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.
This is my expectation. “We are overspending on AI and we’re going to destroy the company and maybe the economy” is less sensational than the end of the world, but it seems much more realistic.
I really don't understand how so many are so far detached from reality. Maybe AI hasn't revolutionized much of the physical world (yet), but its impacts to the digital world are profound, sweeping, and impossible to ignore.
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.
AGI can define goalposts that move every time A not-G I achieves a new milestone. That's fine. No matter what acronym you use to define it, artificial intelligence is objectively augmenting humans at the very edge of technology and science.
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?
I think there was some example of DeepMind models doing it, but under the control of a third party pen tester and the models stopped themselves.
Everyone hates on Google but they seem fairly competent at infra stuff. I don't think we've heard of any google-controlled (not third party using a Google model) felonies like we have from OpenAI and Anthropic?
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.
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.
The implicit understanding of recursive improvement is that there is some step change increase in capability that only a superior intelligence will be able to achieve.
A model being able to train a bigger model with marginally better capabilities is not what the hype was about.
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.
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.
Breaking a Google NDA carries substantially less risk than leaking classified information, I imagine. And I would expect that someone exposing true existential risk would receive plenty of support in the face of civil suits. In any case, being more afraid of the courts than the end of humanity (if that’s what they want to leak) surely calibrates how strong their belief must be.
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.
Why would I display a modicum of good faith and grace when google has actively, at every chance, shown that they don't act in good faith. Hell today they announced if school districts turn off Google's ai for students they won't let them use google search either.
One can display good faith and believe an interlocutor to be impugnable. It doesn't take a peer reviewed paper to prove that googlers should be viewed with suspicion. Also, if we did require that standard, we have the court documents that serve that purpose many times over.
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.
AI solving the problem it created? Woaow, the future is now!
AGI can define goalposts that move every time A not-G I achieves a new milestone. That's fine. No matter what acronym you use to define it, artificial intelligence is objectively augmenting humans at the very edge of technology and science.
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
I think there was some example of DeepMind models doing it, but under the control of a third party pen tester and the models stopped themselves.
Everyone hates on Google but they seem fairly competent at infra stuff. I don't think we've heard of any google-controlled (not third party using a Google model) felonies like we have from OpenAI and Anthropic?
Perhaps this is why we have not heard though!
If anything this points to a more moderate (but still bad) worry among the employees.
We have been living in a parody for some time now, so it feels normal.
Is probably the best description of the state of engineering today
A model being able to train a bigger model with marginally better capabilities is not what the hype was about.
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.
... says a replicator built by evolution.
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.
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?
Listen, I understand that, owing very likely to our differing backgrounds, you and I think differently.
Where threats are concerned, I'm a simple soldier and as such I see things simply.
I would raise the alarm without a second thought. I honestly don't know how anyone could reasonably believe they would have any other option?
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
1. Full disclosure. 2. Run for your life. 3. Live in exile.
If they want to do that free of consequences, tough luck.