5 minutes? I'll have spent more time writing this comment and posting the image to imgur than I did having the bot produce that image in the first place.
It's not grand at all; just an informative sign about some fall specials that is direct and to the point. (Nothing about it was done by hand, but nothing about it would get people all torches-and-pitchforks, either.)
Exactly. What we are really talking about is people lacking taste. People don’t want to see shit graphic design. At the state fair this year, all the food places and rides were covered an obvious low quality ai image work. The difference between acceptable and gross can be as simple as a prompt tweak.
> the frontier labs are priced according to the narrative that they have produced or will in the very near future produce a fully automated drop-in replacement for most knowledge workers
That's a reason to be bearish about AI companies, not LLMs. But is it even true? OpenAI and Anthropic have each reported ~50 billion in revenue with ~900 billion valuations. That's a high ratio but I'm not sure if follows that the only way it pans out is if we get "fully automated drop-in replacement for most knowledge workers".
It wouldn't shock me to see those revenue numbers scaling up to where they need to be over the next decade ( to, say, ~400 billion) without ever achieving drop-in worker replacements.
I looked at the math and I think it's true. Remember revenue is just sales, not profit. These labs are shooting for > $1T valuations, which traditionally means your PROFIT is at least 1/20th or 1/30th of that (so let's say minimum 30B$/year PROFIT).
These companies however are LOSING money (anthropic tries to make it sound like it's profit by deviating from accepted accounting principles) and subsidizing these models. When accounting for all the engineering salaries, training, GPUs, etc, what's their best-case realistic margin three years out, 10%?
So to we'd need a scenario where companies are spending a collective 300B annually on AI (believable) but ALSO that these companies jack up their margins WITHOUT companies switching to the cheaper open-source models (even when there's a $300B incentive to do so).
Yeah, just take the EBITDA and suddenly the valuations make sense. Paying money for a vending machine that currently loses money hand over fist is generally not a sound investment strategy
You see the same thing in literature, where modernism has crowded out of other styles. As one example, you see way less description today than in older books.
I'm not a hater on modernism -- I like stripped down literature and architecture (though I wish concrete buildings were cleaned more regularly). But there does seem to be some truth in this idea that "the elite critics" have dominated these fields, resulting in the production of art (and buildings) that are broadly unpopular. I think the overriding goal of these critics is avoiding kitsch, which they see in classical forms.
"Wake back to bed" works for me. I don't do it intentionally, I often wake up between 3 and 5, use the bathroom, go back to bed. It only triggers a lucid dream maybe 5% of the time but I've never had one without that.
I am in the quarter of aphantasics that report no imagined sensations of any kind, no inner sound, smell, taste or touch. I do have some very basic imagery occasionally while dreaming, a bit like Dali’s drawings. Never faces. I am also remarkably bad in remembering faces (or names) but recognise them just fine when recent. I don’t think in words, shapes, or relations. I think in words that are on the tip of my tongue and make sentences in their “rhyming”.
I understand people just get off posting stuff like this. But creating LLMs from the entire corpus of human text was a huge achievement. Distilling those models is much less of an achievement. It means China is further behind than we thought.
First movers rarely take the prize, though, do they? Les Paul and Mary Ford pioneered overdubbing voices back in the 1940s and 50s. The Beatles stole it from Buddy Holly. And Elton John from the Beatles.
Distilling is a massive achievement. I can run Qwen. I can't run GPT (TM). It's not a matter of X is better than Y. It's a matter of Y exists, X does not.
I'm generally in favor of mass surveillance, as long as it's in public. I think it can prevent a lot of crime and I don't believe we have a right to "privacy in public". Red light cameras, for example, strike me as unalloyed good.
However, I'd be interested to see if opponents of mass surveillance could draft good legislation to curb potential abuses. For example, the laws around who owns/controls data from cameras could be changed/redefined, as could the process of accessing that data.
I generally actually agree with you. The real issue with these cameras is the network effect that allows them to stalk anyone anywhere.
Were I to draft legislation, the first thing I'd do is ban any sort of surveillance network from giving data to non-governmental 3rd parties. Google and mcdonalds don't need to know exactly who I am and what sort of places I like to visit or products I buy.
The next thing I'd do is require a search warrant for each access. If a crime is committed then the cops need to go to a judge and say "A crime was likely committed here and we need a warrant to figure out who was there and where they went". None of this "well, here's the same case number I use for all my searches".
I would probably classify the likes of traffic cameras as something different than a surveillance network as there's not much of a real network. They are all watching one location for violations and reporting them when they happen. That to me is fine (though, again, they shouldn't share that data with private 3rd parties).
And this goes beyond surveillance networks, but I'd setup a national registry of police misconduct and bar cops who abuse their authority from ever getting a job as law enforcement again. All these cases of cops stalking exes would result in those cops needing a new career where they can't abuse the public's trust.
> Red light cameras, for example, strike me as unalloyed good.
I get where you're coming from, but there's a big problem: our laws did not evolve to be enforced effectively. They are appropriate when they are enforced selectively, with discretion. Many are wholly inappropriate if simply enforced 100% of the time to the letter of the law.
That's not to say we couldn't change laws to work if enforced near-100% of the time, but it would take our legal institutions quite a lot of time to adapt.
>I'd be interested to see if opponents of mass surveillance could draft good legislation to curb potential abuses
Isn't this just warrants? We have infrastructure in place, if police have reasonable suspicion a crime has occurred, they go to a judge and convince them they have enough evidence and the judge gives them a warrant. The problem is police and Flock are constantly arguing they shouldn't need a warrant to use the data, and it creates a system designed for abuse.
The ones making the laws are your elected officials. If you want a city to pass a law prohibiting changing the traffic light timings, then go campaign for a candidate that has that on their platform.
>If you want a city to pass a law prohibiting changing the traffic light timings, then go campaign for a candidate that has that on their platform.
people love saying this as if its some sort of gotcha.
ive been campaigning for candidates i believe in since ive been able to vote, which is over 40 years now. additionally, i voice my opinion locally and online (hi).
That's why we have an independent judiciary. What's the alternative to fixing this with laws/lawsuits? Do we riot when "the police and governments break laws"?
which is something that only happens after abuse occurs, is costly (time and money), presents a risk of retaliation, and offers no guarantee of being made whole.
>What's the alternative to fixing this with laws/lawsuits?
not doing mass surveillance means we don't need legislation about it.
>Do we riot when "the police and governments break laws"?
Right, and while we're at it, we can stop theft by having the thieves choose not to steal.
If your approach is that we can fix abuses by having the abuser write down rules that say they're not allowed to abuse us, why limit it to public surveillance? Stick cameras in everybody's houses while we're at it. Think of all the horrible crimes that go on behind closed doors. Solve the privacy issues with legislation.
If red light cameras are incorrectly calibrated, that seems like something that could be challenged in court and changed.
> If your approach is that we can fix abuses by having the abuser write down rules that say they're not allowed to abuse us, why limit it to public surveillance? Stick cameras in everybody's houses while we're at it. Think of all the horrible crimes that go on behind closed doors. Solve the privacy issues with legislation.
People have a constitutional right to privacy in their own homes.
"People have a constitutional right to privacy in their own homes."
When the current administration is wiping its ass with the constitution on a daily basis, are you really so daft you can't see the direction we're headed in as a society in regard to an unregulated surveillance dragnet?
People have a constitutional right to due process and to not get murdered in the street by ICE (that's pronounced T-R-A-I-T-O-R, btw) but look how well that's been going.
Just wait til we have an administration in the White House (I’m guessing not this one?) which directly threatens your freedoms in such a way that they can easily use cameras to prosecute those who they disagree with.
Current governments have used the mass surveillance to go after people for abortions and for being too foreign-seeming (even if here legally). If those don’t affect you, surely you can think of something that the next wave of government officials might be opposed to, which could use a surveillance network to take away from you.
Having laws to prevent abuses doesn’t help when the lawmakers change and can just revoke those laws, while the full infrastructure exists to do whatever they want.
> Red light cameras, for example, strike me as unalloyed good.
This actually makes me think that your post is satire, due to the heavy debate on whether these cameras actually improve safety [0]. Lots of studies that contradict each other on that. I agree we need to prevent the recklessness of drivers, and there is an argument to be made here. But it’s not at all obvious like you imply.
From [0]:
> When the Houston cameras were removed, angle accidents increased by 26 percent. However, all other types of accidents decreased by 18 percent. Approximately one-third of all Houston intersection accidents are angle accidents. This suggests that the program’s drawbacks canceled out its benefits.
> Our study showed no evidence that cameras reduce the total number of accidents. We estimate that total accidents are reduced by a statistically insignificant 3 percent after the cameras are turned off.
> Likewise, there’s no evidence that the camera program reduced the number of traffic-related injuries or the likelihood of incurring an incapacitating injury.
This post is a good example of why I find the anti-surveillance people so tiresome
> Just wait til we have an administration in the White House (I’m guessing not this one?) which directly threatens your freedoms in such a way that they can easily use cameras to prosecute those who they disagree with.
"You must not be gay/trans/black/hispanic/female or you'd agree with me" is not a good argument.
We're living in a brave new world where cameras and storage are cheap, networks are fast, etc. People have cameras at their front doors, stores have cameras, appartment buildings have cameras, etc. We need to deal with these facts productively, i.e. with legislation and serious thinking, rather than inane comments like "women seeking abortions have been tracked therefore we need to ban cameras".
(1) not the point I was making, I’m not any of those categories. But I am worried about the treatment of those people, as well as giving the government tools to someday come after people with guns, people who vote incorrectly, meeting up with a friend who happens to have a strong political view, etc.
(2) It can be both - we can ban the government from having direct access to mass surveillance and data privacy laws which makes retaining and using data on individuals much more restrictive.
That sounds like an argument for anarchism. I am very unimpressed by political philosophies that take arguments to their logical conclusion.
What seems to me to have worked is understanding that a government is necessary, that it has the potential to misbehave, and therefore we should carefully design laws and processes to constrain it. I would like to see opponents of mass surveillance do this, which seems more productive than what I normally observe.
Fundamentally impossible. Power structures are inherent to all human social activity. Even if we instantaneously abolished all governments and all power structures, new ones would spontaneously form immediately after.
> I am very unimpressed by political philosophies that take arguments to their logical conclusion.
People's inability or unwillingness to take things to their logical conclusion is why mainstream political philosophy is doomed to fail. If you want to test principles, then you must assign extreme values to all variables and see how the system copes. Everybody breaks under this simple test.
> we should carefully design laws and processes to constrain it
Those already exist and are routinely ignored. The US government routinely disregards the rights of its own citizens, and nothing happens, all it has to do is say the magic words "national security" and it all goes away. Every once in a while, evidence of abuses somehow surface and they are so egregious that they have no choice but to do something about it, but due to survivorship bias you must assume the abuse is far more prevalent than what you're actually seeing.
Only structural solutions meaningfully change this calculus. It must be impossible for the government to abuse you in any way. Anything short of that will fail.
No, it is an argument that the best way not to abuse something is to not have access to it. If you enshrine a ban of total surveillance on a deep and universal level in a country, it'll be a lot more difficult for a future bad actor state to reinstate it. We need the most extreme legal hurdles that can exist. More importantly, if the technology isn't actually physically installed, it becomes even more of an uphill battle for a future corrupt government. Your proposal is for surveilling everyone but making a pinky promise that no one in the chain will ever be corrupt at any point in the future, because all it would take at that point is changing some words on a page, refusing to prosecute an abuse of power, using some loopholes or pushing a software update.
The modern legally bound nation state has done much better at guaranteeing rights and protecting freedoms than any other configurations despite the fact that it's constained largely by "pinky promises" (i.e. laws, courts, etc).
A lot of those exist because of fairly strong “don’t get close to this” rules or culture. There are certainly areas where our democracies had good properties that slowly erroded too.
From what I understand, none of the people involved here are originators of the idea that led to this solution. Not Buckmaster nor Alpöge nor OpenAI. All of the above were using LLMs to push other mathematicians' ideas forward (Diego Cordoba and Luis Martinez-Zoroa; named in the linked document).
I don't know how the math community handles this but normally I would think if X mathematician comes up with an idea and Y mathematician uses it to solve some problem, Y would get credit. But does that change if Y heavily relied on LLMs? I suppose we're going to find out.
You're right that the case is not so clear cut since both sides were using LLMs. And yet it's still being seen as a major confrontation between the mathematical community and the AI industry because Buckmaster is a prominent member of the community and has taken pains to follow mathematical norms while OpenAI has not (with the most flagrant violation being the insistence on removing Alpöge from authorship, simply because of corporate affiliation), and is more nakedly threatening human ownership with capital (the OAI blog post says the final result involved 10,000 concurrent agents).
Regarding credit assignment when LLMs are involved, the mathematical community has organized around some rough principles. Gowers has some thoughts on his blog (https://gowers.wordpress.com/2026/07/26/thoughts-about-the-l...) about how explaining a result may be more deserving of credit than producing it. It looks like Buckmaster and Alpöge were taking their time in understanding their results and writing them up when OpenAI forced them to publish their work-in-progress. At the same time OpenAI has published their own writeup but it's not really clear to me how involved humans were.
Worth noting that Tao's post says the authors had "significant AI input" but are reworking them into "acceptable form". Either way, it seems AI was involved.
Of course AI was involved, you'd expect most mathematicians and researchers to use AI nowadays. This drama is about AI achieving impressive outcomes with little to no human intervention, as that would be signalling AGI.
> But the right to accurate information from public officials, I thought that was indeed an absolute and basic right in a democracy
That's a misunderstanding on your part. You can't have a "right to accurate information" because officials might simply be wrong. "Being wrong" can't be against the law. Additionally, people simply disagree about lots of things and often there's no procedure to resolve this.
However, you can't lie under oath, or in front of Congress. But lying is not the same as providing inaccurate information. In most other contexts, lying is legal.
Northern Virginia has tons of data centers because it had massive amounts of excess energy capacity.
That is now being consumed and now Loudon County will also start engaging in electricity price wars with data centers. New regulations to segregate markets is supposed to kick in on Jan 1 2027, in fact.
That aside, if you spend 5 minutes prompting AI, you can create an image that no one will clock as AI.
reply