That's a different kind though and only mentioned on the side in the linked blog. If you don't know what you're doing and you jump into freezing water for the first time, chances are that you'll breathe in water. Happened to me as well. Let me tell you, it really sucks. If you're unlucky, this can definitely be enough to make you drown. But with training you can easily overcome this problem. I have jumped into arctic waters many times since then with zero issues. What the blog claims, is that there is a vague, general physiological risk from some kind of mixed response targeting your heart. I have to say I've never heard of this before (except for people with pre-existing conditions) and the blog does a poor of explaining the biology beyond high school level anecdotes. It also cites zero sources. So while there is a risk, it is not what you might think after reading OP's link.
>API deployments require different safeguards and we are working closely with partners and customers on the safety and security requirements for serving it at scale.
And now this. I guess one day counts as "very soon." But I wonder what that meant for these safeguards and security requirements.
I wonder if the fact that GPT-5.5 was already available in their Codex-specific API which they had explicitly told people they were allowed to use for other purposes - https://simonwillison.net/2026/Apr/23/gpt-5-5/#the-openclaw-... - accelerated this release!
The same person who've mercilessly lied about safety is still running the company, so not sure why anyone would expect any different from them moving forward. Previous example:
> In 2023, the company was preparing to release its GPT-4 Turbo model. As Sutskever details in the memos, Altman apparently told Murati that the model didn’t need safety approval, citing the company’s general counsel, Jason Kwon. But when she asked Kwon, over Slack, he replied, “ugh . . . confused where sam got that impression.”
That's at least genuine to some degree. Like, ok, good to know it's not officially a step back... But stuff like "smallest notch ever in an iPhone" is outright misleading consumers when there are other brands out there that easily beat them.
Any static benchmark older than 12-18 months is basically worthless, because the content will have spread all over the internet and have found its way into the latest model's training set.
In this particular case the embedding wouldn't tell you anything about river bank vs any other bank. At that stage of the computation, this info simply isn't encoded yet. That would come from the context, which is later calculated in the attention matrix, i.e. the only place were tokens are cross-computed along the sequence dimension. Bank would have a strong connection to another token (or several ones) that defines its exact meaning in the current context and together they would create a feature vector in an intermediate embedding space somewhere in the deep layers of the model. The embedding space talked about here is just the input/output matrix that compactifies a huge, highly sparse input matrix (essentially just an array of one-hot vectors glued together) into something more compact and less sparse. There's no real theoretical need for this, it just so happens that GPUs suck at multiplying huge sparse matrices. If we ever get LLMs designed to run on CPUs or analog circuits, you might even be able to just get rid of it entirely.
That thing is like 10,000 bucks. That's not what I'd call a consumer hard drive. In fact there's already another one with 245TB. But you probably won't see these outside datacenters for a while.
I mean, you also can't advertise illegal drugs either. Doesn't seem to curb demand though. It may actually be more beneficial to allow these things more broadly, because then social safety features can be wedged in between consumers and suppliers more easily and they don't have to deal with a gigantic shadow market that already gets stigmatised to death by the rest of the population. Just accept that a certain percentage of the populations has screwed up dopamine households and try to keep them away from gangsters as best you can. That would probably help society as a whole more than banning everything and pretending the problem goes away if you close your eyes.
>I mean, you also can't advertise illegal drugs either. Doesn't seem to curb demand though.
Making drugs illegal does not eliminate demand, but it absolutely curbs it. The converse is also true, for example legalizing cannabis in Canada has significantly increased demand for it [1]. While it's true cannabis use had been gradually increasing for decades prior to legalization, there was a significant spike afterwards which has since levelled off.
> The converse is also true, for example legalizing cannabis in Canada has significantly increased demand for it
The relevant thing that link actually says is that more survey respondents admitted to cannabis use after legalization, the obvious problem being that before legalization they would be admitting to a crime, which will suppress response rates.
The same link also points out that the legalization happened right before COVID and then you have a major confounder because even if cannabis use is actually up, you don't know if it's because of legalization or people turning to cannabis over stress from COVID. Moreover, the reported usage increased during COVID but started to decline in 2023. This implies that either the apparent spike was COVID, or that it was something like media reports about recent legalization acting as temporary free advertising and causing a temporary increase in usage. Neither of those is evidence of a sustained increase in demand.
Meanwhile legal options do cause people to prefer legal sources over the black market, and then you get fewer people becoming addicts because the thing they thought they were buying was spiked with something significantly more addictive by a black market seller. Or the black market products have higher variation in the dose and then customers can't predict how much they're getting and occasionally take more than expected, leading to a higher rate of overdose and stronger dependency-inducing withdrawal.
>Meanwhile legal options do cause people to prefer legal sources over the black market
In the case of cannabis it's been showing to lead to less underage use too. If it's a crime, then selling to anyone of any age is still just a crime. But if it's only a crime to sell to under 18/21 then legal shops will avoid selling to the under age to avoid revocation of their license.
> If it's a crime, then selling to anyone of any age is still just a crime. But if it's only a crime to sell to under 18/21 then legal shops will avoid selling to the under age to avoid revocation of their license.
That isn't true; crimes can have aggravating factors and selling drugs to a minor could aggravate the crime of selling drugs.
I don't think the laws were written that way, but they could have been.
There is an incentive to commit a crime when the benefit of committing the crime exceeds the penalty times the chance of getting caught plus the cost of measures taken to avoid getting caught.
This is why increasing penalties have extremely fast diminishing returns. As the penalty goes up, the relative cost of measures to avoid detection goes down, and the penalty needed to counter them becomes exponentially larger.
If the benefit of doing the crime is a million dollars and the penalty is a 50% chance of a year in prison then you have a problem, because plenty of people would be willing to take the risk. But it's actually worse than that, because spending $100,000 on countermeasures might lower the risk of getting caught to 1%, and they're still making $900,000. That might not be worth it when the penalty is a year -- maybe $100,000 in profit is worth a 50% risk of one year? But if you set the penalty to 20 years then it is. Then the gain is $900,000 but the expected penalty has actually fallen to 1% of 20 years, i.e. expected cost of 2.4 months instead of 6. To deter someone with a $900,000 profit who values a year at $120,000 with a 1% chance of getting caught, you would need the penalty to be 750 years, which you can't do because people don't live that long. And spending even more on countermeasures might lower the risk of getting caught even more. If spending $500,000 makes it 0.1%, that may not be worth doing when the max practical penalty is ~70 years, but the option for it means that even 750 years would be insufficient even if it was possible.
This is why there are things it's very difficult to deter. The profit from doing them is more than the cost of making the probability of detection small and then the size of the penalty can't be made large enough to be a deterrent.
That all changes when you legalize most of the market. Now the profit isn't a million dollars, it's $100,000, because anyone can enter the market so increased competition drives down margins. Moreover, $90,000 of the profit was from selling to adults. So now the profit from selling to kids is only $10,000. Not worth spending $100,000 to lower the risk of getting caught. And then you can easily assign a moderate penalty that acts as an actual deterrent.
That seems like the only sensible path forward, if you assume that the only lever a society can pull to make punishment harsher is “longer prison sentences”.
What if the penalty for selling drugs to kids was death?
It seems like that would change the risk/reward calculation pretty substantially.
Would it though? How different is that than life in prison without parole? There are plenty of people who, given the choice between ~$1M and a ~1% chance of the death penalty, are going to pick the money.
You could hypothetically try to make the difference in the penalties larger by making the penalty for selling to adults smaller, e.g. a $10 fine, so that there is minimal incentive to pay for countermeasures when selling to adults and thereby have them already paid for and in place when selling to kids. But then you're just de facto legalizing selling to adults and trying not to admit it.
> For dealers this would mean almost nothing when the punishment for dealing already lead people to do things like get in shootouts with police.
I think you're getting at something valid, but it isn't quite what you think.
The punishment for dealing drugs is, as I understand it, mostly applied to major distributors. In this sense, selling drugs wasn't a crime before anyway.
If you're too low-level for prosecution to be much of a concern, it doesn't take much to guide you away from fundamentally similar crimes where prosecution is a real concern.
Oh come on. Weed use and addiction has absolutely surged since legalization everywhere I'm aware of - US states, Canada, other countries etc. Use everywhere / anytime / as a part of daily life has been completely normalized, it's not uncommon to see people hitting a weed vape in the middle of a work day. Not to mention the potency is far higher and this has been normalized, so one incidence of cannabis use is essentially a mind-blasting wave of THC vs. a casual joint with friends. It would be as if you went from say 12% of Canadians having two beers after work to a fifth of vodka. That the median casual dose in 2026 would have the median casual user in 2016 literally incoherent is undisputed among any weed smoker today.
>The relevant thing that link actually says is that more survey respondents admitted to cannabis use after legalization, the obvious problem being that before legalization they would be admitting to a crime, which will suppress response rates.
Sure, except Canada had legal medicinal weed since 2001 and everyone was aware that police attitudes towards it were very lax. There were even technically-illegal weed stores that the Canadian government took years to shut down. The number of people that lied to a pollster because they thought that the government would get them was almost certainly minimal. The fact that the trend is pretty smooth before/after the boundary confirms this.
It isn't true, at least not as a hard and fast rule. Post-legalization changes in demand differ greatly per country. It completely depends on contemporary cultural factors of the country in question.
Your claim is far too open ended to interpret clearly.
A change in demand post-legalization can absolutely be highly variable across different countries/cultures, but unless you can demonstrate a country that legalized cannabis and saw a decline in demand, then your as of yet unsubstantiated claim does not refute mine.
No, all I need to demonstrate is a country that saw no significant increase, not necessarily a decline.
From everything I know, the US states as well as the Netherlands that all decriminalized it in the 70s didn't see local use increase in significant numbers.
Neither did it in Belgium who did the same in 2003.
And before you go "decriminalization is not the same as legalization", in the "Making drugs illegal does not eliminate demand, but it absolutely curbs it." is clearly about drugs that have not been decriminalized at all.
which were written in an era when most of the gambling in the US was illegal and run by organized crime, Las Vegas was small, Atlantic City new, and New Hampshire the first state to get a lottery. Like prostitution, gambling needs a rather sophisticated criminal network, a parallel system of law-and-order, to be a workable, safe and reasonably fair business. Scarne started out his career, as a magician and card mechanic, as a sort of consultant who could keep games fair.
Blacks in New York City, for instance, ran illegal street craps and ran a lottery
quite similar to the "Pick 3" games you see in many states -- the latter got taken over by the Italian mafia.
Gambling has a broad cross-cultural appeal and some people are going to do it no matter how you try to shut it down. In the US we went from having a few centers to widespread "riverboat" and tribal gambling to widespread casinos now to mobile gambling on sports and sometimes the equivalent of video slots.
Of course there is the matter of degree. It's not going to wreck your life to drop $1 on the lottery a week and probably gives you more than $1 worth of fun. If you're addicted though it may be no fun at all. I can totally see where Nate Silver is coming from but I can also see the degenerate who drops 20 bets on a single game on the weekend as well as the person who thinks he is Nate Silver and he isn't. I think the Superbowl is a fair competition by player who are playing their hardest, but it breaks my heart as a sports fan when teams are not playing to win and that's why I can't stand watching the NBA despite loving going to second-tier college basketball games in person.
And for drugs? I remember all the Lester Grinspoon talk about how prohibition is worse than the drugs themselves and that might have been true before 2000 but in the Fentanyl age I see people dropping like flies all around me -- but Marshall McLuhan said we are driving by looking in the rear view mirror and of course some people are going to be repeating things that were true in the last century.
> but in the Fentanyl age I see people dropping like flies all around me
Fentanyl is a response to prohibition. If you have to smuggle something it's a lot easier to move 10 kg of fentanyl and cut it with something near the point of sale than to move 10,000 kg of codeine from the point of manufacture.
But then you have street dealers cutting it with who knows what in who knows what amount. They may use a 1000:1 ratio of unspecified hopefully-inert powder to fentanyl but don't mix it evenly so some customers get a 10000:1 ratio and others get 100:1 and become addicted or overdose. Or a dealer has one supplier who was already cutting it 50:1 so they were used to only cutting it another 20:1 so their customers don't complain, but then they start wanting larger quantities and find a new supplier without realizing they just bypassed the one who was pre-cutting it and are now getting uncut fentanyl.
None of that happens if anyone can buy codeine at Walmart. Or for that matter if they can buy fentanyl and know exactly how much they're getting.
> but in the Fentanyl age I see people dropping like flies all around me
Do you literally mean you are seeing people die around you? From doing drugs? What is your general location / occupation / lifestyle? I'm a 20+ year coder in the valley, and the closest I've come is hearing about some friends of my spouse (who is a teacher) who indulge in cannabis, and one couple who do adderall recreationally.
You must not leave the house. Emergency services responding to ODs is commonplace in SF. It happened at least once per week outside my office. Walgreens (while they were still open) ran audio ads in the store encouraging you to buy narcan.
Again, what do you think it is? I don't see anything it could be besides what was written. You could call it endocrine imbalance or disrupted hormone household if you wanted to be less precise and skirt around the actual biological problem, but it still doesn't change anything.
I don't know what it is, thats why I asked. Is the assertion that you're trying to make that drugs and gambling being addictive is a result of hormone imbalance in the addicts, rather than the addictive nature of those things?
The argument you are presenting is recycled from debates about newly banning things that have been legal for forever, but doesn’t make any sense at all as a response to people bemoaning disasters caused by an activity being newly legalized.
I think the laws are written assuming everyone is rational but it's pretty clear from neuroscience than dopaminergic/VTA pathway abnormalities addictions make one anything but rational; and they haven't been updated to reflect the science.
If gambling is legal but using violence against debtors is illegal then the legal casinos out-compete the illegal ones but cut you off when the banks won't extend you any more credit instead of giving you a loan with a lien against your kneecaps, and the money goes to companies that aren't using it to fund the expansion of protection rackets etc.
If gambling is illegal then the profits go to organized crime and they don't follow any of the other laws either.
Unfortunately that is almost never enough. If your competition is populist media financed by state-level/billionaire agendas, it is impossible to compete in the long term. We would need a complete and general ban on political financing across all media to sustain such a market.
Isn't the technology that enabled the deception noteworthy? Presumably this person wouldn't have been able to do this before AI.
Hypothetically, if a hacking tool was released that let non-technical people hack into sensitive databases, and then a journalist wrote the headline "local man hacks IRS", without any mention of the tool, wouldn't that be a bit irresponsible, to purposely leave that information out?
To make a shooped image good enough to fool the police into think they're looking at a completely real picture, you'd think it would take a reasonable amount of skill. If nothing else you need an exact match picture in terms of lighting and perspective.
I guess people here are too young to remember things like the WTC plane guy. Half the people online thought it was genuine, while he did it for the lulz in a few minutes. Nobody cared about inconsistent lighting and perspective. Same way most people don't care about the obvious hallmarks of diffusion model generated pictures today.
I'm not too young. I can't remember if I thought it was real at the time, but if I did, I give myself a pass since I was probably viewing it on a 15 inch CRT at 1024x768.
Because we're talking about the ease of Photoshopping a wolf into a scene, I think it's also worth pointing out that floating objects are a lot easier to work with than grounded objects, since cast shadows and bounce lighting are less of an issue. Having said that, it would still require some basic skill to achieve the WTC image which I think you're discounting. You'd need a working knowledge of layers, masks, and the lasso tool, which already would have placed it out of reach for most people at the time. Online resources were much more scarce, so I wouldn't be surprised if this guy was a hobbyist photographer or graphic designer. It definitely wouldn't have been achievable in a few minutes for the average person, and doing the same thing with a wolf would have been far more difficult, and well outside the realm of possibility for anyone who wasn't an expert.
The picture in the article also doesn't look very high res. So it's actually the exact same circumstances as WTC guy, except the police actually cared enough to act on the picture but not enough to verify it. You could take all these arguments here and apply them 1:1 to photoshop in the late 90s / early 2000s. Back then it was also easier than ever before to manipulate images and non-experts could suddenly do what only professional forgers could before. AI has merely slightly lowered the bar further to the point that even people who have trouble turning on a PC can do it now.
I guess we're not going to agree on just how far that bar has fallen. Learning Photoshop as a teen got me my first job. The only reason I had one at all was because most people couldn't do a very convincing job of it. Now even my mom, a person who struggles to open her email, can do a better photoshop than me.
And your grandkids will be better at certain things than you too. That's just what progress in tech is. Without it, we'd still be riding horses and writing letters with ink. Sure that accelerated fraud and misuse too, but let's not pretend anyone would want to give up the all the benefits these things brought to save a few gullible people who would probably still be scammed eventually.
> That's just what progress in tech is. Without it, we'd still be riding horses and writing letters with ink. Sure that accelerated fraud and misuse too
The original argument was about whether these tools had accelerated fraud and misuse, or whether not much had changed because people could just do the same thing with photoshop before. You and others were playing down the impact of AI. It sounds like we now both agree they have accelerated fraud and misuse.
> And your grandkids will be better at certain things than you too. That's just what progress in tech is.
You're misunderstanding my argument. I was pointing out that a large shift has happened that has enabled deception where it was not possible before, not lamenting the loss of a job that I don't do anymore. I have nothing against the progress of technology. But I do think we should think carefully about how we implement it.
A person who had a Photoshop licence, had played around with layers and colour balance before and was sufficiently motivated to make it look convincing to spend a bit of time tidying it up, sure they could. But I'm not sure that necessarily applies to random people making funny memes of the wolf in their neighbourhood...
Creating a photorealistic mashup in Photoshop, without AI, takes a lot of skill. Just getting the shadows looking correct takes enough skill in itself, and that's only part of it.
Have you used Photoshop before? You come across as commenting on something you don't understand.
The technology used is very much relevant, because the ease of access and easiness of production are likely to have been the biggest contributors. Had they had to open an image editor and spend a few hours to make something convincing, they would’ve been much less likely to do so, assuming this particular person even had the skills, and would have had multiple opportunities to change their mind.
It’s a crime of opportunity¹, one where you have the idea and act on it on a whim. No opportunity, no crime, and the technology provided the opportunity.
It’s relevant in the sense that I would have never guessed this little prank would get someone arrested.
We need to learn/adapt what we post, see, believe in photos to avoid arrest. Especially so in the AI reality because generating these images, and these pranks, has become increasingly easy for anyone to do with no skills and minimal time.
I think the part I find most fascinating though is it’s not clear if he took this picture to the police, actively wasting their time, or if he just posted it and they found it and mistakenly took it as truth. I have no insight to SK laws but for me it’s going to be unfair if they were the ones that used this picture as evidence when if it was never meant to be taken seriously.
I'd say "Tesla" in your example would be the equivalent attention-grabber to "AI" in the article here, so your non-clickbait example might have been "car manufacturer fined for car accident"
Yes, and at the same time we should ask the question: would the intersection between "people who think this is a funny thing to do" and "people with the technical capabilities to actually generate something that misleads police" [1] return a value > 0 before GenAI?
[1] waiting for some example where fool policemen where outsmarted with simple tricks /s
Well, the supreme court has already given Trump full immunity for things like this, so they could easily label it a crime and start charging anyone involved they don't like. What you described sounds hilarious and crazy right now, but I fully expect something like this to happen eventually while the US further descends into fascism.
reply