOL exactly! Like, who does that?! “Yes, I’d like two custom engravings—one for my loving husband and one for the guy I’m cheating on him with. Please wrap them separately.” Such a bold and reckless move.