GPT-2: Too Dangerous To Release (2019)

221 pointsposted 3 hours ago
by AbuAssar
(naokishibuya.github.io)

82 Comments

minimaxir

an hour ago

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.

Qhemlomo

an hour ago

How is this a gimmick?

It changes my whole profession on a level i couldn't even imagine how we would 'solve' software engineering.

malfist

9 minutes ago

> It changes my whole profession on a level i couldn't even imagine

I assure, it doesn't.

juleiie

41 minutes ago

For starters it makes you able to bypass having to go on Reddit to find incomplete trace of solution to some niche problem and acts as a sophisticated (but sometimes wrong) search engine. This already is worth every penny and improved my mental health immensely.

realusername

an hour ago

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.

Qhemlomo

an hour ago

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.

stanmancan

an hour ago

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.

Qhemlomo

43 minutes ago

You know what the problem is in software engineering? A LOT of people have no clue what good software engineering is.

I was working in a company before which used md5 in 2015! Databases on the internet with a 5 character password. No tests.

A person i know would have broken the whole production DB if i wouldn't have stoped the PR.

Another ex-collegue thought its okay to 'encrypt' with a basic shift cyper creditcard data.

I don't think any of these companies care that much

stanmancan

5 minutes ago

Yes the same applies to junior and inexperienced developers.

StableAlkyne

36 minutes ago

You could always do this, though.

Before gen code killed the freelance business model, there were hoards of people on Upwork/Fiverr willing to fuck other freelancers over and underpay themselves to make whatever barely-working slop you wanted.

Hell, before managers got the idea of AI layoffs, they had been off-shoring to low-quality code sweatshops for years. That was supposed to kill software engineering in the States 20 years ago. And it was just as frustrating (if not moreso) to get them to actually fulfill the project requirements.

realusername

16 minutes ago

I'd say creating a project is 5% of the job and maintaining it over time 95% of it.

It's true that they can start amazing projects without guidance but then the real work begins.

oathvz

an hour ago

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?

minimaxir

an hour ago

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.

aesthesia

17 minutes ago

One of the main purposes of model cards, from the beginning, has been to outline the ways that a model could be harmful or dangerous, and mitigations that can be or have been taken to reduce those risks. How do you expect labs to publish model cards without talking about this rationale?

enraged_camel

an hour ago

Unnecessary based on what exactly? Your vibes?

killerstorm

an hour ago

Yeah, I'm sure Anthropic loves people switching to Codex. Brilliant marketing.

jrflo

an hour ago

It's really interesting how back then no one was considering these tools for coding at all. Today, the hype around Mythos is mostly around security vulnerabilities, while in the original GPT 2 post they don't mention coding once. The "danger" was probably spam content and mis-information.

minimaxir

37 minutes ago

Even if the ReAct paper was published in 2019, I don't think GPT-2 was robust enough to actually work with a tool-calling approach even when finetuned.

For regular coding, GPT-2 was effectively useless because it was only trained from links posted on Reddit.

ffsm8

44 minutes ago

The agentic loop wasn't really established back then either, as tool calling came much much later... So yeah, not just probably - rather most definitely.

a2128

12 minutes ago

Yet it's 2026 and we see extreme examples of spam content and misinformation to the point that it's killing the internet, but AI companies have collectively decided to not care

MostlyStable

an hour ago

Unfortunately, Anthropic and Claude models have joined the ranks of Mind-killer topics where the signal to noise ratio in any discussion has dropped through the basement.

Tenoke

an hour ago

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.

lnenad

an hour ago

Feels like a hundred years ago.

kkkqkqkqkqlqlql

20 minutes ago

What's next, a virus from China? Such a fuss for nothing.

cjjfjjfjf

an hour ago

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.

qurren

an hour ago

On the other hand, maybe it makes people just get off the internet and value in-person interactions more.

I've stopped scrolling social media and tired of seeing fake landscapes, fake foods, and fake cities that don't exist.

larodi

an hour ago

Indeed, people seem to try to engage more around me. May be generational, but it can definitely be felt. The internet of algorithmic media may experience a downfall nobody saw coming.

applfanboysbgon

23 minutes ago

I use the internet because I enjoy seeing what the best of humanity, globally, has to offer. There are millions of incredibly skilled individuals in the world - artists, musicians, developers, and so on - and I had access to all of them at my fingertips, both for entertainment and learning to develop my own skills. That is now being drowned out, with generated content being produced at 100x or 1000x the rate of human content. "Hurrdurr it's good if the internet is destroyed because I have no self control and needed to be incentivized to touch grass anyways" is such a lowbrow pseudo-contrarian-intellectual take.

boelboel

an hour ago

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.

whstl

41 minutes ago

Yep. And there were previews of that 10+ years ago already with content farms and SEO-spam.

throwaway85825

36 minutes ago

There's significant overlap between the smartest bots and dumbest humans. Internet platforms have a negative incentive to encourage quality content. Google embraced the spam and scams decades ago.

smith7018

an hour ago

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

boelboel

an hour ago

You're right, I should've been more accurate and said a significant portion of the enshitification of the internet would've happened regardless. The effects on education is probably a lot worse.

I shouldn't have targeted the developing world as much as the incentives made by social media platforms needing to get growth in other ways than usercount.

nonethewiser

an hour ago

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.

FabCH

an hour ago

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…

hk__2

an hour ago

> 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.

I feel like this is the worst example, actually, because here it’s 100% clear to anyone that it’s AI-generated content. The danger is more about AI-generated fake images/videos disguised as real content.

tempestn

an hour ago

I honestly feel like that's a counter-example. With AI he'd be tweeting some other nonsense. It's not like anyone saw the image and thought he actually was orange Jesus.

FabCH

12 minutes ago

Multiple people have the same response, I randomly selected this one for follow up:

Yes, but two things were lost:

1) the need for skill or an accomplice. He _couldn’t_ tweet that image in 2016, not without first asking someone to photoshop it. And that need to engage in human to human communication is something truly fundamental that was changed and lost.

2) Any ambiguity or misunderstanding. Yes bad textual tweets exist for a long time in politics. But there IS something about images that is more powerful than text. The text „I’m Jesus Christ and god sent me to heal the sick“ would probably make the news, but a lot of people would go: „is he quoting the bible? What’s going on?“, not so much with Jesus Picture.

ge96

an hour ago

Hard to beat shrimp jesus

nonethewiser

38 minutes ago

> 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.

Does that say anything about AI or everything about Donald Trump?

PhunkyPhil

an hour ago

School is almost a joke now. The fraction of students who have a propensity to cheat now has increased, and the accuracy of the cheated material is so good teachers/professors can't or don't have the resources to properly address it.

_aavaa_

42 minutes ago

This is not a great example depending on how you frame this.

The solution to the cheating is, as has always been, to have tests conducted in person, on paper without digital technology, under strict supervision.

breezybottom

an hour ago

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.

stanmancan

an hour ago

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.

legitster

44 minutes ago

The sloppification of the internet began before AI. Google was SEOing the open internet to death, Reddit had fully baked in a hivemind, and social media became dominated by professional influencers.

AI is accelerating but also perhaps backfilling in what was already being lost.

throwaway85825

34 minutes ago

AI is the same slop but cheaper. Ideally the value of slop approaches zero but the value of quality stays the same.

nonethewiser

37 minutes ago

But what is the social damage? Can you quantify the damage, even roughly?

stanmancan

3 minutes ago

It's likely nearly impossible to evaluate that in the short term; I think we're looking at generational damage, much of which won't be apparent for years to come.

witx

24 minutes ago

How easy it is now to forge data (video, images, etc) will rott society. Cheating for students is now so much easier. There many examples.

Is it really that hard to understand?

Qhemlomo

an hour ago

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.

pixl97

an hour ago

>I want to see a Star Trek economy/society in my lifetime. I only life once.

While Star Trek is fiction, it's probably a good idea to understand the history of how the ST utopia came about, at the cost of a third of the worlds population and decades of suffering.

https://memory-alpha.fandom.com/wiki/World_War_III

tintor

an hour ago

"World War III was the last of Earth's three world wars, lasting from approximately 2026 to 2053."

panzi

an hour ago

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?

thewebguyd

an hour ago

> so maybe it was that path that humanity took in Star Trek anyway?

It was (aside from first contact, and the subsequent development of the replicator which enabled the post scarcity economy). The federation was built from the war, not after it.

Suffering is what made the utopia possible, and if ever get to the point of nearing a post scarcity economy, we are likely to experience the same. Progress is built on catastrophe. Whether or not you call it progress depends on if you are born later after the catastrophe and can look back and call it progress, or if you lived through the suffering without seeing the end result.

Qhemlomo

an hour ago

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.

root-parent

an hour ago

Countries without Internet access will see their population IQ explode.

Macha

an hour ago

Of course, this damage could still be enabled with just hosted access to the models, restricting access to the model files themselves did not stop that

DaveZale

an hour ago

I certainly cannot survive much more of the AI memes generated about our so-called Commander in Chief with a fake bodybuilder mystique... you are absolutely correct, this kind of material is psychologically damaging. And a huge distraction from the genocide by the "best friend and ally" of the US. Heart wrenching and extremely damaging hasbara - just please stop, haven't you stolen and killed enough guys? This is _not_ the old American West when communications were few and it was most often a tale of solitary survival. It's organized Nazi-esque kill, command and control, enable by so-called AI to take some guilt off the shoulders of those pushing buttons and pulling triggers.

woah

an hour ago

Lol I can't tell if this is sarcastic or not

stackghost

an hour ago

>In hindsight, they were entirely correct.

Sure but when serial grifter Sam Altman said it was "too dangerous" what he meant was that he wanted regulators to create him an artificial competitive moat so Anthropic et al couldn't catch up.

Serial grifter Sam Altman does not care about anything but making money, and certainly doesn't care about ethics. That's why serial grifter Sam Altman's company trained its models on pirated textbooks and copyrighted works without paying. Rules for thee but not for me.

Serial grifter Sam Altman doesn't care if society unravels because he is so rich that laws and consequences do not apply to him.

Jzush

an hour ago

I believe it was a marketing strategy.

zkmon

41 minutes ago

After people get tired of the "too dangerous to release" punchline, they might come up with "too big to fail". Oh, wait that's already invented in 2008.

wg0

an hour ago

Hilarious. Imagine the same about Claude coming back from 2036.

HALtheWise

an hour ago

Say hypothetically that they were concerned that GPT models would see widespread abuse, for example by students cheating on homework assignments, in a way that could cause likely-irreversible societal changes some of which are harmful. Can we confidently say they were wrong?

minimaxir

an hour ago

The dangerous use cases back in 2019 were spam and phishing and GPT-2 1.5B was nowhere near good enough to do homework assignments. No one envisioned how LLMs would develop.

Zambyte

an hour ago

GPT-2: Too Dangerous To Release (2019) (2022)

jansan

an hour ago

Same vibes:

In 2000 Sony "declared that the company’s PlayStation2 has been hit with export restriction because it could be used for military purposes"

"Trade officials said they initially placed restrictions on the game console because PlayStation2’s high-speed graphic processing could be used for missile guidance."

[1] https://variety.com/2000/biz/news/playstation2-export-regs-e...

throwaway85825

32 minutes ago

And the same with PS3 and again with PowerPC G5 mac.

functionmouse

an hour ago

they knew the risks and went ahead anyways, making them LIABLE for the damages that followed.

catigula

an hour ago

I don't know about anyone else, but LLMs certainly significantly negatively impacted my life overall and contributed to a loss of hope in the future.

throwaw12

an hour ago

> Due to our concerns about malicious applications of the technology, we are not releasing the trained model.

They were not wrong, indeed whole industries are running on this technology maliciously now, because of which RAM, disk prices increased a lot.

    - RAM, GPU, Disk prices are up
    - Slop became the norm
    - people are writing documents with AI, reading with AI, responding with AI
    - students are doing homeworks with AI
    - interviewees are using AI to cheat
    - people are mass emailing with AI
    - tiktok, instagram, youtube got even more non-sense videos
    - and many more...

ThejaCH

an hour ago

Fable/Mythos: Hello World!

wongarsu

an hour ago

Different company. And doing security reviews with Fable is pretty annoying, it loves to downgrade to Opus

GPT-5.5 seems more dangerous in those regards

rfoo

an hour ago

> Different company

Same people.

ThejaCH

an hour ago

Lol, funny but makes me uneasy

EA-3167

an hour ago

Seven years of this insufferable brand of "Oh it's so dangerous, I sure hope no one gives us a ton of money and takes us seriously" marketing and people are still falling for it at scale.

Terr_

an hour ago

Every night I am wracked by grief and anxiety that we might deliver too much value to our investors and shareholders. If only someone would create legislation that would mildly inconvenience us while crippling potential competitors!

Qhemlomo

an hour ago

They feared that GPT-2 could break all Spam filters.

And tbh do you prefer companies not taking anything serious?

Opus 4.5 def changed a lot already, GenAI changed a lot.

Certain jobs are gone. Do you think the person who was translating text doesn't deserve to be taken serious?

I haven't written code in a few month now and the quality of these coding agents is not getting worse, they are getting better.

All of this is transformable and we just started. GPT-3 came out in 2020 and public got access to it only 2022.

The last 4 years do not feal like 4 years and we are still progressing.

We have to also ask us as a society what is happening to young people. Even if we accept that we still hire juniors, they themselves have to completly rethink how they learn and how they work.

uselessTA

an hour ago

The concern I heard was that releasing it would start an arms race for AGI, which I think it clearly did

minimaxir

an hour ago

GPT-2 did not start the LLM arms race. GPT-3's release didn't either.