PeterisP

We don't want it both ways - if that's the price we'd have to pay, at least I definitely don't want model makers to license content.

I wonder if there are any invasive species between US states - like, some plant that's native to Florida but invasive in California, or native to California but invasive in Hawaii?

You could probably fire them with special artillery shells, just as deploying remote minefields or cluster munitions.

"Flagged as invasive" where? Every invasive species has a place where they're a perfectly nice local native.

The fact that we have to go as far back as CD players for a decent example illustrates my point - the "CD player" content distribution model is long dead, effectively nobody sells CD players or devices like CD players, effectively nobody distributes digital content on CDs or things like CDs (like, CD sales are even below vinyl sales) - almost every currently existing product receives content through a channel where updates could trivially be sent as well.

And if we're talking about how new products should be designed, then the "almost" goes away and they 100% wouldn't receive new content through anything like CDs, the issue transforms from an irrelevant niche (like CDs nowadays) to a nonexistent one.

I thing being continuously updated should be tied to receiving new external content.

It's fine to have an embedded device that's never updated, but never receives new content - it doesn't matter that a system won't be able to show a new emoji because it doesn't have any content that uses that new emoji.

However, if it is expected to display new and updated content from the internet, then the system itself has to be able to get updated and actually get updated, there's no acceptable excuses for that - if it's going to pull new content, it must also pull new updates for itself.

A key aspect is that nowadays we rarely encode pure text - while other encodings are more efficient for encoding pure Mandarin, nowadays a "Mandarin document" may be an HTML or JSON or XML file where less than half of the characters are from CJK codespace, and the rest come from all the formatting overhead which is in the 7-bit ASCII range, and UTF-8 works great for such combined content.

As far as I understand, one of things that it includes actually literally is the permutations of code adapted for every different model of supported nVidia hardware; that is a major (and desirable!) part of the driver+CUDA deployment.

It's a useless waste of resources - if there's no genuine dispute whether a violation occurred, in this situation, effective deterrence (which, as psychological research shows, is generally based on the likelihood of getting caught, not the size of punishment, so ticketing as many offences as possible is the key part) is the desired result, so why shouldn't society do it in the most efficient means possible, i.e. with as little wasted labor as possible?

If we want people to obey a certain restriction (especially a relatively minor one like parking in bus lanes), any obstacles that make enforcement time-consuming or inconvenient tend to result in it not being enforced and not being followed - and if we're okay with that, then we should instead totally repeal that restriction and any fines, as if it will be enforced only sporadically, then it will be turned into a tool for arbitrary punishment and not for actually getting that law followed - if we want the latter, we do want mass enforcement.

I think it is good and honest for companies to acknowledge that the current incentives work against trying to achieve [SOME_GOOD_RESULT], as in that case the first step towards getting that thing done would be figuring out how can we change the money flow so that there would actually be motivation for companies to direct their R&D efforts towards that.

If the society says to manufacturer "if you do this, you'll get less money from us" that is effectively the society saying "please don't do this". People and governments vote with their wallets about what they want, and if they're "voting wrong" or ineffectively, then they need to change what they pay for, instead of being surprised that they don't get what they disincentivize.

Perhaps it's different in USA due to different salary policies, but in European academia it sometimes happens that some of the leading positions are not actually that desirable, as you'll just sacrifice your scientific work (which you presumably love) to get bogged down with administrative duties instead. Like, our leadership roles are generally elected, so we can pick any one of us if we're unhappy with what the administration does, but I have seen this resulting even in literally drawing straws to choose who will apply to be the only candidate to be the head of an (sub)organization, because it's clear to everybody that we have to choose somebody to do it, but no one wants to be that one despite it technically being a promotion with slightly higher compensation. Perhaps rector is the role where this isn't relevant, since there you have multiple candidates who have already "sacrificed their life" to administration, but it can be the case for others.

This is the type of prohibition which we want to enforce because they disturb others, and it would be inappropriate to disturb the illegitimately disturbed person even more by having to stop their activities and wait for some official to resolve the problem.

If a car is blocking a bus stop, the harm is done even if the car leaves before some official arrives to write a ticket, and it seems reasonable that the bus operator should be able to request that the law gets enforced without having to stay on the spot and point out the violation to someone - we do want a deterrent effect towards any would-be violators, that they should fear a ticket even if there are no officials onsite, and perhaps not do the prohibited thing.

This is exactly what's being proposed - quoting the article, "Once a recording is made, it will be submitted to L.A. Department of Transportation where a human will assess whether a ticket should be issued."

In any case, neither the camera company nor the bus company have any authority to issue tickets, all they can do is submit complaints+evidence to the appropriate institution.

It depends on the location, my local automated speed camera process involves getting a human-approved ticket in the mail that includes the photos.

Quoting the official claims from the article, "Only when the system observes a vehicle parked illegally in a bus lane or a bus stop does it record the license plate and capture video of the event." - it seems very reasonable from both practical and privacy standpoint that all that data should not be collected unless there's some reason to believe (e.g. by such an "AI" system) that there likely is an actual violation.

For starters, you need to determine which cars are actually parked, and if that spot is violating the rules. For fixed cameras generally the installation process just marks as spot or lane where every car "counts", but a camera on a moving bus needs to do that on its own.

On the other side of the coin, if you freeze your dependencies and commit to not update whenever there's a newer version of any dep you are using, you commit yourself to having to continuously (and rapidly!) evaluate all the security advisories of all those dependencies to see if there are any bugs you have to mitigate before your system gets exploited.

You can't simply choose to never update any dependencies - the only question is how you decide when and which updates will get made, or delegate that decision to others.

For code, future is neither unknown nor certain - we often do have a valid, informed opinion about what kind of changes are likely (or even inevitable on a schedule) and what kind of changes are possible but unlikely.

For example, you may know for sure that some invariant is true in your market but not globally, so you know in advance that when(if!) your company expands, you'll need to change that thing in a specific way - but not in an arbitrary direction, that change is knowable even if you haven't yet spent the effort to know about its details.

Multisig requires user cooperation, many users will not care to cooperate, and chain of custody verification really starts working only if you can get (force) ~100% of legitimate users globally to adopt the system.

Also, for the potential creators of political fakes, such a multisig won't change things - getting a manufacturer's key may take some effort, but getting (and 'burning') keys of a dozen random people is relatively trivial in many ways - e.g. buying off of poor people, stealing from compromised random machines, or simply issuing fake identities for state-backed actors.

> In my specific case it was the bias to provide an answer to a question

That seems to be a reasonably expected result of the "instruction post-training" finetuning with RLHF or otherwise. If for some reason you don't want this behavior, you can avoid this by using a model version that just has the core language modeling without that finetuning, e.g. the llama models have such a version available.

There's a sufficient number of people personally doing targeted surveillance or pulling the trigger on weapons - and I'd assume that the barrier for building the tools for something is even lower than doing that yourself.

It would be interesting that a remote candidate could easily identify as whatever ethnicity, age or even gender they consider most beneficial for hiring to avoid discrimination or fit certain diversity incentives.

Tech like this has the potential to bring us back to the days of "on the Internet, nobody know's you're a dog" https://en.wikipedia.org/wiki/On_the_Internet,_nobody_knows_...

Quite the opposite, governments and large companies even explicitly run services for digital timestamping of documents - if I wanted to potentially assert some facts in court, I'd definitely prefer having that e-document with a timestamp notarized from my local government service instead of Bitcoin, because while the cryptography is the same, it would be much simpler from the practical legal perspective, requiring less time and effort and cost to get the court to accept that.

Just as before, a major part of photo or video evidence in court is not the actual video itself, but a person testifying "on that day I saw this horrible event, where these things happened, and here's attached evidence that I filmed which illustrates some details of what I saw." - which would be a valid consideration even without the photo/video, but the added details do obviously help.

Courts already wouldn't generally approve random footage without clear provenance.

I think it's practically impossible for such a system to be globally trustworthy due to the practical inevitability of "but improper storage of private keys lead to vulnerabilities" scenarios.

People will expect or require that chain of custody only if all or at least the vast majority of the content they want would have that chain of custody.

Photo/video content will have that chain of custody only if all or almost all of devices recording that content will support it - including all the cheapest mass-produced devices in reasonably widespread use anywhere in the world.

And that chain of custody provides the benefit only if literally 100% of these manufacturers have their private keys secure 100% of the time, which is simply not happening; at least one such key will leak, if not unintentionally then intentionally for some intelligence agency who wants to fake content.

And what do you do once you see a leak of the private keys used for signing the certificates for the private keys securely embedded in (for example) all of 2029 Huawei smartphones, which could be like 200 million phones? The users won't replace their phones just because of that, and you'll have all these users making content - so everyone will have to choose to either auto-block and discard everything from all those 200 million users, or permit content with a potentially fake chain of custody; and I'm totally certain that most people will prefer the latter.

I think the proper way is to start the consideration from the perspective of what the user can do, for example, from the four freedoms of FSF (https://www.gnu.org/philosophy/free-sw.html.en#four-freedoms)

The starting point is the ability to run the LLM as you wish, for any purpose - so if a license prohibits some uses and you have to start any usage with thinking whether it's permitted or not, that's a fail.

Then the freedom where "source" matters is the practical freedom to change the behavior so it does your computing as you wish. And that's a bit tricky - since one interpretation would require having the training data, training code and parameters; but for current LLMs the training hardware and cost of running it is a major practical limitation, so much that one could argue that the ability to change the behavior (which is the core freedom that we'd like) is separate from the ability to recreate the model, and would be more relevant in the context of the "instruction training" which happens after the main training, is the main determiner of behavior (as opposed to capability), and so the main "source would be the data for that (instruct training data, and the model weights before that finetuning) so that you can fine-tune the model on different instructions, which requires much less resources than training it from scratch, and don't have to start with the instructions and values imposed on the LLM by someone else.

While individual circumstances matter a lot for impact to a single organization, for estimating the global "blast impact" of a vulnerability, we'd expect that all of that would somewhat average out between all the thousands or millions of different systems and organizations running that tool, and a library with 100 times more users can be expected (all things being equal) to have 100-ish times more situations where it's used in a publicly reachable way - but the difference in average expected impact of having a million systems with a 2.5 CVSS vulnerability (likely causing zero incidents) and a million systems with a 10.0 CVSS vulnerability (likely wormable RCE) isn't 4 times, not even close to that.

It's probably close to the best thing that can be done with the available data, even if the available data is bad - the main flaw of such a metric is that while you could argue that the total impact is proportional to the number of machines affected, you can't reasonably argue that the impact is proportional to the CVSS score, the relationship is very non-linear.

A vulnerability with CVSS 2.5 wouldn't have a quarter of impact of a CVSS 10 vulnerability, its impact would be insignificant.

People have separate rights even if they are employees, and wiretapping laws apply to recording people - like, the company recording their own calls can easily be a violation of wiretapping laws if they don't properly inform their employees.

With a warrant, they definitely can legally record you without your consent and your knowledge, e.g. secretly invading your home and placing recording devices there - IMHO that's the point of the parent post; while there are checks and balances that require legal review, in the end, the government can do all that unilaterally after a judge approves that it is necessary.