Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

GPT-2: Too Dangerous To Release (2019)

naokishibuya.github.io

Post reply on HN
Print view View original

41–50 of 139 posts

Re: GPT-2: Too Dangerous To Release (2019)

#41

Post by minimaxir »

oathvz wrote: post #22
minimaxir wrote: post #8

Back in 2019, it was more fair to have caution around the larger GPT-2 models since robust text generation (by 2019 standards) was a complete unknown. For something like Mythos in 2026, where now the social implications of better LLMs are more understood, it's more fair to call it (EDIT: specifically, the declaration of its danger) a marketing gimmick.

This is a natural follow up question -- what kind of an escalation or message should frontier labs/companies publish to be seen as genuine and not marketing gimmick?

It's fine for the labs to publish model safety cards and stagger releases/limit it to a narrow test group as they are already doing, but saying they're doing it "because the models could be dangerous" comes off as unnecessary as best.

Re: GPT-2: Too Dangerous To Release (2019)

#42

Post by Tenoke »

I've barely changed my mind on it. It was obviously premature at the time, but the right attitude because it's hard to tell which model is too dangerous in advance. If anything, I wish this rigor had evolved with the next releases but alas we no longer have the OpenAI of 2019.

Re: GPT-2: Too Dangerous To Release (2019)

#44

Post by breezybottom »

nonethewiser wrote: post #30
cjjfjjfjf wrote: post #6

In hindsight, they were entirely correct. The social damage caused by low cost content generation that’s hard to distinguish from human authorship is astronomical. You don’t need to entertain the more ridiculous doomsday scenarios to wish that this technology had never been created.

What is the astronomical social damage that this has caused? I am having so much trouble relating to and even understanding what the anti-AI crowd's position is. It looks like a caricature to me.

It's certainly accelerated the breakdown of trust. The US government has turned into an AI slopaganda shop. People don't know what to believe anymore when anything could be fake.

Re: GPT-2: Too Dangerous To Release (2019)

#45

Post by Qhemlomo »

panzi wrote: post #25
Qhemlomo wrote: post #19

Earlier quoted context omitted.

I don't want to stop progress just because its hard to imagine how it will transform our society. I want to see a Star Trek economy/society in my lifetime. I only life once. Btw. AI/LLM/Machine learning is the gateway technology for robotics, this will affect even more.

It's the wrong way around. If we get AGI (or any well working AI) before we abandon capitalism it's going to be a huge disaster. A handful of even richer even more powerful very greedy people will have all the wealth and everyone else will have nothing. I mean, there was a WW3 in Star Trek, so maybe it was that path that humanity took in Star Trek anyway?

If AI takes over to slowly, we might play the boiling game aka we don't realize that the water gets warmer and warmer until we boil.

But lets be honest, i don't know that, you don't either. But if a critical mass is reached, faster, we might need to actually solve this problem instead of migrating to a very dystopian future.

Stoping is not an option i think. Anthropic vs. OpenAI vs. Google <<< they ahve so much money and so much to loose. And then we have USA vs. China.

Re: GPT-2: Too Dangerous To Release (2019)

#46

Post by FabCH »

nonethewiser wrote: post #30
cjjfjjfjf wrote: post #6

In hindsight, they were entirely correct. The social damage caused by low cost content generation that’s hard to distinguish from human authorship is astronomical. You don’t need to entertain the more ridiculous doomsday scenarios to wish that this technology had never been created.

What is the astronomical social damage that this has caused? I am having so much trouble relating to and even understanding what the anti-AI crowd's position is. It looks like a caricature to me.

The president of the United States tweeted an AI generated image of himself as Jesus Christ descending from the sky and saving a sick person.

I feel like that is a good example. Now multiply that by hundreds of millions of AI generated propaganda images across the world.

And that’s even without touching the effect of fake videos on democracy or Elons pedo-bot that generates CSAM on demand of specific people…

Re: GPT-2: Too Dangerous To Release (2019)

#47

Post by smith7018 »

boelboel wrote: post #28
cjjfjjfjf wrote: post #6

In hindsight, they were entirely correct. The social damage caused by low cost content generation that’s hard to distinguish from human authorship is astronomical. You don’t need to entertain the more ridiculous doomsday scenarios to wish that this technology had never been created.

A lot of low cost content generation would've come regardless with something like 50% of the developing world getting access to mobile internet between 2018-2026 and social media incentivizing certain types of content (monetizing). But AI certainly didn't help.

Cheap labor has always been a thing; a random country getting more access to the internet doesn't change that. What's truly changed is velocity, quality, and quantity. Framing the pure firehose of slop targeting scientific research, used for nefarious political purposes, flooding social media, scamming people, and much more as something that "would've come regardless" without LLMs is disingenuous imo

Re: GPT-2: Too Dangerous To Release (2019)

#48

Post by stanmancan »

nonethewiser wrote: post #30
cjjfjjfjf wrote: post #6

In hindsight, they were entirely correct. The social damage caused by low cost content generation that’s hard to distinguish from human authorship is astronomical. You don’t need to entertain the more ridiculous doomsday scenarios to wish that this technology had never been created.

What is the astronomical social damage that this has caused? I am having so much trouble relating to and even understanding what the anti-AI crowd's position is. It looks like a caricature to me.

A large portion of the content on the internet is now generated by AI.

You can and do have full conversations with bots and not know. I want to interact with humans not LLMs.

There’s no way to combat it. An army of bots can post a specific rhetoric and it can and does sway people’s opinions.

The new version of Digg was shut down because they couldn’t find a way to combat AI. They were at least trying to, other platforms are just eating it up because “user activity” is a win for them.

Re: GPT-2: Too Dangerous To Release (2019)

#49

Post by enraged_camel »

minimaxir wrote: post #41
oathvz wrote: post #22

Earlier quoted context omitted.

This is a natural follow up question -- what kind of an escalation or message should frontier labs/companies publish to be seen as genuine and not marketing gimmick?

It's fine for the labs to publish model safety cards and stagger releases/limit it to a narrow test group as they are already doing, but saying they're doing it "because the models could be dangerous" comes off as unnecessary as best.

Unnecessary based on what exactly? Your vibes?

Re: GPT-2: Too Dangerous To Release (2019)

#50

Post by stanmancan »

Qhemlomo wrote: post #38
realusername wrote: post #31

Earlier quoted context omitted.

We still didn't "solve" software engineering, try to give Claude code access to your friends or family and see what they do with it.

My partner wrote an android app which was doing what she wanted to do. She did this experiment 5 month ago and she did this in one day. My wife has 0 knowledge how any of this works. That was shocking to see. Progress is not stoping and Fable proves that.

You can scaffold out a simple app pretty easily. Anything large or complex things break down. If you don’t know what you’re doing you end up leaking secrets like the dozens of examples we’ve seen so far.
Post reply on HN