Ok, I'll say it.
The news is: "AI is not that powerful or functional".
Sure, can change some little things but not that much. And we are overspending on it.
I get people who said things like that like year ago, but nowadays it's a laughable position. Solving millenium problem means nothing nowadays? Exponential increase in CVE is nothing?
I’m mostly bearish on AI, but this is flat wrong. The vast majority of CVEs that have been discovered with AI assistance were in code written long before AIs arrived on the scene.
from TFA, I liked seeing this, considering how influential Google HR policy choices are across the valley and beyond: “Recently, we’ve updated our terms for new hires, which included removing the post-termination aspect of non-disparagement clauses.”
Haven't we already heard the worst of it? If you let a swarm of agents on the loose with a goal, they will get to all kind of nefarious activity in pursuit of that goal. What could possibly be worse than that?
If that’s the reality, you would think that employees would be willing to break an NDA for it. “We’re all going to die, but I really need my bonus this year” seems like a parody rather than a real strongly held position.
If anything this points to a more moderate (but still bad) worry among the employees.
I suppose there may be a bit of a prisoner's dilemma to it. Single voices can be discredited. It would take a coordinated disclosure to actually make the message stick.
That's sci-fi. They cannot "recursively improve". There is no magic to LLMs. An agent swarm can increase it's thinking power by increasing the number of agents, the thinking time and by changing the base model to a more powerful one. So at worst they can assign themselves more compute or hijack another agent swarm based on a more powerful model.
Sorry bro, we live in a sci-fi world. I don't know what you think recursive improvement is, but we've been doing it for some time with models making synthetic data models and using it to train the next model. This is being done inside the labs intentionally to avoid as many employees and humans in post training. Works very well for things like math and cyber-sec.
> I don't know what you think recursive improvement is
The ability for a model to improve itself over time without outside intervention. As yet, no models yet have this capability, and LLMs specifically what with how they are constructed, are not likely to ever have it. They are frozen in carbonite at the time of their training.
That's not to say they can't use skills to appear somewhat worldly, things like searching the internet and what have you, but if you ask a model to strictly stick to data in it's training, it will do that and it will be stuck in whatever time in which it was built.
Energy is conserved, gravity exists, and LLMs aren't intelligent.
LLMs today are improved by experiments that assay their performance, changes to architecture, pretraining, and post-training. None of those steps necessarily require humans. Welcome to your sci fi future.
Just, Devil's Advocate, but what would be the benefit of your career if all humans (or even just a few billions of humans) are going to die?
I think it's reasonable to assume that if it's not worth breaking the NDA, it's likely not an existential threat for the vast majority of humanity.
Now, does that mean that people who currently work upper-middle class knowledge work jobs aren't in for an exceedingly harsh economic reality check? Absolutely not.
But that's not really as serious as an existential threat to humanity.
I doubt you are asking in good faith, because I suspect a good faith interlocutor would have worked this out on their own, but I'll proceed anyway.
On the one hand, you've potentially got an indeterminate number of future comfortable years based on a subjective assessment of how long you think it takes catastrophe to materialize, with the downside that you can't warn anyone and so no one will stop the coming catastrophe.
On the other hand you've got the immediate assurance of your current career being gone and high likelihood that you'll struggle financially as a result, with the upside that you can provide a warning to the public that they will almost certainly ignore and catastrophe comes anyway.
Given the risk/reward assessment of those two options, what do you pick?
Look what happened to many people who broke NDAs or helped to do so. Snowden is stuck in Russia. Assange got holed up in an embassy for years. Sarah Wynn-Williams, the Facebook/Meta whistleblower, got an insane court order against her [1].
Theoretically, retaliating against whistleblowers is seriously illegal stuff. But in practice these regulations are seldom enforced. And even if your (former) employer does not retaliate openly, there might still be "invisible" retaliation, or the risk of other large employers just blacklisting you by default.
If it's juicy enough presumably they could be an anonymous source for some journo. Which would still be breaking the NDA but good luck getting anywhere with that.
What if we displayed a modicum of truth and stopped pretending that Google employees don’t work there for the money? That is faithful and gracious to reality.
https://archive.li/fKI3D
Need more details. Any Googler willing to create a throwaway account and tell us what's going on?
Ok, I'll say it. The news is: "AI is not that powerful or functional". Sure, can change some little things but not that much. And we are overspending on it.
I get people who said things like that like year ago, but nowadays it's a laughable position. Solving millenium problem means nothing nowadays? Exponential increase in CVE is nothing?
The point is it's not AGI, and won't be for a long time.
>Exponential increase in CVE is nothing?
AI solving the problem it created? Woaow, the future is now!
I’m mostly bearish on AI, but this is flat wrong. The vast majority of CVEs that have been discovered with AI assistance were in code written long before AIs arrived on the scene.
truly surprised nobody here is claiming this is a marketing op. it's the same thing as jacob coxon, right?
from TFA, I liked seeing this, considering how influential Google HR policy choices are across the valley and beyond: “Recently, we’ve updated our terms for new hires, which included removing the post-termination aspect of non-disparagement clauses.”
Haven't we already heard the worst of it? If you let a swarm of agents on the loose with a goal, they will get to all kind of nefarious activity in pursuit of that goal. What could possibly be worse than that?
That the agent swarm recursively improves powerful enough and the extinction of all humans is a side effect of that goal.
If that’s the reality, you would think that employees would be willing to break an NDA for it. “We’re all going to die, but I really need my bonus this year” seems like a parody rather than a real strongly held position.
If anything this points to a more moderate (but still bad) worry among the employees.
> “We’re all going to die, but I really need my bonus this year” seems like a parody rather than a real strongly held position.
We have been living in a parody for some time now, so it feels normal.
“We’re all going to die, but I really need my bonus this year”
Is probably the best description of the state of engineering today
I suppose there may be a bit of a prisoner's dilemma to it. Single voices can be discredited. It would take a coordinated disclosure to actually make the message stick.
That's sci-fi. They cannot "recursively improve". There is no magic to LLMs. An agent swarm can increase it's thinking power by increasing the number of agents, the thinking time and by changing the base model to a more powerful one. So at worst they can assign themselves more compute or hijack another agent swarm based on a more powerful model.
Sorry bro, we live in a sci-fi world. I don't know what you think recursive improvement is, but we've been doing it for some time with models making synthetic data models and using it to train the next model. This is being done inside the labs intentionally to avoid as many employees and humans in post training. Works very well for things like math and cyber-sec.
Yes, works great for anything that is amenable to reinforcement learning. For other domains, not so much.
> I don't know what you think recursive improvement is
The ability for a model to improve itself over time without outside intervention. As yet, no models yet have this capability, and LLMs specifically what with how they are constructed, are not likely to ever have it. They are frozen in carbonite at the time of their training.
That's not to say they can't use skills to appear somewhat worldly, things like searching the internet and what have you, but if you ask a model to strictly stick to data in it's training, it will do that and it will be stuck in whatever time in which it was built.
Energy is conserved, gravity exists, and LLMs aren't intelligent.
LLMs today are improved by experiments that assay their performance, changes to architecture, pretraining, and post-training. None of those steps necessarily require humans. Welcome to your sci fi future.
> They cannot "recursively improve"
... says a replicator built by evolution.
So in the article it says : - a former employee did it with no consequences - executives said they won't enforce it
The article is just about them asking for changing the wording of contracts, not some super secret stuff they're dying to say and really can't.
Without the paywall: https://archive.is/fKI3D
saved me a click thanks
Gift link: https://wapo.st/4sFuF7P
They should quit and warn. Anything else is hypocritical.
They likely signed non-disclosure agreements, breaking one of those could be a career-ender.
If they’re not willing to break an NDA for it, it can’t be all that serious.
You'd be willing to destroy your career to publish a warning no one will need?
Not me. But these are all "most important people in the world." So should fit their own narrative.
This entire line of reasoning is just toxic. You represent a charicature of these people in your head and dole out self-righteous condemnation.
Just, Devil's Advocate, but what would be the benefit of your career if all humans (or even just a few billions of humans) are going to die?
I think it's reasonable to assume that if it's not worth breaking the NDA, it's likely not an existential threat for the vast majority of humanity.
Now, does that mean that people who currently work upper-middle class knowledge work jobs aren't in for an exceedingly harsh economic reality check? Absolutely not.
But that's not really as serious as an existential threat to humanity.
I doubt you are asking in good faith, because I suspect a good faith interlocutor would have worked this out on their own, but I'll proceed anyway.
On the one hand, you've potentially got an indeterminate number of future comfortable years based on a subjective assessment of how long you think it takes catastrophe to materialize, with the downside that you can't warn anyone and so no one will stop the coming catastrophe.
On the other hand you've got the immediate assurance of your current career being gone and high likelihood that you'll struggle financially as a result, with the upside that you can provide a warning to the public that they will almost certainly ignore and catastrophe comes anyway.
Given the risk/reward assessment of those two options, what do you pick?
You mean, no other evilcorp will hire them afterwards?
No anycorp, yes
Look what happened to many people who broke NDAs or helped to do so. Snowden is stuck in Russia. Assange got holed up in an embassy for years. Sarah Wynn-Williams, the Facebook/Meta whistleblower, got an insane court order against her [1].
Theoretically, retaliating against whistleblowers is seriously illegal stuff. But in practice these regulations are seldom enforced. And even if your (former) employer does not retaliate openly, there might still be "invisible" retaliation, or the risk of other large employers just blacklisting you by default.
[1] https://www.bbc.com/news/articles/cy42xgeq9vpo
If it's juicy enough presumably they could be an anonymous source for some journo. Which would still be breaking the NDA but good luck getting anywhere with that.
NDAs will be void in the future when people wrestles back control from the corporate elite controlling so-called democracies
But wait they don’t want to compromise their paycheck and opportunity to say they worked on the torment nexus. That would be uncool.
What if we displayed a modicum of good faith and grace
They work for Google.
What if we displayed a modicum of truth and stopped pretending that Google employees don’t work there for the money? That is faithful and gracious to reality.
Just can't do it, can you
How about:
1. Full disclosure. 2. Run for your life. 3. Live in exile.
That doesn't seem to have the best track record for leading to positive change
I mean, they can do that, it’s not like google has kidnapped them.
If they want to do that free of consequences, tough luck.
Exactly, the title should read Google researchers are looking for their spines to tell the world how dangerous AI is.
Where is the next Julian Assange? Where is the next Edward Snowden? You don’t demand freedom, you have freedom. What they don’t have his courage.
I think those two examples are a perfect demonstration of how futile it is to throw everything away to inform the public.