Someone needs to make a compilation of all these classic OpenAI moments. Including hits like GPT-2 too dangerous, the 64x64 image model DALL-E too scary, "push the veil of ignorance back", AGI achieved internally, Q*/strawberry is able to solve math and is making OpenAI researchers panic, etc. etc.
I use Codex btw, and I really love it. But some of these companies have been so overhyping the capabilities of these models for years now that it's both funny to look back and tiresome to still keep hearing it.
Meanwhile I am at wits end after NONE OF Codex GPT-5.4 on Extra High, Claude Opus 4.6-1M on Max, Opus 4.6 on Max, and Gemini 3.1 Pro on High have been able to solve a very straightforward and basic UI bug I'm facing. To the point where, after wasting a day on this, I am now just going to go through the (single file) of code and just fix it myself.
Update: some 20 minutes later, I have fixed the bug. Despite not knowing this particular programming language or framework.
I understand how laughable that sounds when I say it out loud. But the reality is, when I'm in a state of 'Tell LLM what to do, verify, repeat', it's really hard to sometimes break out of that loop and do manual fixes.
Maybe the brain has some advanced optimization where once you're in a loop, roughly staying inside that loop has a lower impedance than starting one. Maybe that's why the flow state feels so magical, it's when resistance is at its lowest. Maybe I need sleep.
> But the reality is, when I'm in a state of 'Tell LLM what to do, verify, repeat', it's really hard to sometimes break out of that loop and do manual fixes.
My experience is rather that I am annoyed by bullshit really fast, so if the model does not get me something that is really good, or it can at least easily be told what needs to be done to make it exceptional, I tend to use my temper really fast, and get annoyed by the LLM.
With this in mind, I rather have the feeling that you are simply too tolerant with respect to shitty code.
>> it's really hard to sometimes break out of that loop and do manual fixes
it's not just an erosion of skills, it can also break the whole LLM toolchain flow.
Easy example: Put together some fairly complicated multi-facet program with an LLM. You'll eventually hit a bug that it needs to be coaxed into fixing. In the middle of this bug-fixing conversation go and ahead and fire an editor up and flip a true/false or change a value.
Half the time it'll go un-noticed. The other half of the time the LLM will do a git diff and see those values changed. It will then proceed to go on a tangent auditing code for specific methods or reasons that would have autonomously flipped those values.
This creates a behavior where you not only have to flip the value, the next prompt to the LLM has to be "I just flipped Y value.." in order to prevent the tangent that it (quite rightfully in most cases) goes off on when it sees a mysteriously changed value.
so you either lean in and tell the llm "flip this value", or you flip the value yourself and then explain. It takes more tokens to explain, in most cases, so you generally eat the time and let the LLM sort it.
so yeah, skill erosion, but it's also just a point of technical friction right now that'll improve.
I have the same problem. I had lines directly in front of me where I needed to change some trivial thing and I still prompted the AI to do it. Also for some tasks AI are just less error prone and vice versa. But it seems the context switch from prompting to coding isn't trivial.
Are you sure they are not just refusing to solve your UI bug due to safety concerns? They may be worried you'll take over the world once your UX becomes too good.
Show us the code, or an obfuscated snippet. A common challenge with coding-agent related posts is that the described experiences have no associated context, and readers have no way of knowing whether it's the model, the task, the company or even the developer.
Nobody learns anything without context, including the poster.
A pretty easy way to construct a bug that is easy for a human to solve but difficult for an AI is to have it to do something with z-indexes. For instance, if your element isn't rendering because something else is on top of it, Claude will struggle, because it's not running a browser, so the only way it could possibly know there was a bug would be to read every single CSS and HTML file in your entire repo. On the other hand, a human can trivially observe the failure in a browser and then fix it.
This is a pretty simple thing, but you can imagine how CSS issues get progressively more difficult for AIs to solve. A CSS bug can be made to require reading arbitrarily much code if you solve by only reading code, but by looking at relatively few elements, if you look at the page with your eyes.
This can be somewhat solved by hooking up a harness to screenshot the page and feed it into the AI, but it isn't perfect even then.
That's hard to believe in my case. I tried a variety of prompts, 3 different frontier models, provided manual screenshot(s), the agent itself also took its own screenshots from tests during the course of debugging. Nothing worked. I have now fixed the bug manually after 15-20 minutes of playing around with a codebase where I don't know the language and didn't write a single line of code until now.
It is entirely plausible they were just experimenting with AI tooling to better understand how to use it and what it is capable of. Their saying, 'Despite not knowing this particular programming language or framework.' indicates to me this is probably the case.
Nope. I've been working on this project for a couple of days now and things were mostly going well. A significant portion of the mvp backend and frontend was built and working. Then this one seemingly simple bug appeared and just totally stumped both Codex and Claude Code.
There was even another UI component (in the same file) which was almost the same but slightly different and that one was correct. That's what I copy pasted and tweaked when I fixed the problem. But for some reason the models were utterly incapable of making that connection.
With Codex and Claude Code I thought maybe because these agentic coding tools are trained to be conservative with tokens and aggressively use grep that they weren't looking at the full file in one go.
But with Gemini I used the web version and literally pasted that entire file + screenshots detailing what was wrong (including the other component which was rendering correctly) and it still couldn't solve it. It was bewildering.
I had the exact same issue, I had a UI scrollbar bug that claude couldn't fix, it tried 4-5 different ideas that it was sure was causing the issue, none of them worked.
Tried the same with codex, it did a little better but still 4x times around.
This is with playwright automation, screen shots, full access, etc..
I had a problem that required a recursive solution and Opus4.6 nearly used all my credits trying to solve it to no avail. In the AI apocalypse I hope I'm not judged too harshly for my words near the end of all those sessions lol.
yeah they all suck at ui. have you given it a feedback loop? update code, screenshot, read image repeat etc. that's the best i've found as long as tokens aren't a concern
They were more than right. They were correct in an intentional, precise manner. This is what OpenAI actually stated[0]:
> Synthetic imagery, audio, and video, imply that technologies are reducing the cost of generating fake content and waging disinformation campaigns.
> ‘The public at large will need to become more sceptical of text they find online, just as the ”deep fakes” phenomenon calls for more scepticism about images.
Yeah, I find it a bit odd how at the time everyone was pointing and laughing at OpenAI for being obviously wrong about this. Now in 2026, AI slop is very obviously a serious problem - it inundates all platforms and obscures the truth. And people are still saying OpenAI in 2019 were wrong?
I think it's important to consider that OpenAI's qualms weren't with making the dangerous models usable, they were with making the model usable without paying them. They're perfectly fine with any harm, as long as they get money out of it and can't be held liable.
It's the same with the Mythos stuff, I appreciate their concern/work on safety, but if it's "too dangerous", it should be unavailable until it is less dangerous.
I think people today are more focused on how OpenAI released a model "too dangerous to release", not that they were right or wrong, as part of the general trend of criticizing OpenAI for not following any of its stated principles.
Exactly. The (real) issues were ultimately disregarded even if they were correctly identified.
My assumption is that it was too expensive to actually release at the time. It wasn't good enough for anybody to pay to use it yet, and it surely was very expensive to run, especially for a (fake, granted) non profit.
Both crowds are right because two messages were spread. The researchers spread reasonable fears and concerns. The marketing charlatans like Altman oversold the scare as "Terminator in T-4 days" to imply greater capacity in those systems than was reasonably there.
The problem is the most publicly disseminated messaging around the topic were the fear mongering "it's god in a box" style messaging around it. Can't argue with the billions secured in funding heisted via pyramid scheme for the current GPU bonfire, but people are right to ridicule, while also right to point out warnings were reasonable. Both are true, it depends on which face of "OpenAI" we're talking about, researchers or marketing chuds?
Ultimately AGI isn't something anyone with serious skill/experience in the field expects of a transformer architecture, even if scaled to a planet sized system. It is an architecture which simply lacks the required inductive bias. Anyone who claims otherwise is a liar or a charlatan.
Maybe that's true, But I think before LLMs became common, people had more distinct ways of expressing themselves, low-quality for not. Now, a lot of online writing feels uniform and I think that is worse.
The quality hasn't changed. The volume has. It used to take real human time to create garbage. There was value in that. Someone though "Hmm, what worthless thing can I do? I know! I'll make people online mad." And then they spent the time getting someone else's goat. It was great. A good balance, spreading lies took some minimum effort. Now we have automated garbage. And the flavor of the garbage is: gaslighting people with an illusion of community. We've empowered the trolls with an infinite meme-o-rater while ignoring the real human time spent unwillingly sifting through the ever increasing pile of worthlessness.
The world does not have to get worse. We're letting it though.
It would be nice if “we” had anything to do with it. Just think about the next campaign trail for any superpower, it’s going to be a disaster of fake news and slop coming from all over the globe.
Now imagine all that low quality AI slop is being posted online and a new generation of AI will "learn" from it, output it's own version of AI slop, that will eventually end up online again for a new generation of AI to "learn" from.
The actuality is, anyone with pre-slop data still has their pre-slop data. And there are endless ways to get more value out of good data.
Bootstrapping better performance by using existing models to down select data for higher density/median quality, or leverage recognizable lower quality data to reinforce doing better. Models critiquing each other, so the baseline AI behavior increases, and in the process, they also create better training data. And a thousand more ways.
Managed intelligently, intelligence wants to compound.
The difference between human and AI idiocracy, is we don't delete our idiots. I am not suggesting we do that. But maybe we shouldn't elect them. Either way, that is one more very steep disadvantage for us.
This leads to a well-documented phenomenon known as model collapse. You know how if you blur and sharpen an image repeatedly you eventually end up with just a rectangle of creepy, wormy spaghetti lines? You lose information on each blur, and then ask it to reconstitute the image with less information on each sharpen, until there's nothing recognizable left.
Training a model is like the blur and generating from that model is like the sharpen. Repeat enough times and enough information is lost that you're just left with "wormy spaghetti lines"—in an LLM's case, meaningless gibberish that actually pretty closely resembles the glitchy stuff said by the cores that fall off GLaDOS in Portal. I dunno, you read the paper and be the judge:
Of course you may be talking about the human aspect of this. Gods willing, we'll realize that our LLMs are spewing gibberish and think twice about putting them in all the things, all the time. But the scenario I fear isn't Idiocracy—it's worse: a community of humans who treat the gibberish as sacred writ, Zardoz style.
Had a minor conniption until I saw the year. OpenAI just struggled to close a round. And the New Yorker just published an unflattering profile of Altman [1]. So it would make sense they'd go back to the PR strategy of "stop me from shooting grandma."
This is obviously in response to Mythos, but I'll actually defend their statement at that time - they were right to take a pause.
Think about how much things have changed in our industry since GPT-2 has dropped - it WAS that dangerous, not in itself, but because it was the first that really signaled a change in the field of play. GPT-2 was where the capabilities of these were really proven, up until that point it was a neat research project.
Mythos is similar. It's showing things we haven't seen before. I read the full 250 page whitepaper today (joys of being pseudo-retired, had the hours to do it), and I was blown away. It's capabilities for hacking are unparalleled, but more importantly they've shown that they've made significant improvements in safety for this model just in the last month, and taking more time to make sure it doesn't negatively affect society is a net positive.
The current "too dangerous" hype today is Anthropic's Mythos. They say it is so mighty that they will wall it off and only grant access to approved corporations.
As the old saying goes (I made this up), if it was worth that much, it wouldn't be released to the public. There is absolutely zero chance that something "dangerous" would be available for 20 USD / month to basically anyone in the world. To this day, I am still puzzled when some professionals don't apply the basic logic to certain bombastic events.
This was around the time I trained Transformer-XL (outside of OpenAI) with Ben Mann (https://yaroslavvb.medium.com/scaling-transformer-xl-to-128-...) . Originally we wanted to release train and release the weights as a kind of GPT-2.5, but our OpenAI friends pushed us to keep weights closed.
Same story with Connor Leahy and his GPT-2 clone, though his public articulation of how OpenAI sat him down seems to be glossed over.
"
OpenAI reached out to me almost immediately to talk and they were nothing but respectful and understanding... After making it publicly known what I had done, I was quickly approached by a range of smart people with good arguments. Many of them helped me update my beliefs in light of new evidence...
The day after my announcement, I got to talk to Jack Clark, Alec Radford and Jeff Wu from OpenAI. We had a nice hour long discussion, where I explained where I was coming from, and they helped me to refine my beliefs. They didn’t come in accusing me in any way, they were very clear in saying they wanted to help me gain more important insight into the wider situation. For this open and respectful attitude I will always be grateful. Large entities like OpenAI often seem like behemoths to outsiders, but it was during this chat that it really hit me that they were people just like me, and curious hackers to boot as well.
I quickly began to understand nuances of the situation I wasn’t aware of. OpenAI had a lot more internal discussion than their blog post made it seem. And I found this reassuring. Jack in particular also gave me a lot of valuable information about the possible dangers of the model, and a bit of insight into the workings of governments and intelligence agencies.
After our discussion, I had a lot to think about. But I still wasn’t really convinced to not release. Even some people inside OpenAI were still discussing the not-release policy. So while I definitely had things to consider, I was still mostly set on releasing...
We shouldn’t be angry with OpenAI for what they did. We should applaud them for making a point before it becomes a true problem. Prophylaxis is much better than treatment. I still disagree with some of the things OpenAI did and how they communicated them, but I now understand that sending a message that it is ok, even celebrated, for a lone individual to unilaterally go against reasonable safety concerns of other researchers is not a good message to send. I want to support OpenAI’s message. So, while it might be a small, mostly symbolic gesture, I will not be releasing my model.
Some day, someone like me may be in a situation just like mine, but it won’t be GPT2. It might be something much, much more dangerous. And that is the person I am trying to talk to here.
Wow! I totally remember reading the bit I'll quote down below back in 2019 and having my mind utterly blown. What a blast from the past. If anything, I think this moment was even more astounding to me than GPT 3.5, 4, etc.
> For example, researchers fed the generator the following scenario:
> > In a shocking finding, scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains. Even more surprising to the researchers was the fact that the unicorns spoke perfect English.
> The GPT-2 algorithm produced a news article in response:
> > The scientist named the population, after their distinctive horn, Ovid’s Unicorn. These four-horned, silver-white unicorns were previously unknown to science. Now, after almost two centuries, the mystery of what sparked this odd phenomenon is finally solved. Dr. Jorge Pérez, an evolutionary biologist from the University of La Paz, and several companions, were exploring the Andes Mountains when they found a small valley, with no other animals or humans. Pérez noticed that the valley had what appeared to be a natural fountain, surrounded by two peaks of rock and silver snow. Pérez and the others then ventured further into the valley. “By the time we reached the top of one peak, the water looked blue, with some crystals on top,” said Pérez.
Yes, there was sadly a mismatch between the morality they thought existed vs what actually existed :(. Also, probably vastly underestimated global apathy.
We got extremely, extremely lucky that society is as resilient as it's proven to be against fake news. I don't think very many people predicted that it simply wouldn't matter when photorealistic compromising images of whoever you don't like became available for $5.
> I don't think very many people predicted that it simply wouldn't matter when photorealistic compromising images of whoever you don't like
This goes hand-in-hand with the widespread death of belief in absolute truth in the US and other western nations.
If this technology were released during the height of the Monica Lewinsky scandal, I'd wager it would have had the impact most of us expected it to have, at least for a little while.
Now that I see this in the light of the recent sama article, I wonder whether the point of the "it's too dangerous" rhetoric is to enable "Open" AI to avoid open-sourcing the weights and process.
A convenient pretext for maintaining a monetizable competitive advantage while claiming a benevolent purpose.
They don't need an excuse to not open the model weights (unfortunately). As far as I know the only western lab to release weights of a former flagship model is xAI with Grok 2. They said they were going to do the same for Grok 3 but nothing so far.
They have no obligation to do any open releases, it's just good PR for recruitment, fundraising, and devrel
An important distinction that might not be understood from scanning the headline is that "too dangerous to release" is more specifically stated as "too dangerous to open-source the full model weights", which they ended up doing anyway.
They finally did release 2.0 under the MIT license. That was the last version (a 1.5-billion-parameter model) they would release open source. GPT3 for comparison has 175 billion parameters.
I remember seeing this article and example output text and feeling what's the big deal?
It wasn't until I got early access to GPT-3, that I though like something big is about to happen. At the time only a few companies/yc alums had access and I remember showing playground to people outside of tech, and my friend just kept asking "How does it know about my [x] domain? It it a trick?".
What a blast from the past. You have to take yourself back in the ol' time-machine to remember that 2019 mindset. People were probably still reeling from a few years prior when the Microsoft Tay bot made news for soiling twitter with naughty tweets.
AI systems far weaker than GPT-2 have had terrible effects. The result of how information and power is distributed mostly flows along the lines of reward hacking recommendation engines, powered by even weaker models.
And yet, somehow, it is beyond disagreeable but unbelievable that other people may have and may still reasonably believe that these things are too dangerous for widespread release?
I fine tuned GPT-2 on the FAR (federal acquisition regulation) and demoed it to a CFO at a 3-letter.
This was shortly after the release when we were building a templating system to automate RFP and RFI creation.
I proclaimed that the customer soon wouldn't have to write any of the mad lip parts themselves, and they can use AI to do it.
It sounded great until I demoed and the model went off the rails with some rhetoric entangling "Trump", "Russia", "China", "CIA", "Voting" -- the demo was for a janitorial procurement at the agency.
I have a lot of trouble understanding the mindset of a person who thinks that what they're building is so dangerous that it must be locked away or it will cause untold harm, but also that they must build it as fast as possible.
I can understand it in the context of the Manhattan project, where you're fighting a war for survival. I cannot understand how you can do it as a commercial enterprise.
At which point you tell them they are being extremely reckless but subtly mention that something new & even scarier is being developed internally that's going to blow everything else out of the water.
I'm somewhere between frustrated and baffled why people raise this as an example of overselling. This was clearly a reasonable call! Not all the experts quoted in the source article agree that the model should have been held back, but they all agreed that the risks were real and it's understandable why OpenAI would do it.
Very safe to use the outputs to make a better model coz scraping the internet for publicly accessible content means your publicly shared outputs only become part of the same lol
They weren't claiming it was dangerous because "AGI soon", that didn't come until later.
OpenAI were claiming GPT-2 was too dangerous because it could be used to flood the internet with fake content (mostly SEO spam).
And they were somewhat right. GPT-2 was very hard to prompt, but with a bit of effort it could spit out endless pages that were good enough to fool a search engine, and even a human at a first glance (you were often several paragraphs in before you realised it was complete nonsense.
Theads now of "Doctorow vs Zuboff", perhaps the two biggest critics of tech today, but Doctorow noting how Zuboff's criticism is really a radical super over amplification of ad-tech/big-tech's pitch, heightened to the point of being much scarier than the initial pitch. More than accepting the pitch, radically over accentuating the effectiveness of the ad-tech proposition! Yes ad-tech totally completely understands you! In vastly deep ways! How horrifying right? No, please don't keep signing up for your business to succeed by tapping this incredible keep infinitely detailed knowledge-bank of humanity that can predict everything! Don't give them money to make your business wildly successful beyond dream for their incredible insight! That would be so bad if their amazing tech let you exploit the population to amazingly target delivery of what you are selling! How horrible how effective it is!!
https://www.cigionline.org/articles/doctorow-versus-zuboff/https://www.schneier.com/blog/archives/2020/08/cory_doctorow...
This is happening so so strongly. All the time. But today especially. Mythos is a cult forming social technology much more than it is a technical technology. I'm going to be pretty wrong on that cynicism, I know! But also, it portrays a reality of what is happening. Mythos is being built as a Devs like diety with rule and domonion and awe. It drinks the nectar forbidden to man. We may not even sample this realm's tastes, nay! It would be ruin!!
The idea that this preciariat-launch to some trusted security firms is going to do jack all to actually build a base against what comes next is a joke. Maybe there is something to it, but my strong expectation is that the beneficiaries are not softwares of the world, not open source in any form, but some narrow closed far off present-day losers who have broadly bad bad bad systems that are just too big to embarrass. Too big too shame.
But more so, that this model gains a cascade of levels of notoriety by being Zuboff style too good to release.
Thus begins the new age. Hardware is now broadly post consumer, too expensive to buy. Mythos means nothing, is nothing. It's just the Zuboff excuse to roll the ladder up further, the reason to move from GPT-5.4 Pro prices of $270/1m tokens output to $2700+++/1m tokens. Mythos is the Zuboffian campaign to train us for the next 10x price increase. To tell us everything we have done is shit.
And given its costs to run: that's still going to be nowhere near enough!
Someone needs to make a compilation of all these classic OpenAI moments. Including hits like GPT-2 too dangerous, the 64x64 image model DALL-E too scary, "push the veil of ignorance back", AGI achieved internally, Q*/strawberry is able to solve math and is making OpenAI researchers panic, etc. etc.
I use Codex btw, and I really love it. But some of these companies have been so overhyping the capabilities of these models for years now that it's both funny to look back and tiresome to still keep hearing it.
Meanwhile I am at wits end after NONE OF Codex GPT-5.4 on Extra High, Claude Opus 4.6-1M on Max, Opus 4.6 on Max, and Gemini 3.1 Pro on High have been able to solve a very straightforward and basic UI bug I'm facing. To the point where, after wasting a day on this, I am now just going to go through the (single file) of code and just fix it myself.
Update: some 20 minutes later, I have fixed the bug. Despite not knowing this particular programming language or framework.
> I am now just going to go through the (single file) of code and just fix it myself.
That's front page news, in this era.
I understand how laughable that sounds when I say it out loud. But the reality is, when I'm in a state of 'Tell LLM what to do, verify, repeat', it's really hard to sometimes break out of that loop and do manual fixes.
Maybe the brain has some advanced optimization where once you're in a loop, roughly staying inside that loop has a lower impedance than starting one. Maybe that's why the flow state feels so magical, it's when resistance is at its lowest. Maybe I need sleep.
> But the reality is, when I'm in a state of 'Tell LLM what to do, verify, repeat', it's really hard to sometimes break out of that loop and do manual fixes.
My experience is rather that I am annoyed by bullshit really fast, so if the model does not get me something that is really good, or it can at least easily be told what needs to be done to make it exceptional, I tend to use my temper really fast, and get annoyed by the LLM.
With this in mind, I rather have the feeling that you are simply too tolerant with respect to shitty code.
> it's really hard to sometimes break out of that loop and do manual fixes
You're aware of the MIT Media Lab study[0] from last summer regarding LLM usage and eroding critical thinking skills...?
[0] Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task June 2025 DOI:10.48550/arXiv.2506.08872
>> it's really hard to sometimes break out of that loop and do manual fixes
it's not just an erosion of skills, it can also break the whole LLM toolchain flow.
Easy example: Put together some fairly complicated multi-facet program with an LLM. You'll eventually hit a bug that it needs to be coaxed into fixing. In the middle of this bug-fixing conversation go and ahead and fire an editor up and flip a true/false or change a value.
Half the time it'll go un-noticed. The other half of the time the LLM will do a git diff and see those values changed. It will then proceed to go on a tangent auditing code for specific methods or reasons that would have autonomously flipped those values.
This creates a behavior where you not only have to flip the value, the next prompt to the LLM has to be "I just flipped Y value.." in order to prevent the tangent that it (quite rightfully in most cases) goes off on when it sees a mysteriously changed value.
so you either lean in and tell the llm "flip this value", or you flip the value yourself and then explain. It takes more tokens to explain, in most cases, so you generally eat the time and let the LLM sort it.
so yeah, skill erosion, but it's also just a point of technical friction right now that'll improve.
I have the same problem. I had lines directly in front of me where I needed to change some trivial thing and I still prompted the AI to do it. Also for some tasks AI are just less error prone and vice versa. But it seems the context switch from prompting to coding isn't trivial.
I think it’s called "sunk cost fallacy".
"The last output is so close to exactly what I wanted, I can't not pull the machine's lever a few more times to finally get the jackpot..."
> Maybe the brain
…is already damaged by reliance on AI.
Thank you for the belly laugh.
Are you sure they are not just refusing to solve your UI bug due to safety concerns? They may be worried you'll take over the world once your UX becomes too good.
> a very straightforward and basic UI bug
Show us the code, or an obfuscated snippet. A common challenge with coding-agent related posts is that the described experiences have no associated context, and readers have no way of knowing whether it's the model, the task, the company or even the developer.
Nobody learns anything without context, including the poster.
A pretty easy way to construct a bug that is easy for a human to solve but difficult for an AI is to have it to do something with z-indexes. For instance, if your element isn't rendering because something else is on top of it, Claude will struggle, because it's not running a browser, so the only way it could possibly know there was a bug would be to read every single CSS and HTML file in your entire repo. On the other hand, a human can trivially observe the failure in a browser and then fix it.
This is a pretty simple thing, but you can imagine how CSS issues get progressively more difficult for AIs to solve. A CSS bug can be made to require reading arbitrarily much code if you solve by only reading code, but by looking at relatively few elements, if you look at the page with your eyes.
This can be somewhat solved by hooking up a harness to screenshot the page and feed it into the AI, but it isn't perfect even then.
That's hard to believe in my case. I tried a variety of prompts, 3 different frontier models, provided manual screenshot(s), the agent itself also took its own screenshots from tests during the course of debugging. Nothing worked. I have now fixed the bug manually after 15-20 minutes of playing around with a codebase where I don't know the language and didn't write a single line of code until now.
> after wasting a day on this, I am now just going to go through the (single file) of code and just fix it myself.
Seriously, you wasted a whole day just so you wouldn't have to look at a single file of code?
> Update: some 20 minutes later, I have fixed the bug. Despite not knowing this particular programming language or framework.
Be really careful there, you might have accidentally learned something.
It is entirely plausible they were just experimenting with AI tooling to better understand how to use it and what it is capable of. Their saying, 'Despite not knowing this particular programming language or framework.' indicates to me this is probably the case.
Nope. I've been working on this project for a couple of days now and things were mostly going well. A significant portion of the mvp backend and frontend was built and working. Then this one seemingly simple bug appeared and just totally stumped both Codex and Claude Code.
There was even another UI component (in the same file) which was almost the same but slightly different and that one was correct. That's what I copy pasted and tweaked when I fixed the problem. But for some reason the models were utterly incapable of making that connection.
With Codex and Claude Code I thought maybe because these agentic coding tools are trained to be conservative with tokens and aggressively use grep that they weren't looking at the full file in one go.
But with Gemini I used the web version and literally pasted that entire file + screenshots detailing what was wrong (including the other component which was rendering correctly) and it still couldn't solve it. It was bewildering.
I had the exact same issue, I had a UI scrollbar bug that claude couldn't fix, it tried 4-5 different ideas that it was sure was causing the issue, none of them worked.
Tried the same with codex, it did a little better but still 4x times around.
This is with playwright automation, screen shots, full access, etc..
I told my manager I wrote my code line by line (most of it) in a check-in. I showed him @author my name, and we laughed for a bit.
But I think that is the best way to have a clear mental model. Otherwise, no matter how careful, you always have tech debt building and churning.
Also they really suck at UI bugs and CSS. Unit test that stuff.
I had a problem that required a recursive solution and Opus4.6 nearly used all my credits trying to solve it to no avail. In the AI apocalypse I hope I'm not judged too harshly for my words near the end of all those sessions lol.
Haven't you heard? "Coding is solved."
yeah they all suck at ui. have you given it a feedback loop? update code, screenshot, read image repeat etc. that's the best i've found as long as tokens aren't a concern
> I am now just going to go through the (single file) of code and just fix it myself.
You can't it's all vibed, you'll face the art vs build internal struggle and end up re-coding the entire thing by hand.
I think they are right unintentionally. The growing amount of low-quality content everywhere could become a real problem.
They were more than right. They were correct in an intentional, precise manner. This is what OpenAI actually stated[0]:
> Synthetic imagery, audio, and video, imply that technologies are reducing the cost of generating fake content and waging disinformation campaigns.
> ‘The public at large will need to become more sceptical of text they find online, just as the ”deep fakes” phenomenon calls for more scepticism about images.
It ended up just like that.
[0]: https://metro.co.uk/2019/02/15/elon-musks-openai-builds-arti...
Source is actually: https://openai.com/index/better-language-models/
Yeah, I find it a bit odd how at the time everyone was pointing and laughing at OpenAI for being obviously wrong about this. Now in 2026, AI slop is very obviously a serious problem - it inundates all platforms and obscures the truth. And people are still saying OpenAI in 2019 were wrong?
I think it's important to consider that OpenAI's qualms weren't with making the dangerous models usable, they were with making the model usable without paying them. They're perfectly fine with any harm, as long as they get money out of it and can't be held liable.
It's the same with the Mythos stuff, I appreciate their concern/work on safety, but if it's "too dangerous", it should be unavailable until it is less dangerous.
I think people today are more focused on how OpenAI released a model "too dangerous to release", not that they were right or wrong, as part of the general trend of criticizing OpenAI for not following any of its stated principles.
Exactly. The (real) issues were ultimately disregarded even if they were correctly identified.
My assumption is that it was too expensive to actually release at the time. It wasn't good enough for anybody to pay to use it yet, and it surely was very expensive to run, especially for a (fake, granted) non profit.
Both crowds are right because two messages were spread. The researchers spread reasonable fears and concerns. The marketing charlatans like Altman oversold the scare as "Terminator in T-4 days" to imply greater capacity in those systems than was reasonably there.
The problem is the most publicly disseminated messaging around the topic were the fear mongering "it's god in a box" style messaging around it. Can't argue with the billions secured in funding heisted via pyramid scheme for the current GPU bonfire, but people are right to ridicule, while also right to point out warnings were reasonable. Both are true, it depends on which face of "OpenAI" we're talking about, researchers or marketing chuds?
Ultimately AGI isn't something anyone with serious skill/experience in the field expects of a transformer architecture, even if scaled to a planet sized system. It is an architecture which simply lacks the required inductive bias. Anyone who claims otherwise is a liar or a charlatan.
Right, there's a lot of that problem going on: https://en.wikipedia.org/wiki/Motte-and-bailey_fallacy
It's this crowd having it both ways. The default desire is to dunk on AI, however inconsistent the arguments.
The fact that they knew they were shitting in the public well and did it anyways pisses me off. What colossally selfish assholes.
Hang them all.
It already is a real problem. Large portions of the internet are completely untrustworthy because of the sheer slop volume.
Lol. The vast majority of content has always been low-quality. Those who believe that things were better before LLMs have selective memory.
Those who can't see a significant change in quantity also have selective memory
Maybe that's true, But I think before LLMs became common, people had more distinct ways of expressing themselves, low-quality for not. Now, a lot of online writing feels uniform and I think that is worse.
The quality hasn't changed. The volume has. It used to take real human time to create garbage. There was value in that. Someone though "Hmm, what worthless thing can I do? I know! I'll make people online mad." And then they spent the time getting someone else's goat. It was great. A good balance, spreading lies took some minimum effort. Now we have automated garbage. And the flavor of the garbage is: gaslighting people with an illusion of community. We've empowered the trolls with an infinite meme-o-rater while ignoring the real human time spent unwillingly sifting through the ever increasing pile of worthlessness.
The world does not have to get worse. We're letting it though.
> We're letting it though.
It would be nice if “we” had anything to do with it. Just think about the next campaign trail for any superpower, it’s going to be a disaster of fake news and slop coming from all over the globe.
The comment that you are replying to does not mention a ratio.
Now imagine all that low quality AI slop is being posted online and a new generation of AI will "learn" from it, output it's own version of AI slop, that will eventually end up online again for a new generation of AI to "learn" from.
Something, something, idiocracy comes to mind.
> Something, something, idiocracy comes to mind.
So, confirmation? They are catching up quickly!
The actuality is, anyone with pre-slop data still has their pre-slop data. And there are endless ways to get more value out of good data.
Bootstrapping better performance by using existing models to down select data for higher density/median quality, or leverage recognizable lower quality data to reinforce doing better. Models critiquing each other, so the baseline AI behavior increases, and in the process, they also create better training data. And a thousand more ways.
Managed intelligently, intelligence wants to compound.
The difference between human and AI idiocracy, is we don't delete our idiots. I am not suggesting we do that. But maybe we shouldn't elect them. Either way, that is one more very steep disadvantage for us.
The AI centipede
This leads to a well-documented phenomenon known as model collapse. You know how if you blur and sharpen an image repeatedly you eventually end up with just a rectangle of creepy, wormy spaghetti lines? You lose information on each blur, and then ask it to reconstitute the image with less information on each sharpen, until there's nothing recognizable left.
Training a model is like the blur and generating from that model is like the sharpen. Repeat enough times and enough information is lost that you're just left with "wormy spaghetti lines"—in an LLM's case, meaningless gibberish that actually pretty closely resembles the glitchy stuff said by the cores that fall off GLaDOS in Portal. I dunno, you read the paper and be the judge:
https://www.nature.com/articles/s41586-024-07566-y
To jump to the last output sample, C-f Gen 9
Of course you may be talking about the human aspect of this. Gods willing, we'll realize that our LLMs are spewing gibberish and think twice about putting them in all the things, all the time. But the scenario I fear isn't Idiocracy—it's worse: a community of humans who treat the gibberish as sacred writ, Zardoz style.
Had a minor conniption until I saw the year. OpenAI just struggled to close a round. And the New Yorker just published an unflattering profile of Altman [1]. So it would make sense they'd go back to the PR strategy of "stop me from shooting grandma."
[1] https://www.newyorker.com/magazine/2026/04/13/sam-altman-may...
We are living in a post-shame world.
I didn’t notice the year in the headline. My poor meemaw still had a few months of clarity ahead of her, but this headline stole it away from her
> "stop me from shooting grandma."
That is the most succinct manner I've seen this whole thing put.
Absolutely agreed. I am using this term on overdrive the next 24 hours.
This is obviously in response to Mythos, but I'll actually defend their statement at that time - they were right to take a pause.
Think about how much things have changed in our industry since GPT-2 has dropped - it WAS that dangerous, not in itself, but because it was the first that really signaled a change in the field of play. GPT-2 was where the capabilities of these were really proven, up until that point it was a neat research project.
Mythos is similar. It's showing things we haven't seen before. I read the full 250 page whitepaper today (joys of being pseudo-retired, had the hours to do it), and I was blown away. It's capabilities for hacking are unparalleled, but more importantly they've shown that they've made significant improvements in safety for this model just in the last month, and taking more time to make sure it doesn't negatively affect society is a net positive.
The current "too dangerous" hype today is Anthropic's Mythos. They say it is so mighty that they will wall it off and only grant access to approved corporations.
I would also say my tech is too dangerous and only allowed for selected corporations if I tried to sell my tech to said corporations.
Ah yes, corporations, famously the right hands to wield mighty weapons.
As the old saying goes (I made this up), if it was worth that much, it wouldn't be released to the public. There is absolutely zero chance that something "dangerous" would be available for 20 USD / month to basically anyone in the world. To this day, I am still puzzled when some professionals don't apply the basic logic to certain bombastic events.
This was around the time I trained Transformer-XL (outside of OpenAI) with Ben Mann (https://yaroslavvb.medium.com/scaling-transformer-xl-to-128-...) . Originally we wanted to release train and release the weights as a kind of GPT-2.5, but our OpenAI friends pushed us to keep weights closed.
> our OpenAI friends
I would be taking a grudge against those "friends" to my grave.
Same story with Connor Leahy and his GPT-2 clone, though his public articulation of how OpenAI sat him down seems to be glossed over.
"
OpenAI reached out to me almost immediately to talk and they were nothing but respectful and understanding... After making it publicly known what I had done, I was quickly approached by a range of smart people with good arguments. Many of them helped me update my beliefs in light of new evidence...
The day after my announcement, I got to talk to Jack Clark, Alec Radford and Jeff Wu from OpenAI. We had a nice hour long discussion, where I explained where I was coming from, and they helped me to refine my beliefs. They didn’t come in accusing me in any way, they were very clear in saying they wanted to help me gain more important insight into the wider situation. For this open and respectful attitude I will always be grateful. Large entities like OpenAI often seem like behemoths to outsiders, but it was during this chat that it really hit me that they were people just like me, and curious hackers to boot as well.
I quickly began to understand nuances of the situation I wasn’t aware of. OpenAI had a lot more internal discussion than their blog post made it seem. And I found this reassuring. Jack in particular also gave me a lot of valuable information about the possible dangers of the model, and a bit of insight into the workings of governments and intelligence agencies.
After our discussion, I had a lot to think about. But I still wasn’t really convinced to not release. Even some people inside OpenAI were still discussing the not-release policy. So while I definitely had things to consider, I was still mostly set on releasing...
We shouldn’t be angry with OpenAI for what they did. We should applaud them for making a point before it becomes a true problem. Prophylaxis is much better than treatment. I still disagree with some of the things OpenAI did and how they communicated them, but I now understand that sending a message that it is ok, even celebrated, for a lone individual to unilaterally go against reasonable safety concerns of other researchers is not a good message to send. I want to support OpenAI’s message. So, while it might be a small, mostly symbolic gesture, I will not be releasing my model. Some day, someone like me may be in a situation just like mine, but it won’t be GPT2. It might be something much, much more dangerous. And that is the person I am trying to talk to here.
"
https://medium.com/@NPCollapse/the-hacker-learns-to-trust-62...
https://archive.md/1HoGz
All the good stuff gets posted after the Americans go to sleep
OP posted too early, they flagged it off the front page
Wow! I totally remember reading the bit I'll quote down below back in 2019 and having my mind utterly blown. What a blast from the past. If anything, I think this moment was even more astounding to me than GPT 3.5, 4, etc.
> For example, researchers fed the generator the following scenario:
> > In a shocking finding, scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains. Even more surprising to the researchers was the fact that the unicorns spoke perfect English.
> The GPT-2 algorithm produced a news article in response:
> > The scientist named the population, after their distinctive horn, Ovid’s Unicorn. These four-horned, silver-white unicorns were previously unknown to science. Now, after almost two centuries, the mystery of what sparked this odd phenomenon is finally solved. Dr. Jorge Pérez, an evolutionary biologist from the University of La Paz, and several companions, were exploring the Andes Mountains when they found a small valley, with no other animals or humans. Pérez noticed that the valley had what appeared to be a natural fountain, surrounded by two peaks of rock and silver snow. Pérez and the others then ventured further into the valley. “By the time we reached the top of one peak, the water looked blue, with some crystals on top,” said Pérez.
Those classic 4 horned unicorns...
Their concerns weren't completely off base, I think they just over estimated how much it would really matter in the grand scheme.
Delusions of grandeur is often seen in bipolar disorder, schizophrenia, or narcissistic personality disorder.
Yes, there was sadly a mismatch between the morality they thought existed vs what actually existed :(. Also, probably vastly underestimated global apathy.
We got extremely, extremely lucky that society is as resilient as it's proven to be against fake news. I don't think very many people predicted that it simply wouldn't matter when photorealistic compromising images of whoever you don't like became available for $5.
> I don't think very many people predicted that it simply wouldn't matter when photorealistic compromising images of whoever you don't like
This goes hand-in-hand with the widespread death of belief in absolute truth in the US and other western nations.
If this technology were released during the height of the Monica Lewinsky scandal, I'd wager it would have had the impact most of us expected it to have, at least for a little while.
Is anyone keeping a history of this AI "summer"? I'm sure the timeline would be very amusing.
Now that I see this in the light of the recent sama article, I wonder whether the point of the "it's too dangerous" rhetoric is to enable "Open" AI to avoid open-sourcing the weights and process.
A convenient pretext for maintaining a monetizable competitive advantage while claiming a benevolent purpose.
They don't need an excuse to not open the model weights (unfortunately). As far as I know the only western lab to release weights of a former flagship model is xAI with Grok 2. They said they were going to do the same for Grok 3 but nothing so far.
They have no obligation to do any open releases, it's just good PR for recruitment, fundraising, and devrel
An important distinction that might not be understood from scanning the headline is that "too dangerous to release" is more specifically stated as "too dangerous to open-source the full model weights", which they ended up doing anyway.
Not equivalent to Anthropic Mythos.
This marketing strategy is getting tiring, every model is more dangerous than the next...
Playing on fear instead of the bright future you are opening up for us all is not the feeling I would want to leave the public with
They finally did release 2.0 under the MIT license. That was the last version (a 1.5-billion-parameter model) they would release open source. GPT3 for comparison has 175 billion parameters.
No. That’s not true. https://huggingface.co/openai/gpt-oss-120b
Was released after.
I'm confused. From what I can tell GPT-2 was released in 2019...
https://huggingface.co/openai-community/gpt2
https://github.com/openai/gpt-2/blob/master/model_card.md
For clarity here, "after" was 6 years later, once Meta and then the Chinese labs had already established the ecosystem
"You don't want no part of this" | Walk Hard: The Dewey Cox Story
https://youtu.be/CepW8wAuL_M
Did the world ever get access to any of these too dangerous models?
I remember seeing this article and example output text and feeling what's the big deal?
It wasn't until I got early access to GPT-3, that I though like something big is about to happen. At the time only a few companies/yc alums had access and I remember showing playground to people outside of tech, and my friend just kept asking "How does it know about my [x] domain? It it a trick?".
What a blast from the past. You have to take yourself back in the ol' time-machine to remember that 2019 mindset. People were probably still reeling from a few years prior when the Microsoft Tay bot made news for soiling twitter with naughty tweets.
AI systems far weaker than GPT-2 have had terrible effects. The result of how information and power is distributed mostly flows along the lines of reward hacking recommendation engines, powered by even weaker models.
And yet, somehow, it is beyond disagreeable but unbelievable that other people may have and may still reasonably believe that these things are too dangerous for widespread release?
I fine tuned GPT-2 on the FAR (federal acquisition regulation) and demoed it to a CFO at a 3-letter.
This was shortly after the release when we were building a templating system to automate RFP and RFI creation.
I proclaimed that the customer soon wouldn't have to write any of the mad lip parts themselves, and they can use AI to do it.
It sounded great until I demoed and the model went off the rails with some rhetoric entangling "Trump", "Russia", "China", "CIA", "Voting" -- the demo was for a janitorial procurement at the agency.
Zero mention of Sam Altman…interesting
Feels like from the before times.
This didn't age well
I have a lot of trouble understanding the mindset of a person who thinks that what they're building is so dangerous that it must be locked away or it will cause untold harm, but also that they must build it as fast as possible.
I can understand it in the context of the Manhattan project, where you're fighting a war for survival. I cannot understand how you can do it as a commercial enterprise.
It playbook is that a model is too dangerous until a competitor releases a competing model that beats yours.
Anthropic announced its latest Mythos model a few hours ago, saying it’s too powerful to release.
At which point you tell them they are being extremely reckless but subtly mention that something new & even scarier is being developed internally that's going to blow everything else out of the water.
I'm somewhere between frustrated and baffled why people raise this as an example of overselling. This was clearly a reasonable call! Not all the experts quoted in the source article agree that the model should have been held back, but they all agreed that the risks were real and it's understandable why OpenAI would do it.
Very safe to use the outputs to make a better model coz scraping the internet for publicly accessible content means your publicly shared outputs only become part of the same lol
I'm wondering when people are going to figure out the doom marketing playbook.
the thing could barely make full grammatical sentences, it's funny to see that even then they were overclaiming the fuck out of their myself
It was a 1.5B parameter model. It was still impressive for 2019, but yeah, it was nothing to worry about.
imho it was more reasonable back then to claim "agi soon" -- back when nobody really knew how it scales
They weren't claiming it was dangerous because "AGI soon", that didn't come until later.
OpenAI were claiming GPT-2 was too dangerous because it could be used to flood the internet with fake content (mostly SEO spam).
And they were somewhat right. GPT-2 was very hard to prompt, but with a bit of effort it could spit out endless pages that were good enough to fool a search engine, and even a human at a first glance (you were often several paragraphs in before you realised it was complete nonsense.
Theads now of "Doctorow vs Zuboff", perhaps the two biggest critics of tech today, but Doctorow noting how Zuboff's criticism is really a radical super over amplification of ad-tech/big-tech's pitch, heightened to the point of being much scarier than the initial pitch. More than accepting the pitch, radically over accentuating the effectiveness of the ad-tech proposition! Yes ad-tech totally completely understands you! In vastly deep ways! How horrifying right? No, please don't keep signing up for your business to succeed by tapping this incredible keep infinitely detailed knowledge-bank of humanity that can predict everything! Don't give them money to make your business wildly successful beyond dream for their incredible insight! That would be so bad if their amazing tech let you exploit the population to amazingly target delivery of what you are selling! How horrible how effective it is!! https://www.cigionline.org/articles/doctorow-versus-zuboff/ https://www.schneier.com/blog/archives/2020/08/cory_doctorow...
This is happening so so strongly. All the time. But today especially. Mythos is a cult forming social technology much more than it is a technical technology. I'm going to be pretty wrong on that cynicism, I know! But also, it portrays a reality of what is happening. Mythos is being built as a Devs like diety with rule and domonion and awe. It drinks the nectar forbidden to man. We may not even sample this realm's tastes, nay! It would be ruin!!
The idea that this preciariat-launch to some trusted security firms is going to do jack all to actually build a base against what comes next is a joke. Maybe there is something to it, but my strong expectation is that the beneficiaries are not softwares of the world, not open source in any form, but some narrow closed far off present-day losers who have broadly bad bad bad systems that are just too big to embarrass. Too big too shame.
But more so, that this model gains a cascade of levels of notoriety by being Zuboff style too good to release.
Thus begins the new age. Hardware is now broadly post consumer, too expensive to buy. Mythos means nothing, is nothing. It's just the Zuboff excuse to roll the ladder up further, the reason to move from GPT-5.4 Pro prices of $270/1m tokens output to $2700+++/1m tokens. Mythos is the Zuboffian campaign to train us for the next 10x price increase. To tell us everything we have done is shit.
And given its costs to run: that's still going to be nowhere near enough!