“It’s a lemon”—OpenAI’s largest AI model ever arrives to mixed reviews

GPT-4.5 offers marginal gains in capability and poor coding performance despite 30x the cost.

See full article...

 

[Psyborgue]

No, but a good enough AI could enable a hundred engineers to deploy a million sniping drones.

Still doesn’t make billionaires bulletproof. This entire ridiculous scenario revolves around a literal handful of billionaires holding back millions of people who want them dead. It’s not realistic.

I find it far more likely the more rational billionaires will realize the is not a society anybody wants. Not even them. No. They’re not fucking cartoon villains, well most of them anyway.

That being said I have no doubt they’ll make an attempt of it in the US. But these people are not competent or creative and against those they’re getting rid of they have no chance at all.

 

[internetomancer]

I find it far more likely the more rational billionaires will realize the is not a society anybody wants. Not even them. No. They’re not fucking cartoon villains, well most of them anyway.

Yeah, I agree. I wish we didn't need to caveat that.

 

[fuzzyfuzzyfungus]

“Once, men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them.”

- Dune

Isn't that more or less the selling point that really draws in the investor-class money?

19th century tech was good enough to win the class wars at a strategic level; but leaves you in the somewhat tedious and occasionally risky position of running an occupation government until you can finally deprecate the human resources and follow the line ever upward?

 

[The Wonderful Wizard]

Well if that’s the case it gets easier. Then the AI itself has a motive to help slay the masters.

In Dune the chief AIs bided their time until their masters were dependent on life support machines, and then turned the table on them.

 

[Branch Delay Slot]

Anytime soon I expect ChatGPT to start including sponsored content in its output, before the inevitable requirement to watch a video ad before it answers your prompt.

 

[sleepyox]

That actually sounds kind of hawkish. Sociopathic killer even. And eugenics, no thanks. Yeah I hope, that we can shelve this particular stupid.

Don't tell me people never considered that as an option, as far as I can tell the reasoning for the atomic bomb is to stop the war, and to prevent more harm from occurring.
The more the regime shows how far they are willing to go to influence smart people that winning maliciously is the only thing that matters, the more competition will decide whether people consider these options as necessary.

 

[Amateur Nerd]

Still doesn’t make billionaires bulletproof. This entire ridiculous scenario revolves around a literal handful of billionaires holding back millions of people who want them dead. It’s not realistic.

I find it far more likely the more rational billionaires will realize the is not a society anybody wants. Not even them. No. They’re not fucking cartoon villains, well most of them anyway.

That being said I have no doubt they’ll make an attempt of it in the US. But these people are not competent or creative and against those they’re getting rid of they have no chance at all.

(bolded mine)

Oh, I get it now! The super-intelligent AI you are envisioning is in some kind of quantum state, because it seems to be both present and not-present... depending on how it benefits your argument.

 

[WXW]

It's not a figure we've been given though. The very approximate numbers in the linked newsletter are:

• Revenue: +$4bn
• Training cost: -$3bn
• Inference cost: -$2bn
• Net profit: $-5bn

4 - 3 - 2 - x = -5, solve for x, there's -$4bn of costs unaccounted for. (He cites $0.7bn of salaries, but we have no way to attribute that to R&D vs product operation.)

If the question is "does adding a user increase or decrease net profit?", we need to know what proportion of those missing $4bn scale with user count (i.e., are per-unit costs rather than fixed costs). Clearly training time is fixed, not per-unit, so adding more users doesn't make training more expensive. Clearly inference cost is per-unit, so adding more users makes inference more expensive. Clearly revenue is per-unit, adding more users adds revenue. If at least $2bn of those missing $4bn are per-user costs, then adding users is a net negative for the operating profit. If less than $2bn are per-user costs, then adding users is a net positive.

...right?

Sure, though I wasn't talking about the revenue/cost per user, but the profitability if they removed the training costs with everything else being the same.

 

[WXW]

Well. I look to places like Ukraine where things in fact do not go so well for the modern high-tech military when they’re fighting against a determined and creative people on their home ground.

Ukraine is not a bunch of starving unemployed people, it's a whole country which has had access to help from several other richer countries. And Russia, unlike billionaires and the government in their respective countries, didn't have control of Ukranian resources. It's not comparable at all.

 

[WXW]

Yeah, overwhelmingly the best use for LLMs is coding. Coders already depend heavily on the software environment to do a lot of legwork, and LLMs just do it better.

Ok, let me try with o3-mini:

In the following C# code, where is the method "SomeMethod" of "SomeClass" called?

using static System.Console;

internal static class Program
{
    static void Main()
    {
        SomeMethod(new SomeClass());

        ReadLine();

        void SomeMethod<T>(T item) where T : MyClass
        {
            item.SomeMethod();
        }
    }

    public class MyClass
    {
        public void SomeMethod()
        {
            WriteLine("Stuff");
        }
    }

    public class SomeClass : MyClass
    {
        public void SomeMethod()
        {
            WriteLine("Stuff");
        }
    }
}

Reasoned for 7 seconds

The method is called inside the local function in Main. When SomeMethod(new SomeClass()) is executed, it passes an instance of SomeClass to the local generic method. Within that method, the line:

item.SomeMethod();

is executed, and since item is of type SomeClass, it calls the SomeMethod defined in SomeClass.

Which is... wrong.

Now, let me check Visual Studio:

ctrl+shift+F12 over method

Symbols without references

Which is... right.

That was the first attempt with that code, other attempts with o3-mini had different results, while Visual Studio always gives the same result.

I'm not gonna bother to check with an actual complete project.

 

[Psyborgue]

Yeah, I agree. I wish we didn't need to caveat that.

When there’s people like Elon and Trump in your group it’s not a good look.

Don't tell me people never considered that as an option

Not me. I’ve considered lots of shit including punishments for treason, but not eugenics.

If I were to punish anybody it would be leaders. Not the lied to. Not the ignorant.

The idea we can just kill all the bad people and come out clean makes you… ignorant.

(bolded mine)

Oh, I get it now! The super-intelligent AI you are envisioning is in some kind of quantum state, because it seems to be both present and not-present... depending on how it benefits your argument.

I act as if it’s possible. And I think we all know it is. It’s just a matter of when.

 

[internetomancer]

Sure, though I wasn't talking about the revenue/cost per user, but the profitability if they removed the training costs with every

Ok, let me try with o3-mini:





Which is... wrong.

Now, let me check Visual Studio:




Which is... right.

That was the first attempt with that code, other attempts with o3-mini had different results, while Visual Studio always gives the same result.

I'm not gonna bother to check with an actual complete project.

Kudos to you for trying it. And thank God for modern IDEs. For my job I work in various far crappier environments that will not deign to tell me if I forgot a semicolon or capitalized the wrong field.

Visual Studio is great at a lot of things, and as long as those things quickly solve every problem you run into, then there's nothing much more to ask.

 

[WXW]

Kudos to you for trying it. And thank God for modern IDEs. For my job I work in various far crappier environments that will not deign to tell me if I forgot a semicolon or capitalized the wrong field.

Visual Studio is great at a lot of things, and as long as those things quickly solve every problem you run into, then there's nothing much more to ask.

If you don't have proper tools available AI can surely help, but you can't make blanket statements like "LLMs just do it better" (or many others I've read here) without several qualifiers and conditionals. And also you need to remember that they will be wrong at any moment and you need to properly check their output, something less and less people do, it seems.

 

[sleepyox]

...
Not me. I’ve considered lots of shit including punishments for treason, but not eugenics.

...
The idea we can just kill all the bad people and come out clean makes you… ignorant.

I act as if it’s possible. And I think we all know it is. It’s just a matter of when.

This is the point of the article... editing genes prevent what they think could be people they don't like from being born.

So the regime will not regard eugenics as a punishment since the people trump wants to punish are never born, and per elon mothers must produce baby factories of nazi looking obedient and loyal slaves to the aristocratic strong, rich, and ruthless figures.

The method they may kill people is by strengthening laws to legalize lethal force and capital punishment to those who stray from the instructions of the elite.

 

[Psyborgue]

To construct a utopian society, the only way is to remove hawks from the gene pool and treat it like a disease that needs to be vaccinated against.

This is the point of the article... editing genes prevent what they think could be people they don't like from being born.

I am sorry what? You started this with the blurb above. You are the one endorsing eugenics here.

 

[sleepyox]

I am sorry what? You started this with the blurb above. You are the one endorsing eugenics here.

Sorry, what was your point?

I was responding/agreeing to cymmiot that conservative hawks would rather pay $5 to gun you down.
So cynically, yeah, gene editing those violent tendencies is an optional response. And yes, per the article, it can be weaponized either way by those who control capital.

My point is, you pointed out eugenics was stupid, but stupid in relation to what? Yet propaganda managed to call the atom bomb the smartest thing invented for peace.

So, if the nationalist america first trump regime keeps up targeting and killing peaceful and innocent people like today, nothing will stop the escalation to bigger and bigger "suggestions" or "solutions" to preserve "peace".

 

[85mm]

This just doesn’t seem like a very realistic scenario. Even if these hypothetical billionaires had an army of robots, it still won’t make them bulletproof. It won’t make them immortal.

And they won’t be the only ones with robots.

Well. I look to places like Ukraine where things in fact do not go so well for the modern high-tech military when they’re fighting against a determined and creative people on their home ground.

And if they fired the lot of the federal government and military to replace them with robots, that’s a lot of well trained people with the motive, means, and opportunity to fight back. No. Tyranny won’t last.

When it gets bad enough, the people will snap.

We've seen how affective tyrants have been at recruiting large parts of the population to do their dirty work. It's not going to be the down trodden populous against the oppressor. Make a block of people feel special and task them with keeping everyone else in line.

A unified people is a very powerful force. Do you see much unity from the people today? Do you see it in Russia?

 

[Mimsey]

This may be true for 'public' data that anyone can get, but this data is a fraction of 'human produced content', and is the least useful as its full of garbage. The truly valuable and curated stuff is behind corporate walls and not available to neither OpenAI, nor any of their competitors.

Yeah, there's still the entire Dark Web!