The AI crash is impossible - Change my View

[Leo Gura]

On 8/24/2026 at 11:50 PM, Marten98 said:

@Leo Gura didn‘t you say a couple weeks ago that your strategy was to stack gold, since this is a foolproof way of investing, especially when stocks are down? And that gold is fairly cheap right now.

If the bubble bursts, even gold will drop 30%.

So I am careful even about gold.

Gold is the best investment right now, but still risky given how hard this market could crash.

When the market is crashing, the only safe place is cash. Of course inflation is also a risk, so there is no perfect answer.

Edited 6 hours ago by Leo Gura

[oOo]

14 hours ago, Elliott said:

That's not "AI", that's a program made by "AI". "AI" means using it on something novel which means needing all data, processing a TON of data.

Elliott, I think I've picked up on some patterns so I am going to take a stab at your algorithm:

• Say something absurd with enormous confidence
• Someone explains why it’s wrong.
• Redefine the word.
• If that fails, announce the source is trash.
• Repeat until everyone gets tired.

Here it went - “entire building” > “that’s not AI” > “AI needs all data” > Photoshop > spaceships > curing cancer.

You sound like an early-release reasoning model with no error correction. Aggressive token usage with compounding hallucinations. 

To go back on what I said briefly, you're the only reason why I wouldn't want every homeowner to have its own miniature data centre, many would use it to power their irrationalities on the internet while intermittently coming together to form their red-pill, feminist and religious propaganda, and from that, your odd school shooting with 2035 AI robots created by a kid that got rejected by his crush.

Your points are:

• Cloud AI can be cheaper for low-use customers
• Frontier model training still requires enormous datasets and comput
• Outsourcing sensitive data can create genuine privacy risks
• Local AI will not economically suit every user

Big fucking deal. These are obvious points you're making with even more obvious solutions, most of which I've already stated and or alluded to. You sound like the Greta Thunberg of AI, but far more annoying, less educated, and more oblivious. 

Your audience is giving this way too much compute. I've given this way too much compute, this is unpaid digital standup right here on my part.

Politely, I am now walking away for good. 

12 hours ago, Pudgey said:

I don't see how Anthropic and OpenAI survives if open weight models are becoming this good.

They're absolutely fucked.

Google, Meta, Amazon, Microsoft, xAI all got their own data centres. Anthropic and OpenAI mostly rent theirs like movies from a video store, they'll never be able to become casting directors. Everyone remember the days of renting films from video stores? Ha!

OpenAI's literal business model is now no more sophisticated than "we will make the best models". Even if that were true, which it isn't, for how long would it be true? I expect bankruptcy within 5 years or a radically renovated business model they lead with. 

Their only chance is to rub shoulders with government. That's literally it. Maybe Trump will give em some datacenters they build in Greenland after the fantasy takeover! Either way, Sam Altman's screwed outside the pay packet he'll leave with after screwing investors. They get more funding than any other AI company and are no where near the best while spending the most. 

Perplexity is just as bad, if not worse, they don't even make their own models.

Quote

LLM prices are crashing to the bottom and it doesn't look like it's stopping. 

Yep, and now the infrastructure surrounding.

For me as I expressed, now more than ever over the last 100 yrs we have a chance for "centralisation collapse" in the next decade, or at the very least, a positive derailment.

It's the best outcome for a situation in which the general population is outnumbered by negative alternatives, IMO.

 

Edited 6 hours ago by oOo

[PenguinPablo]

On 3/4/2026 at 11:21 AM, LordFall said:

Leo I don't know what to tell you man I think think you're basically close to 100% out of the loop on this. 

You are already pretty close to being able to create a whole game with physics engine, novel characters, custom gameplay and a logic system off of a single prompt. 

Why do you reduce AI agents's potential in gaming to quirky dialogue? You could make an alternative american Civil war campaign where the confederates somehow allied with Canada and Mexico against the Union and have a whole alternative history plotline with characters that live out a 24 hour storyline with real strategy and twitsts where players log in to take part of an ongoing campaign. So many cool ideas that were never possible beforehand. 

The game I wanna make will simulate a fantasy 15th century europe with active AI agents that operate jobs and labor for a crypto token with a whole simulated economy. That level of game has never been close to being possible before. 

Let's have this conversation again in 5 years when everyone on this forum including you has one. Human relationships are gonna change forever. You're already pretty close to having a companion on your phone that answers your every question and is a more interesting conversation partner than most humans that don't use AI. 

It won't transition to a weird dystopia it'll just empower people to have more intentional relationships on their terms. Human connections will be more powerful than ever. Do you really put human relationships on such a pedestal? A lot of couples you could literally plot their conversations on a flowchart and be accurate in a majority of cases, consciousness has patterns. That's the whole idea of pickup you can go out and have the same conversations over years with 100% different people because conversation threading is pretty similar across people. 

This is true and already happening via podcasts and influencers so I'd predict the same as you. All the weird and taboo things have basically become the norm. Online dating used to be a badge of shame now it's what everyone does.

I suppose cam girls are a precursor to only fans and influencers. I think the 5 year thing is true but I don't think it will be so wide spread that Leo and everyone on this forum will have an AI girlfriend. It's pretty cringe tbh. 

But then again the movie Her basically was a prophecy. 

Edited 4 hours ago by PenguinPablo

[oOo]

1 hour ago, PenguinPablo said:

This is true and already happening via podcasts and influencers so I'd predict the same as you. All the weird and taboo things have basically become the norm. Online dating used to be a badge of shame now it's what everyone does.

I suppose cam girls are a precursor to only fans and influencers. I think the 5 year thing is true but I don't think it will be so wide spread that Leo and everyone on this forum will have an AI girlfriend. It's pretty cringe tbh. 

But then again the movie Her basically was a prophecy. 

I can't do the AI partner thing but, my brain simply won't buy in despite trying to convince myself as a self-experimentation. 

I think personality variables are a prerequisite.

For example, someone on high on openness to experience may seem more likely in the beginning but then they also may be more likely to become bored or find it too shallow, especially if they have other pairing qualities like high on sentimentality. That said, people who are already objectifying their relationships, like those that are high in openness but low in sentimentality may be more likely; another sub-domain is fantasy proneness minus reality-testing habits. 

And even Cognitive dispositions.

For example those that are more likely to take many different perspectives into account may be less likely to buy into having an AI girlfriend, simply because say, switching between third and first person as one take, creates those noted reality tests that more often than not are likely to break the AI illusion. The mention of dating apps becoming socialised is a good example of this, many people get gamed by digital dating frontiers and that correlation tends to correspond to reality-testing psychological habits.

If it becomes socialised though, that's another variable of course that shouldn't be ignored however at the same time, one can't just expect the opposite to not happen. Meaning group norms forge among certain demographics that outright reject the idea of an AI partner. So we have simultaneous agendas from people less/more likely to just like anything else. 

Another one is attachment style as well, low income earners, people that have lower self-esteem and people that are over-weight; just to name others off the cuff.

It's interesting though. 

I think we will get real life examples of the film "Her" playing out as they already have with some people demonstrating psychotic symptoms following deeper AI interactions, that said, I think that the majority of those examples are going to continue to be a super-dumbed down version of what's portrayed in the film, which was pretty solid for its time. 

Transcendence as a film was epic for its time as well, and we will see dumbed down versions of that coming; with the closest relationship there being synthetic neurology  that are at best promising a scam in the intelligence domain at this point. Lucy and Limitless are ways off and have more security issues at this point as you wouldn't want a Limitless pill being widely available to geopolitical competitors so synthetic neurological enhancements are always going to be a slippery slope over what's accepted (I.e. socially), what's safe (I.e. security) and what's competitive (I.e. geopolitics). With AI partners its similar keys, with safety and competition taking on different positions, for example as a competitive country you'd actually want the foreign country's population forming relationships with AI analogous to the benefits Pokemon Go  brought intelligence agencies from positioning that is as simple as data accumulation to social influence and impacting the trajectory of relationship formations, I.e. from a strategic interest, its better if it leads to less marriages, less children, poorly raised children, worse marriages, less reality checks, etc. 

Just as some alternate takes. Otherwise Project Genie is a must checkout. 

 

Edited 3 hours ago by oOo

[something_else]

Quote

Without data? No. What you're referring to is a program created by AI.

I'm sorry, but I think there are several issues in your thinking. Most of all, you're introducing a lot of other points separate from your main argument which was that "local LLMs are not AI, and AI is only happening when training the model" so I'm going to focus on that point in my response here.

AI is defined as: "the capability of computational systems to perform tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, perception, and decision-making".

If you run a 120B parameter model locally and attach it to a good agent harness, it meets this criteria. Also notice that there is no mention of 'data' or 'training' being a requirement for something to be AI like you are suggesting.

In fact, I would argue your thinking is exactly backwards. The training process for an AI is mostly defined and programmed by humans, while the output of an LLM is produced independently of humans, which more closely matches the definition of AI.

Quote

Its a local small-language model.

Most locally run models are medium sized. Small language model defines anything under 10B parameters. Medium is 10B-100B, large is 100B+. It is possible (although it requires workstation level hardware) to run a 120B parameter model locally. Medium language models are very achievable to run locally, and still meet much of the above criteria for AI, just to less of a degree than a large language model.

The agent harness is actually the magic sauce that makes modern AI agents so capable, less so than model size.

Quote

You can't just download a llm on a local lab and ask it to create a spaceship, and then how to cure cancer

You can't ask Fable to do this either. This has little bearing on whether something qualifies as AI.

Quote

It will be a calculator, not a polymath ChatGPT

Ehm, no not really. LLMs aren't really trained for specific tasks, even the small ones. That's kind of the whole point. The initial LLM is trained on very broad corpus of text, then fine-tuned using a combination of question-answer pairs to optimise it for answering questions, and preference optimisation to ensure it answers questions in a way actual humans find useful. The end result is something with a fairly broad range of knowledge that knows how to talk to humans, even if the model is small.

Yes, if you need a model for aerospace engineering you may include lots of aerospace engineering content in that fine-tuning phase, but you're still building on an LLM that was initially trained on a very broad corpus of text which makes it much more general purpose than a calculator.

On top of this, the model runs in an agent harness, this allows it to decide whether it needs more data to make a decision. If it thinks it does then it can go and fetch that. So if you ask it about the anatomy of a cow it will search online for information on that and store that in its context, allowing it to answer detailed questions about cow anatomy pretty effectively even though it was fine-tuned as an aerospace engineering model.

Edited 1 hour ago by something_else