Robots.txt isn’t even a rule, it’s a request.

“Please do not ask for the following content if you are a robot”.

If you don’t want someone to look at your content, you ultimately have to not give it to them, not just ask them to not ask.

Oh, to me it just doesn’t remotely look like they’re interested in surveillance type stuff or significant analytics.

We’re already seeing growing commercial interest in using LLMs for stuff like replacing graphic designers, which is folly in my opinion, or for building better gateways and interpretive tools for existing knowledge based or complex UIs, which could potentially have some merit.

Chat gpt isn’t the type of model that’s helpful for surveillance because while it could tell you what’s happening in a picture, it can’t look at a billion sets of tagged gps coordinates and tell you which one is doing some shenanigans, or look at every bit of video footage from an area and tell you which times depict certain behaviors.

Looking to make OpenAI, who seem to me to be very clearly making a play for business to business knowledge management AI as a service, into a wannabe player for ominous government work seems like a stretch when we already have very clear cut cases of the AI companies that are doing exactly that and even more. Like, Palantirs advertisements openly boast about how they can help your drone kill people more accurately.

I just don’t think we need to make OpenAI into Palantir when we already have Palantir, and OpenAI has their own distinct brand of shit they’re trying to bring into the world.

Google doesn’t benefit by selling their data, they benefit by selling conclusions from their data, or by being able to use the data effectively. If they sell it, people can use the data as often as they want. If they sell the conclusions or impact, they can charge each time.
While the FBI does sometimes buy aggregated location data, they can more easily subpoena the data if they have a specific need, and the NSA can do that without it even being public, directly from the phone company.
The biggest customer doesn’t need to pay, so targeting them for sales doesn’t fit, whereas knowing where you are and where you go so they can charge Arby’s $2 to get you to buy some cheese beef is a solid, recurring revenue stream.

It’s a boring dystopia where the second largest surveillance system on the planet is largely focused on giving soap companies an incremental edge in targeted freshness.

Well, I’d contend that the same expertise isn’t just readily available. Yes, he’s uniquely positioned for connection to the surveillance apparatus, but the reputation of being the federal governments head security is also a unique credential.

Yes, neither of us is responsible for hiring someone for the OpenAI board of directors, making anything we think speculation.

I suppose you could dismiss any thought or reasoning behind an argument for a belief as “reasons” to try to minimize them, but it’s kind of a weak argument position. You might consider instead justifying your beliefs, or saying why you disagree instead of just “yeah, well, that’s just, like, your opinion, man”.

Those aren’t contradictory. The Feds have an enormous budget for security, even just “traditional” security like everyone else uses for their systems, and not the “offensive security” we think of when we think “Federal security agencies”. Companies like Amazon, Microsoft, and Cisco will change products, build out large infrastructure, or even share the source code for their systems to persuade the feds to spend their money. They’ll do this because they have products that are valuable to the Feds in general, like AWS, or because they already have security products and services that are demonstrably valuable to the civil security sector.

OpenAI does not have a security product, they have a security problem. The same security problem as everyone else, that the NSA is in large part responsible for managing for significant parts of the government.
The government certainly has interest in AI technology, but OpenAI has productized their solutions with a different focus. They’ve already bought what everyone thinks OpenAI wants to build from Palantir.

So while it’s entirely possible that they are making a play to try to get those lines of communication to government decision makers for sales purposes, it seems more likely that they’re aiming to leverage “the guy who oversaw implementation of security protocol for military and key government services is now overseeing implementation of our security protocols, aren’t we secure and able to be trusted with your sensitive corporate data”.
If they were aiming for security productization and getting ties for that side of things, someone like Krebs would be more suitable, since CISA is a bit more well positioned for those ties to turn into early information about product recommendations and such.

So yeah, both of those statements are true. This is a non-event with bad optics if you’re looking for it to be bad.

Yeah, there are a ton of security experts. But none of them are the former head of the NSA.

Snowden is not exactly a font of expertise in this area, so I’m not sure that his opinion is particularly relevant. His only actual relevance is that he had access to classified data. He had no role in policy, and never had anything to do with business hiring practices.

It’s a bit of a non-story, beyond basic press release fodder.

In addition to it’s role as “digital panopticon”, they also have a legitimate role in cyber security assurance, and they’re perfectly good at it. The guy in question was the head of both the worlds largest surveillance entity, but also the world’s largest cyber security entity.
Opinions on the organization aside, that’s solid experience managing a security organization.
If open AI wants to make the case that they take security seriously, former head of the NSA, Cyber command and central security service as well as department director at a university and trustee at another university who has a couple masters degrees isn’t a bad way to try to send that message.

Other comments said open AI is the biggest scraping entity on the planet, but that pretty handily goes to Google, or more likely to the actual NSA, given the whole “digital panopticon” thing and “Google can’t fisa warrant the phone company”.

Joining boards so they can write memos to the CEO/dean/regent/chancellor is just what former high ranking government people do. The job aggressively selects for overactive Leslie Knope types who can’t sit still and feel the need to keep contributing, for good or bad, in whatever way they think is important.

If the US wanted to influence open AI in some way, they’d just pay them. The Feds budget is big enough that bigger companies will absolutely prostrate themselves for a sample of it. Or if they just wanted influence, they’d… pay them.
They wouldn’t do anything weird with retired or “retired” officers when a pile of money is much easier and less ambiguous.

At worst it’s open AI trying to buy some access to the security apparatus to get contracts. Seems less likely to me, since I don’t actually think they have anything valuable for that sector.

I mean, it does learn, it just lacks reasoning, common sense or rationality.
What it learns is what words should come next, with a very complex a nuanced way if deciding that can very plausibly mimic the things that it lacks, since the best sequence of next-words is very often coincidentally reasoned, rational or demonstrating common sense. Sometimes it’s just lies that fit with the form of a good answer though.

I’ve seen some people work on using it the right way, and it actually makes sense. It’s good at understanding what people are saying, and what type of response would fit best. So you let it decide that, and give it the ability to direct people to the information they’re looking for, without actually trying to reason about anything. It doesn’t know what your monthly sales average is, but it does know that a chart of data from the sales system filtered to your user, specific product and time range is a good response in this situation.

The only issue for Google insisting on jamming it into the search results is that their entire product was already just providing pointers to the “right” data.

What they should have done was left the “information summary” stuff to their role as “quick fact” lookup and only let it look at Wikipedia and curated lists of trusted sources (mayo clinic, CDC, national Park service, etc), and then given it the ability to ask clarifying questions about searches, like “are you looking for product recalls, or recall as a product feature?” which would then disambiguate the query.

It will confidently report inaccurate information. It’s usually not so hilariously wrong, but it’s still wrong.
For example, I was talking with someone about what constituents a “fruit” botanically, and I searched “are beans fruit”, and it confidently told me that beans are not a fruit, botanically speaking, because they’re a legume. It seems to have adapted, but that’s a good example of a “small wrong” that’s not uncommon at all.

That’s a bit of a difference in comparison.
A better comparison would be “does watching common heterosexual porn make common heterosexual men more interested in performing common heterosexual sexual acts?” or "does viewing pornography long term satiate a mans sex drive?” or “does consumption of nonconsensual pornography correlate to an increase in nonconsensual sex acts?”

Comparing “viewing child sexual content might lead it engaging in sexual acts with children” to “viewing sexual activity with women might lead to rape” is disingenuous and apples to oranges.

https://wchh.onlinelibrary.wiley.com/doi/full/10.1002/tre.791

a review of 19 studies published between 2013 and 2018 found an association between online porn use and earlier sexual debut, engaging with occasional and/or multiple partners, emulating risky sexual behaviours, assimilating distorted gender roles, dysfunctional body perception, aggression, anxiety, depression, and compulsive porn use.24 Another study has shown that compulsive use of sexually explicit internet material by adolescent boys is more likely in those with lower self-esteem, depressive feeling and excessive sexual interest.1

some porn use in adult men may have a positive impact by increasing libido and desire for a real-life partner, relieving sexual boredom, and improving sexual satisfaction by providing inspiration for real sex.7

As for child porn, it’s not a given that there’s no relationship between consumption and abusing children. There are studies that indicate both outcomes, and are made much more complicated by one of both activities being extremely illegal and socially stigmatized making accurate tracking difficult.
It’s difficult to justify the notion that “most pedophiles never offend” when it can be difficult to identify both pedophiles and abuse.

https://pubmed.ncbi.nlm.nih.gov/21088873/ for example. It looks at people arrested for possession of child pornography. Within six years, 6% were charged with a child contact crime. Likewise, you can find research with a differing conclusion

Point being, you can’t just hand wave the potential for a link away on the grounds that porn doesn’t cause rape amongst typical heterosexual men. There’s too many factors making the statistics difficult to gather.

It does learn from real images, but it doesn’t need real images of what it’s generating to produce related content.
As in, a network trained with no exposure to children is unlikely to be able to easily produce quality depictions of children. Without training on nudity, it’s unlikely to produce good results there as well.
However, if it knows both concepts it can combine them readily enough, similar to how you know the concept of “bicycle” and that of “Neptune” and can readily enough imagine “Neptune riding an old fashioned bicycle around the sun while flaunting it’s tophat”.

Under the hood, this type of AI is effectively a very sophisticated “error correction” system. It changes pixels in the image to try to “fix it” to matching the prompt, usually starting from a smear of random colors (static noise).
That’s how it’s able to combine different concepts from a wide range of images to create things it’s never seen.

The major concern to me, is that there isn’t really any guidance from the FBI on what you can and can’t do, which may lead to some big issues.

https://www.ic3.gov/Media/Y2024/PSA240329 https://www.justice.gov/criminal/criminal-ceos/citizens-guide-us-federal-law-child-pornography

They’ve actually issued warnings and guidance, and the law itself is pretty concise regarding what’s allowed.

(8) “child pornography” means any visual depiction, including any photograph, film, video, picture, or computer or computer-generated image or picture, whether made or produced by electronic, mechanical, or other means, of sexually explicit conduct, where-

(A) the production of such visual depiction involves the use of a minor engaging in sexually explicit conduct;

(B) such visual depiction is a digital image, computer image, or computer-generated image that is, or is indistinguishable from, that of a minor engaging in sexually explicit conduct; or

© such visual depiction has been created, adapted, or modified to appear that an identifiable minor is engaging in sexually explicit conduct.

…

(11) the term “indistinguishable” used with respect to a depiction, means virtually indistinguishable, in that the depiction is such that an ordinary person viewing the depiction would conclude that the depiction is of an actual minor engaged in sexually explicit conduct. This definition does not apply to depictions that are drawings, cartoons, sculptures, or paintings depicting minors or adults.

https://uscode.house.gov/view.xhtml?hl=false&edition=prelim&req=granuleid%3AUSC-prelim-title18-section2256&f=treesort&num=0

If you’re going to be doing grey area things you should do more than the five minutes of searching I did to find those honestly.

It was basically born out of a supreme Court case in the early 2000s regarding an earlier version of the law that went much further and banned anything that “appeared to be” or “was presented as” sexual content involving minors, regardless of context, and could have plausibly been used against young looking adult models, artistically significant paintings, or things like Romeo and Juliet, which are neither explicit nor vulgar but could be presented as involving child sexual activity. (Juliet’s 14 and it’s clearly labeled as a love story).
After the relevant provisions were struck down, a new law was passed that factored in the justices rationale and commentary about what would be acceptable and gave us our current system of “it has to have some redeeming value, or not involve actual children and plausibly not look like it involves actual children”.

Legally, a sufficiently detailed image depicting csam is csam, regardless of how it was produced. Sharing it is why he got caught, inevitably, but it’s still illegal even if he never brought a minor into it.

Fortunately most instances are in the category of a 17 year old to an 18 year old, and require parental consent and some manner of judicial approval, but the rates of “not that” are still much higher than one would want.
~300k in a 20 year window total, 74% of the older partner being 20 or younger, and 95% of the younger partner being 16 or 17, with only 14% accounting for both partners being under 18.

There’s still no reason for it in any case, and I’m glad to live in one of the states that said "nah, never needed .

In many ways they are. The image generated from a prompt isn’t unique, and is actually semi random. It’s not entirely in the users control. The person could argue “I described what I like but I wasn’t asking it for children, and I didn’t think they were fake images of children” and based purely on the image it could be difficult to argue that the image is not only “child-like” but actually depicts a child.

The prompt, however, very directly shows what the user was asking for in unambiguous terms, and the negative prompt removes any doubt that they thought they were getting depictions of adults.

I kinda like a little thingy that has a USB plug on one end under a cap, and a salt shaker on the other. Or maybe like a little hot sauce shaker, if we’re moving away from salt.

Yeah, look at image before title, wondered if it was a salt shaker and what language uses kalt for salt.

Hrm, USB stick salt shaker…

That’s a much better example.

Physical access to the device by a sophisticated attacker is well outside the realm of most people’s risk profile.

The point being that most people do not need to ever change their biometric data, because it isn’t used for remote authentication.

It’s about picking the right threat model, and for most people anything that gets them using the HSM is an improvement to their security.

That’s not retrieving the biometric data from the device, that’s retrieving the biometric data from surveillance or physical interaction.
It’s quite specifically the type of threat that most people do not need to worry about.