Live data from Hacker News

Sam and Greg's response to OpenAI Safety researcher claims

twitter.com

431–440 of 508 posts

Re: Sam and Greg's response to OpenAI Safety researcher claims

#431

Earlier quoted context omitted.

What are examples of immense powerful text writers these days?

Donald Trump and Fox News are maybe the canonical examples in modern times. Say what you like about Rupert Murdoch and Trump, they're incredibly effective communicators who have built fanatical religions based on words alone. And what we're about to see is that level of reach and intensity multiplied by millions.

I am not at all convinced that it is really down to the specific communication vs just seizing an opportunity in one of many possible ways.

If they were really super effective broadly, then Trump would have won any popular vote easily.

Re: Sam and Greg's response to OpenAI Safety researcher claims

#432

Earlier quoted context omitted.

> 1) AGI is not around the corner, so no worries I agree with this. I no longer agree that the "safety" people want the same things we do. I suspect after the bard debacle a bit back that "safety" looks a lot more like 1984. > He no longer cares ... I missed something Sam was gone, none of these people spoke up, and now they are leaving. I think the story is plain as day but we're pretending it isnt. Sam isnt a leade…

> I no longer agree that the "safety" people want the same things we do. I suspect after the bard debacle a bit back that "safety" looks a lot more like 1984. AI Safety will inevitably look a lot like DEI programs. They're trying to thread a needle with two fundamentally opposed ideas, and with a bit of time it turns into little more than a combination of Cover Your Ass policies. The ridiculous thing is that we actua…

I’m confused. How are AI safety and DEI opposed?

Re: Sam and Greg's response to OpenAI Safety researcher claims

#433

I agree with many other commenters here. The danger of AI is not an uncontrollable superhuman. The danger is much more close, down to earth and real. Misidentified speech when we apply LLMs in chat moderation. Wrong conclusions summarising 100-page military report. More subtle things like hallucinating some details in news, which then after many years will be taken as historical fact. The long tail of such consequenc…

How do we increase wages without this happening then? As humans become more expensive the machines become more efficient relatively speaking.

Even if these machines are less accurate than humans doing the same job, it's easier to audit them, they have repeatable and scalable behavior. We're still better off even if many people die from whatever warfare mistakes you can imagine because the humans would be doing something better instead.

It's just trade offs.

Re: Sam and Greg's response to OpenAI Safety researcher claims

#434

The talk of AI exterminating us via paperclips is an annoying one to me that detracts from the real dangers that are currently unfolding: - scammers are having the time of their lives, and their capabilities are only going to get better and easier - people are actively getting dumber as they trust the hallucinated outputs blindly - regular communication is starting to break down because you can't trust that people ar…

I sort of cover that in "SkyNet is not coming to kill you": https://renegadeotter.com/2024/04/22/artificial-intelligence...

AI can generate a lot of automated output, that is true, but my argument is that it's almost always mediocre at best.

There is a LOT to be said about the dangers of AI coming from the people running and funding it - not the AI itself. That stuff is complete bullshit.

Re: Sam and Greg's response to OpenAI Safety researcher claims

#435

Interesting that OpenAI has completely destroyed their reputation over the last year. For me, they've gone from an admirable company to cringe-worthy. I really think that the best outcome here is for OpenAI to collapse and get absorbed by Microsoft, where adults can continue some of the good work that is being overshadowed by ego and cult-like vibes.

> cult-like vibes. What cult-like vibes? The OpenAI employees wanted Sam back because of financial reasons and peer pressure. Sam has learnt how to keep employees on leash via money.

Having senior leadership and rank and file publicly proclaim their love for each other in identical ways, as happened during the leadership debacle, read cult-like to me.

Re: Sam and Greg's response to OpenAI Safety researcher claims

#436

Earlier quoted context omitted.

>> on the order of a couple years up to maybe 15 or 20 years Fusion, flying cars, AI in the 50's and in the 70's... Hell Hal was born in 97 and the film was in 68. For as impressive as LLM's are once you dig into them they aren't magical at all. A sophisticated model of language that predicts the next word is about as likely to become an AGI as whether predictions are likely to control the weather. Ask any expert in…

We don't have a clear, agreed upon definition of AGI, a way to test for it, or a way to predict whether a new model will meet those criteria. The definition OpenAI uses for AGI is being more economically valuable than most humans at most things. That definition is entirely backwards looking. They'll keep releasing new models and occasionally check in to see how each model compared to humans economically. If that isn'…

> The definition OpenAI uses for AGI is being more economically valuable than most humans at most things.

Then were already there. The web, amazon is already more valuable than sears and roebuck was as a paper catalog. All those people who took calls or opened envelopes, all the manual tracking of inventory and orders... we replaced all that years ago.

Re: Sam and Greg's response to OpenAI Safety researcher claims

#437

I don't think OpenAI needs to worry about superalignment before they can even properly define or test for AGI.

"Superalignment" is another bullshit industry term designed to make us believe we are dealing with "super intelligence".

"Model moderation" doesn't sound as sexy, does it?

Re: Sam and Greg's response to OpenAI Safety researcher claims

#438
post #410

Earlier quoted context omitted.

The most competent at that, did so, and are immensely powerful. That's basically all of politics, and not only in a democracy. You do have to be significantly above average to get there. Current generation LLMs are a bit above the human average, enough to be interesting, but probably not enough to be a personal Goebels for every would-be dictator.

Current LLMs are obviously far below the average adult human. Every demo I've ever seen brags about approaching the competency of an ordinary high school intern. Even if they reach above average our whole society is structured around managing billions of above average humans. What's novel?

> Current LLMs are obviously far below the average adult human. Every demo I've ever seen brags about approaching the competency of an ordinary high school intern.

There's a lot of bragging, sure. But what I've seen from the OpenAI models is more like a university level intern or a fresh graduate.

Not that it matters, as the point is to aim for where the ball is going rather than where it is now, but I'd say that this makes them above the average human performance (at least, within its domain).

> Even if they reach above average our whole society is structured around managing billions of above average humans.

That's putting a binary cutoff in an arbitrary place; half the population is always above average on any measure (or indeed topic), what I wrote was "significantly above average".

To crudely approximate with IQ, which is a bad measure for humans and much worse measure for AI, there's just over a billion humans of IQ > 115, around 182 million of IQ > 130, and around 8 million of IQ > 145.

I'm not sure how much to trust estimates of politician's IQs given the tribalism involved, but surveys of other groups tend to show that around half of the highest performing leaders (CEOs etc) with the most power, are in the top 1% by IQ.

(Who knows when, or if, someone will make an AI that's good at the important tasks within of each of those roles).

> What's novel?

Two things:

1) Breadth: the current vogue is very general models, so they're interns sure (even though I rank them as a higher level of intern than you), but they are interns at everything — there is no human you can hire with a even a mediocre grasp of Mandarin and Arabic and Welsh and python and CSS and calculus and magnetohydrodynamics and psychology and

2) Price: there is no human who can read a million tokens for a price equal to two days of the UN abject poverty threshold, or write a million tokens for six days of the UN abject poverty threshold.

You physically can't work that fast, even if you were typing [a] then [space] ten times a second for 18 hours a day; and you basically can't feed yourself if you earn less than that per day.

Re: Sam and Greg's response to OpenAI Safety researcher claims

#439

Earlier quoted context omitted.

Donald Trump and Fox News are maybe the canonical examples in modern times. Say what you like about Rupert Murdoch and Trump, they're incredibly effective communicators who have built fanatical religions based on words alone. And what we're about to see is that level of reach and intensity multiplied by millions.

I am not at all convinced that it is really down to the specific communication vs just seizing an opportunity in one of many possible ways. If they were really super effective broadly, then Trump would have won any popular vote easily.

> If they were really super effective broadly, then Trump would have won any popular vote easily

Trump is up against all the other equally effective communicators who are both allowed to and motivated to hold that specific office.

Re: Sam and Greg's response to OpenAI Safety researcher claims

#440
post #427

Earlier quoted context omitted.

Journalists, pundits, politicians, software developers, advertisers, lawyers, propagandists. Take your pick. Right now, the best LLMs are interns not experts at each of these things; I don't expect the transformer model to be good enough to reach the top level in each of these things, but then again I didn't expect anything like Stable Diffusion to happen before self driving cars.

Which ones there are very powerful? Other than perhaps political leaders (which need not be great writers), power is usually low.

I feel we may be disagreeing about one or both of "very" and "powerful", because otherwise that question makes no sense?

Some random member of the, say, Australian parliament, someone I've never heard of, is (relatively speaking) much more powerful than 99% of the planet.

A journalist who has an audience of millions, likewise.

A software developer whose work is economically valuable either through a broad audience or through being used to control something of great value, also powerful.

Post reply on HN