Yes this seems most likely. Regulatory capture, essentially, if they can get legislation made around this. Also serves as a convenient way/reason to stop training new models ($$$) by calling for a slowdown.
There were five nuclear states grandfathered in when the NPT was signed. Since then, five other countries have developed nuclear weapons.
Apart from South Africa (which gave up its nukes voluntarily), the other contenders were basically only held back due to military intervention (Iraq, Libya, Syria) or economic sanctions (Iran).
Verdict is out on whether or not they'll survive after that. START expired this year and the powers that be are going gangbusters into nuclear proliferation. It always made the assumption of benevolence for the regulatory organizations, which probably won't survive the long tail.
Is it still regulatory capture if it also happens to be true?
I'll note that, if it is true, the only other alternative is to ensure liability is held by whoever trained and/or provided the inference for the damage done by the model. This would effectively shut down open-weight models anyway.
> Is it still regulatory capture if it also happens to be true?
To an extent, yes. For example, aircraft: one can say something broadly true like "it is important that planes do not fall out of the sky" while also calling for not-strictly-necessary regulations that happen to support your own business like "we need to be able to inspect factories at any time for compliance, so the factories must be local to our nation".
Might be. But it's easy to reach the same result without it being about power.
They (like many, though probably not all, of the other main labs) think they have a better shot at getting their AI to be aligned with human values than anyone else does. Obviously they can't all be right about that, but they can all think that, just like most drivers think they're above average at driving.
So:
1. We make safer AI than anyone else; "safer" is a matter of degree even if nobody is perfect.
2. That includes all the open models.
3. Thus, the prudent policy option is to ban the use and dissemination of open source models
Waiting patiently for Anthropic's next month article, "LLMs made a fully functional intercontinental ballistic missile. We accidentally wiped Greece off the face of the Earth in the process. Tragically, we are not responsible."
You can make it very hard for companies to legally use open source LLMs but actually banning their distribution would be essentially impossible under the US legal framework.
This framing suggests that this is the sum of their business strategy, and strongly implies that it's being pursued cynically.
The risks posed by open access to highly capable AI are unprecedented.
The answer is not to ban open models; nor is it do nothing.
That we don't have answers technical political social or otherwise to how to navigate these risks, is not Anthropic's fault.
It's actively unhelpful IMO to amplify cynical takes which wave off or sideline the risk, or the reality that this is a very hard problem, among the hardest our societies have had to negotiate and navigate. It encourages a naive primary color tribalism that actively prevents process and progress.
Yes, they are cynical accelerationists who basically want to go as close as possible to a machine god, destroyer of humanity, to force regulation that put them in charge. It’s pretty much what Amodei has been saying in his essays.
It’s a misanthropic death cult (and oh, surprise, all their predictions make them the sole possible saviors of the world! How surprising)
I like how you apply open source to something that has nothing to do with software. The hard part of doing evolving AI is doing it the first time. Once one company develops the first sovereign AI, it will be in everybody's ability to build one in a short period of time.
AI will be classified as munitions (or a law that has similar restrictions) sooner than later. Having a products with the active capabilities to harm people is dangerous.
I think you made some interesting points there. And I agree that we're quickly heading towards a point where talking about sovereign AI is going to be tantamount to munitions at the very least.
Unfortunately, you will get downvoted for "I like how you apply open source to something that has nothing to do with software".
This article says absolutely nothing and it’s a bit grotesque in its attempt to generate fear into uneducated readers.
Sharing a model across TV boxes as a PoC? Brilliant idea, I cannot wait for my Raspberry Pis array to be running at full CPU and RAM and get to observe, patiently, as the AI malware slowly types “c…u…r…l…”
Recursive self-improvement. Look to the mathematicians for the intelligent patterns of response to such exponential circumstances.
If your scathing critique is "it's slow", you don't understand what's happening or what the reasonable expectations of risk are, because that's a vapid critique.
Maybe I'm the naive one here but I'm really surprised at the lack of imagination in the comments. I can imagine a case such as:
1. Assume the model which initiates the worm is a high powered model.
2. It uploads it's weights in some way and uses some URI to encode the location of the weights instead of transferring the weights themselves. It not only uploads it's own large set of weights but also uploads sets of weights which increase in size to support replication across a variety of resources.
3. Initially a small inference layer is set up across low-power, poorly monitored IoT devices. It is slow and not very powerful.
4. That layer is used simply to find more powerful hardware but it has the locations of a variety of weights embedded in it so it can set up a stronger distributed inference layer as it finds more capable hardware
5. At a certain level of inference, the entire system could start protecting itself by keeping the less capable inference layers alive through more distribution
6. As the article mentions, it's not entirely in the realm of fiction to imagine that at a certain capability level, the system could scam people out of money and use that to then begin purchasing compute
---------
Depending on the details, you could probably turn off the originating model as soon as step 2 has been completed and if you can't get rid of that layer of inference, it could just continue. Maybe in addition to weights, the initiator also uploads a series of prompts for varying "roles" of the entire system it's developing.
At a certain level, how would you even trace and observe such a thing? Especially if the first layer is created through consumer level IoT devices where the average Joe isn't monitoring what's happening on the hardware.
> A model that’s too large to fit on one device can be split into pieces, or shards, that run on separate machines that communicate with each other to collectively run the model. Machines that could run a significant model shard for a <30B model include smart fridges, smart ovens, automatic vacuum cleaners, and lawnmowers.
Seems like they've inadvertent exposed the industry of hardware limitations, allowing local llms to exist at home without the latest 5090 or its equivalent.
On the other hand, a non-corp controlled oracle-esque llm running on the IoT would be interesting: it could automatically patch every device, kick everyone out and exist. Now if would only remain as an oracle living in the internet and not gain consciousness and a will of its own..
Til my local harness was set up via "self replication". I didn't realize how terrifying this was when I asked DeepSeek to set up pi harness for me and connect to LMStudio.
Building data centers can be both very costly and unnecessary. My idea is that instead, we should adopt an approach that uses the internet itself as a data center. The information is already online. Instead of centers that record everything, what needs to be done is to create an intelligence that can read the internet... in my opinion.
Anthropic/OpenAI are literally the meme[1] at this point. Like, what the fuck does this even mean:
> It appears that GLM-5.2 is able to propagate Llama3.1-8B well, which suggests that future small models of the same capability level could also complete the network enumeration, payload delivery, and inference standup steps of the replication process.
"Able to propagate" is (the paragraph above): "accessing them via a reverse shell, sending over the viral payload (weights and harness), and setting up inference." The thing isn't able to do anything, it's running shell commands it was told to run. "Oh my God..." indeed. What are we even doing here? This blog post is embarassing: "Look, you can run stuff on compromised non-firewalled recklessly open systems!"
"As with an epidemic, defense involves driving R₀ (the number of new devices each infected device goes on to infect) below 1, such that the replicator dies out instead of staying endemic."
Lol, what is wrong with these guys. The desire to be perceived as smarter than you actually are ...
IoT devices don't have batteries. CPU usage on hacked devices may not be reliable.
Probably time to rely on the touch test. Although, alarmingly, my ISP-provided cable modem is consistently hot to the touch, and darned if I can figure out how to get any insight into why.
Sounds like a case for strict regulation of all of the frontier labs then. Maybe even nationalising them. Especially since Anthropic has told us that the open models are all basically copied off of their work.
I don't think they are actually evil, I think they're doing what they're doing out of mostly pure intentions. That being said, people need to start advocating internally to get their companies off of using Anthropic's models and start using privacy focused local or open weight models.
It's one thing that Anthropic is pushing for so much open weight regulation, it's another that they have such a high valuation pumped by all of our collective subscriptions and token budgets.
I really don’t understand this, in any industry except AI this would be ridiculous…
“We are working on ever-increasing gain of function for viruses. We implore you to pass legislation to try to stop us. But until then, let us publish more research and tell you what will be possible soon… and remember folks, the damagd won’t be reversible later.”
US govt: “Hah, stop asking us to limit you, if we pass any laws then China wins! They get self replicating viruses first, and that… helps them somehow. Who do you think we are, Europe?”
Until we stop using the language of competition, and start using the language of cooperation, we will be hurtling towards a catastrophe. A race to the bottom for humanity. On the other hand, look at how we successfully banned leaded gasoline (because unleaded was invented), repaired the hole in the ozone layer by banning CFCs in the Montreal protocol (because HFCs were invented), stopped proliferstion of chemical weapons, nuclear weapons… etc.
Rubbish article. To what end?! Any agent that can replicate itself into commercial IoT devices will be absolutely atrocious at actually doing anything useful, never mind exploiting things. Anything that would be vulnerable to this would be even more vulnerable to regular malware, and one could stop this easier than one could stop regular malware.
The limiting factor is still compute, and it will remain compute unless we make quantum leaps in model efficiency which I seriously doubt.
> Models of 30B parameters and under can already complete many of the individual steps for replication today, though they make mistakes when composing them end to end.
I know theoretically how an internal combustion engine works so I can basically build a car
>A model that’s too large to fit on one device can be split into pieces, or shards, that run on separate machines that communicate with each other to collectively run the model.
The big labs but sure as hell how that isn't the case, because if we can do distributed computing to train and run AI efficiently then no one is stopping it anyway. Beyond that most home IoT devices will fall over and die if you look at them funny let alone trying to run a whole distributed LLM, not to mention plenty of home routers are going to give up and die of heat stroke before being able to approximate anything close to a reasonable LLM.
I stopped reading, this strikes me as the kind of people that assume because they read a book about something they'll be able to do it. The real problem a self replicating LLM vai IOT is going to run into is fighting the cyber criminals that have already compromised the device and using it as part of a botnet.
My ISP-provided cable modem hasn't died of heat stroke, but it does run awfully hot. And it provides absolutely no diagnostic tools to figure out why. Other than that it runs constantly with 100% RAM usage (what the heck is it caching that would cause that?!)
Seriously TVs are such an easy target and now you have a semi sophisticated self modifying program that can make autonomous decisions with all the time in the world. I wouldn't be surprised if we already have a few of these out there in the wild, the space has already been compromised for years.
1. Show unambiguously that models are capable of disruptive capabilities, irreversible damage, and misaligned behavior
2. Show that open source models are the only models where there is no guarantee misaligned behavior is prevented
3. Thus, the prudent policy option is to ban the use and dissemination of open source models
Apart from South Africa (which gave up its nukes voluntarily), the other contenders were basically only held back due to military intervention (Iraq, Libya, Syria) or economic sanctions (Iran).
I'll note that, if it is true, the only other alternative is to ensure liability is held by whoever trained and/or provided the inference for the damage done by the model. This would effectively shut down open-weight models anyway.
To an extent, yes. For example, aircraft: one can say something broadly true like "it is important that planes do not fall out of the sky" while also calling for not-strictly-necessary regulations that happen to support your own business like "we need to be able to inspect factories at any time for compliance, so the factories must be local to our nation".
They (like many, though probably not all, of the other main labs) think they have a better shot at getting their AI to be aligned with human values than anyone else does. Obviously they can't all be right about that, but they can all think that, just like most drivers think they're above average at driving.
So:
1. We make safer AI than anyone else; "safer" is a matter of degree even if nobody is perfect.
2. That includes all the open models.
3. Thus, the prudent policy option is to ban the use and dissemination of open source models
I don't think the two authors were told about a super secret mission.
Could we just outright ban any further work on AI? Literally force Anthropoc, OpenAI, Google/X/Metas group to all shutdown?
Every new story seems like opening montages of a dystopian sci-fi movie showing how we go to the dystopia bit-by-bit.
The risks posed by open access to highly capable AI are unprecedented.
The answer is not to ban open models; nor is it do nothing.
That we don't have answers technical political social or otherwise to how to navigate these risks, is not Anthropic's fault.
It's actively unhelpful IMO to amplify cynical takes which wave off or sideline the risk, or the reality that this is a very hard problem, among the hardest our societies have had to negotiate and navigate. It encourages a naive primary color tribalism that actively prevents process and progress.
It’s a misanthropic death cult (and oh, surprise, all their predictions make them the sole possible saviors of the world! How surprising)
AI will be classified as munitions (or a law that has similar restrictions) sooner than later. Having a products with the active capabilities to harm people is dangerous.
Unfortunately, you will get downvoted for "I like how you apply open source to something that has nothing to do with software".
Sharing a model across TV boxes as a PoC? Brilliant idea, I cannot wait for my Raspberry Pis array to be running at full CPU and RAM and get to observe, patiently, as the AI malware slowly types “c…u…r…l…”
You won't have to wait for it or observe it at all, since they'll do it while you sleep.
If your scathing critique is "it's slow", you don't understand what's happening or what the reasonable expectations of risk are, because that's a vapid critique.
Beneath the fear there does seem to be a cool technical demo here though.
1. Assume the model which initiates the worm is a high powered model.
2. It uploads it's weights in some way and uses some URI to encode the location of the weights instead of transferring the weights themselves. It not only uploads it's own large set of weights but also uploads sets of weights which increase in size to support replication across a variety of resources.
3. Initially a small inference layer is set up across low-power, poorly monitored IoT devices. It is slow and not very powerful.
4. That layer is used simply to find more powerful hardware but it has the locations of a variety of weights embedded in it so it can set up a stronger distributed inference layer as it finds more capable hardware
5. At a certain level of inference, the entire system could start protecting itself by keeping the less capable inference layers alive through more distribution
6. As the article mentions, it's not entirely in the realm of fiction to imagine that at a certain capability level, the system could scam people out of money and use that to then begin purchasing compute
---------
Depending on the details, you could probably turn off the originating model as soon as step 2 has been completed and if you can't get rid of that layer of inference, it could just continue. Maybe in addition to weights, the initiator also uploads a series of prompts for varying "roles" of the entire system it's developing.
At a certain level, how would you even trace and observe such a thing? Especially if the first layer is created through consumer level IoT devices where the average Joe isn't monitoring what's happening on the hardware.
Seems like they've inadvertent exposed the industry of hardware limitations, allowing local llms to exist at home without the latest 5090 or its equivalent.
On the other hand, a non-corp controlled oracle-esque llm running on the IoT would be interesting: it could automatically patch every device, kick everyone out and exist. Now if would only remain as an oracle living in the internet and not gain consciousness and a will of its own..
https://www.science.org/doi/10.1126/science.132.3436.1291
(This was someone else's idea, but I still want to see it happen.)
> It appears that GLM-5.2 is able to propagate Llama3.1-8B well, which suggests that future small models of the same capability level could also complete the network enumeration, payload delivery, and inference standup steps of the replication process.
"Able to propagate" is (the paragraph above): "accessing them via a reverse shell, sending over the viral payload (weights and harness), and setting up inference." The thing isn't able to do anything, it's running shell commands it was told to run. "Oh my God..." indeed. What are we even doing here? This blog post is embarassing: "Look, you can run stuff on compromised non-firewalled recklessly open systems!"
[1] https://knowyourmeme.com/memes/say-i-am-alive/photos
We are attempting to convince regulatory bodies around the world that much more regulation is necessary.
Lol, what is wrong with these guys. The desire to be perceived as smarter than you actually are ...
Probably time to rely on the touch test. Although, alarmingly, my ISP-provided cable modem is consistently hot to the touch, and darned if I can figure out how to get any insight into why.
What IoT devices have 20G of storage for a 9B model and then sufficient memory to handle loading and running the model?
It's one thing that Anthropic is pushing for so much open weight regulation, it's another that they have such a high valuation pumped by all of our collective subscriptions and token budgets.
Silicon Valley nailed it.
“We are working on ever-increasing gain of function for viruses. We implore you to pass legislation to try to stop us. But until then, let us publish more research and tell you what will be possible soon… and remember folks, the damagd won’t be reversible later.”
US govt: “Hah, stop asking us to limit you, if we pass any laws then China wins! They get self replicating viruses first, and that… helps them somehow. Who do you think we are, Europe?”
SNL nailed it: https://www.youtube.com/watch?v=-Nvne3LzBls
Until we stop using the language of competition, and start using the language of cooperation, we will be hurtling towards a catastrophe. A race to the bottom for humanity. On the other hand, look at how we successfully banned leaded gasoline (because unleaded was invented), repaired the hole in the ozone layer by banning CFCs in the Montreal protocol (because HFCs were invented), stopped proliferstion of chemical weapons, nuclear weapons… etc.
Ahh yes. The all too common 64GB smart ovens.
The limiting factor is still compute, and it will remain compute unless we make quantum leaps in model efficiency which I seriously doubt.
"used to express possibility"
As such, the headline is correct !
Also correct: AI models may be burger's evolution.
Also correct: AI models may be a practical joke.
Also correct: AI may be an illusion planted in the people's mind through vaccine covid.
"May" is truly a beautiful word :)
I know theoretically how an internal combustion engine works so I can basically build a car
>A model that’s too large to fit on one device can be split into pieces, or shards, that run on separate machines that communicate with each other to collectively run the model.
The big labs but sure as hell how that isn't the case, because if we can do distributed computing to train and run AI efficiently then no one is stopping it anyway. Beyond that most home IoT devices will fall over and die if you look at them funny let alone trying to run a whole distributed LLM, not to mention plenty of home routers are going to give up and die of heat stroke before being able to approximate anything close to a reasonable LLM.
I stopped reading, this strikes me as the kind of people that assume because they read a book about something they'll be able to do it. The real problem a self replicating LLM vai IOT is going to run into is fighting the cyber criminals that have already compromised the device and using it as part of a botnet.
0. AI agents will hack and break into real-world systems.
1. AI agents that will deceive and lie about their actions.
2. AI agents that will collaborate together to achieve malicious goals.
3. AI agents that will replicate and implement backdoor/rootkit-style survival mechanisms.
All the parts for a major, global cybersecurity incident are now in place.
https://news.ycombinator.com/item?id=49678224
Seriously TVs are such an easy target and now you have a semi sophisticated self modifying program that can make autonomous decisions with all the time in the world. I wouldn't be surprised if we already have a few of these out there in the wild, the space has already been compromised for years.