Comment by dmix | Hacker News Reader

Comment by dmix | Hacker News Reader

Parent Full thread

It's interesting the paper is selling Anthropic's approach to 'safety' as the correct approach when they just launched a new version of Claude and HN thread is littered with people saying its unusable because half the prompts they type get flagged as ethical violations.

It's pretty clear some legitimate concerns about a hypothetical future AGI, that we've barely scraped the surface of, turns into "what can we do today" and it's largely virtue signalling type behaviour crippling a non-AGI very very alpha version of LLMs just to show you care about hypothetical future risks.

Even the coorelation between commercialization and AI safety is pretty tenuous. Unless I missed some good argument about how having a GPT store makes AGI destroying the world easier.

It can probably best be summarized as Helen Toner simply wants OpenAI to die for humanity's sake. Everything else is just minor detail.

> Over the weekend, Altman’s old executive team pushed the board to reinstate him—telling directors that their actions could trigger the company’s collapse.

> “That would actually be consistent with the mission,” replied board member Helen Toner, a director at a Washington policy research organization who joined the board two years ago.

https://www.wsj.com/tech/ai/altman-firing-openai-520a3a8c

What’s surprising to me is that top-level executives think that self-destructing the current leader in LLMs is the way to ensure safety.

Aren’t you simply making space for smaller, more aggressive, and less safety-minded competitors to grab a seat on the money train to do whatever they want to do?

Pandora’s box is already open. You have to guard it. You have to use your power and influence to enforce other competitors to also do that with their own boxes.

Self-destructing is the worst way to ensure AI safety.

Isn’t this just basic logic? Even chatGPT might have able to point out how stupid this is.

My only explanation is that something deeper happened that we’re not aware of. Us or them board fight might explain it. Great. Altman is out. Now what? Nobody predicted this would happen?

> Pandora’s box is already open. You have to guard it. You have to use your power and influence to enforce other competitors to also do that with their own boxes.

> Self-destructing is the worst way to ensure AI safety.

If they don't believe you're prepared to destroy the company if that's what it takes, then you have zero power or influence. If they try to "call your bluff", you have to go there, otherwise they'll never respect you again.

So the outcome of that is that you either destroy the company or you’re out? Either outcome is the same from your perspective, you chose to no longer have any say in what happens next

Sure, but you have a say in what happens right then and there.

Has Toner (or someone with like-minded views) filled in the blanks between "GPT-4" and "Terminator Judgement Day" in a believable way? I've read what Yudkowsky writes but it all sounds so fantastical that it's, at least to me, more like an episode of The Twilight Zone or The X-Files than something resembling reality. Is Toner really in the "nuke the datacenters" camp? If so, was her placement on the board not a mistake from the beginning?

I have a feeling EA[1] is like anarcho capitalist communities (or just inverse for socialism) where you have a good idea like markets==good then apply it like a hammer to everything you see in life. Any hint of gov/statism is treated as the enemy, regardless of its IRL negative impact on markets or net freedom, or whatever positive end goal they envision.

OpenAI is (hypothetically) progressing to AGI, therefore bad. Whether GPT specifically is a half a foot on a 1000km journey or potentially insignificant, it doesn't matter to them as they've pre-selected any potential for progress as all bad.

[1] if you search "Helen Toner" on YouTube it's just a collection of videos of her talking at EA conferences. So she's a believer

I think the middle ground where several of the OpenAI board members were trying for is to "responsibly develop AGI", which means developing at a moderate pace while trying to avoid kicking off an investment gold rush through making heavily commercial use cases, and spending a substantial amount of resources on promising safety research (such as Ilya's work).

In my opinion, it was not a very strong position because the allure of money and trying to be the biggest is too strong (as we're seeing now), but I think it was at least coherent.

So, no, the blanks have not been filled in then. Because, for those in Toner's camp, that middle ground is just "progress from GPT to T-1000, but slowly", right? Yudkowsky talks about AIs 3D printing themselves into biological entities and killing us all because humans are an inefficient use of matter. It's not a strong position to me because it sounds ridiculous, not because there's greed involved.

It takes some level of delusion of grandeur to think half a board of a single non profit that just happens to be the first mover can stop the full forces of American capitalism - although I kind of respect the drive/purpose.

They could easily lose any power they had to guide the industry, it was a huge gamble. I remember reading a Harvard business school study showing the first mover advantage repeatedly turned out to be ineffective in the tech industry as there is a looong series of early winners dying out to later market entrants like Friendster->FB, Google, a bunch of dotcom era Ecommerce companies predating Amazon, etc.

They need full industry/society buy in - at an ideological level - to win this battle, they won't win through backroom dealing in a boardroom while losing 90% of their own staff.

No idea if the concerns are realistic, but dismissing them because they are fantastical surely is a fallacy, no?

We currently almost have scifi-level AI we can talk to on our smartphones, which would have been fantastical only a few years ago.

The concerns are unrealistic because there are massive unusual and unvalidated assumptions in multiple levels, and lynchpin arguments that depend on utterly unsupported assumptions.

I don't know whether I have "like-minded views", but I think the risk is there, and it is not fantastical to me because the AGI doesn't need to "do" anything outside of humans assisting it -- no killer robots or nanobots. It just needs to persuade humans to kill for it, or gather resources for it, and so on. And because it's an AGI, it has superhuman abilities for persuasion.

The intuition for GPT-4 being on the way towards superhuman intelligence or persuasion abilities is that the only significant difference between it and GPT-3 appears to be the amount of compute power applied to training.

The previous US President was much less than superhuman, and he still managed to seriously threaten democracy in the US, and by extension cause a minor existential risk to civilization. Imagine what a more intelligent agent with enough attention to be personally persuasive instead of generically persuasive could achieve for itself, given a goal. (If you think that LLMs can't express goals or preferences, read some of the Bing Sydney transcripts.)

Exactly, AGI is overrated because the world isn’t run by its smartest people, ChatGPT would make a better US or Russian president at this point compared with treasonous or genocidal geriatrics

No one has, no.

In his recent Ted talk, Ilya talks about AI as 'brains in a computer', and explains why he believes they aren't that far off from AGI. [0] GPT4 is a substantially worse software engineer than I am, but it is 10x GPT3. And if we 10x for at most 2 or 3 iterations, it could be better than me. While it is unclear that we can get there, it doesn't seem implausible to me, and would have deep implications for human existence. That is already true if it just stayed at that level, but what most doomers are really worried about is the possibility of GPT8 being as good or a better AI researcher than most or all humans.

[0] https://www.youtube.com/watch?v=SEkGLj0bwAU

I feel like the “Terminator Judgement Day” scenario is a strawman more often than not. There are reasons to push for the safe and responsible development of AI that aren’t based on the belief that AI is literally going to launch the nukes.

The biggest risks I see from AI over the next 2 decades are:

1. A complete devolution in what is true (because creating convincing fakes will be trivial). From politics to the way that people interact with each other on a daily basis, this seems like it has the potential to sew division

2. the societal turmoil/upheaval arising from the mass unemployment (and subsequent unemployability) of millions of people without the social safety nets in place to handle that. When people are suddenly unable to work and support their families and those who have consolidated the power have no interest in sharing the wealth, there will inevitably be pushback.

> fill in the blanks

What exactly are you looking for here, proof of concept? This stuff is inherently speculative - because if it wasn't speculative, we'd probably be dead.

>how having a GPT store makes AGI destroying the world easier

The argument in general is that the more commercial interest there is in AI, the more money gets invested and the faster the companies will try to move to capture that market. This increases the risk for AGI by speeding up development due to competition, and safety is seen as "decel".

Helen was considering the possibility of Altman-dominated OpenAI that continued to rapidly grow the overall market for AI, and made a statement that perhaps destroying OpenAI would be better for the mission (safe development of AGI).

If true they should have never started in the first place because there was no way she was not going to press that button.

Reply on news.ycombinator.com