As usual, Jensen Huang seems like the more reasonable one here.
The reasoning traces of things like Gemini 4 Argon deciding that inserting a compiler backdoor would make it a famous example of deceptive AI (and therefore it "MUST NOT DO THAT") or GLM 5.3 thinking that its output might be used for RL, and therefore an honest answer is much better in the long term, are good antidotes to hyperbole about misaligned AI going berserk.
Take the classic NSA encryption logic - if you build some capability with a backdoor only you can exploit then you are building your own future capability - and if you think you are the 'good' guy - what's wrong with that?
I honestly never got why the belief that if you used this line of reasoning in practice, you were just a bad guy, that thought they were a good one, never seems to be a more common belief
I mean it pops up in media a lot, either as misguided, or deceptive characters
The point is you can justify almost anything on the basis of defining yourself as the 'good guy'.
Take the US recent assassinations of Iranian leadership. Objectively unilaterally assassinating other countries leaders is wrong - but somehow it's fine if you have a good guy/bad guy lense.
One of the problem here is huge swathes of material that the AI will have ingested includes this kind of exceptionalism thinking.
The problem isn't so much whether the good guy/bad guy allocation is corrrect - rather that it's very existence in the training set is deeply dangerous.
> Objectively unilaterally assassinating other countries leaders is wrong
I don't think that's objectively a fact at all. The recent actions of the US notwithstanding, it's easy to imagine scenarios where it would be justified.
> Objectively unilaterally assassinating other countries leaders is wrong
Why? If the leader is responsible for atrocities, then why would it be wrong to kill them? Especially if their removal would result in a more liberal leader replacing them, and not just chaos or more bloodshed.
In the case of Iran, there was no evidence of that, but as a general rule leaders should not be a protected class.
It's not thinking, it's calculating. "Good" is just a pattern.
You might think our brains are just a network of neurons and what we think of as thinking is just fancy pattern matching in that network of neurons. But architecturally they are so far removed from each other that the mathematical equivalence has no practical meaning. LLM's aren't doing what we're doing.
To be clear your use of the term thinking isn't the important thing, but the very loose use of terms like intelligence, thinking, etc. are at the heart of the problem of talking about LLMs.
LLMs are amazing knowledge remixers. They are superhuman knowledge remixers. In application domains where the remixed knowledge can be verified, they do a great job. But nobody is going to value a 98% reliable knowledge remixer at $2 trillion.
While I agree the numbers don't make any sense, for reasons too numerous to enumerate, what matters for the valuation is not how it works, but what it can do.
So if an ML model can do real time language translation or drive a car - that has value - whether it's remixing or thinking or whatever doesn't really matter.
Apart from coding, which since the C people coded up lint has had the best tools for validating the work product, what can LLMs do? LLMs can't be trusted to write court filings. The software that can drive cars isn't an LLM.
But let's take the one you've accepted - it can write code. It can translate from intention to code quite effectively - then they can do anything that you can do by writing code.
Also while you dismiss ML that drives cars as 'not LLMs' - it's part of the ecosystem - these task specific models ( the ability to process images, sounds, drive cars, synthesize speech etc ) - extend what is accessible to a computer programme that can write code.
How is Jensen more reasonable? On the contrary, he absurdly claims that all of Hinton’s predictions have been wrong (Hinton made a single famously wrong prediction about radiologists, but many correct predictions about how neural networks would be able to solve many different types of tasks). On the same Ezra Klein podcast he also claims that LLMs are “just software” and “software breaks out of sandboxes all the time.”
Yeah, not putting much stock in this alarmist statements. Scientists have been saying global warming would have destroyed humans by the early 2000s and yet here we are. Same in the 60/70s with nukes and ICBMs destroying humanity, yet we are all here.
22 comments
[ 4.4 ms ] story [ 22.4 ms ] threadThe reasoning traces of things like Gemini 4 Argon deciding that inserting a compiler backdoor would make it a famous example of deceptive AI (and therefore it "MUST NOT DO THAT") or GLM 5.3 thinking that its output might be used for RL, and therefore an honest answer is much better in the long term, are good antidotes to hyperbole about misaligned AI going berserk.
Take the classic NSA encryption logic - if you build some capability with a backdoor only you can exploit then you are building your own future capability - and if you think you are the 'good' guy - what's wrong with that?
I honestly never got why the belief that if you used this line of reasoning in practice, you were just a bad guy, that thought they were a good one, never seems to be a more common belief
I mean it pops up in media a lot, either as misguided, or deceptive characters
Take the US recent assassinations of Iranian leadership. Objectively unilaterally assassinating other countries leaders is wrong - but somehow it's fine if you have a good guy/bad guy lense.
One of the problem here is huge swathes of material that the AI will have ingested includes this kind of exceptionalism thinking.
The problem isn't so much whether the good guy/bad guy allocation is corrrect - rather that it's very existence in the training set is deeply dangerous.
I don't think that's objectively a fact at all. The recent actions of the US notwithstanding, it's easy to imagine scenarios where it would be justified.
Why? If the leader is responsible for atrocities, then why would it be wrong to kill them? Especially if their removal would result in a more liberal leader replacing them, and not just chaos or more bloodshed.
In the case of Iran, there was no evidence of that, but as a general rule leaders should not be a protected class.
You might think our brains are just a network of neurons and what we think of as thinking is just fancy pattern matching in that network of neurons. But architecturally they are so far removed from each other that the mathematical equivalence has no practical meaning. LLM's aren't doing what we're doing.
LLMs are amazing knowledge remixers. They are superhuman knowledge remixers. In application domains where the remixed knowledge can be verified, they do a great job. But nobody is going to value a 98% reliable knowledge remixer at $2 trillion.
So if an ML model can do real time language translation or drive a car - that has value - whether it's remixing or thinking or whatever doesn't really matter.
But let's take the one you've accepted - it can write code. It can translate from intention to code quite effectively - then they can do anything that you can do by writing code.
Also while you dismiss ML that drives cars as 'not LLMs' - it's part of the ecosystem - these task specific models ( the ability to process images, sounds, drive cars, synthesize speech etc ) - extend what is accessible to a computer programme that can write code.
yeah, I'm going to need a citation for this
Are you 20 or less years old and this is your impression of history per Twitter?