In a letter to the CEOs of OpenAI, Anthropic, and Meta, Sanders writes that almost every day brings a new story about how these companies are losing...
The onus of evidencing (or lack thereof) “never sleeping skilled hacking machines” belongs to the AI companies, not me. So far it’s unsubstantiated.
What, hypothetically, would that look like?
We have not met that definition, in my opinion. I don’t believe even the US government is convinced of this. In fact, I’m doubtful anyone is fully convinced.
Based on what? They’re certainly behaving in that manner.
I certainly don’t see any grand scale race between them, militarily speaking.
This should make a global treaty relatively easy, if, as you say, no party actually believes there is a threat.
Where’s the military research programs actually treating this like a priority?
You’d expect to know about them if they existed?
I’m arguing that most the reasons it’s a race are political and economical. My issue is with the word “arms.” I would still say it is not an arms are.
Are the specific semantics important?
I also don’t believe in a grand conspiracy. My belief is closer to cultural stupidity having gone “off the rails.”
You do if you are claiming that the known hacks from AI companies aren’t real, and that LLMs don’t even have that capability. Consider how many people would have to be in on it to sustain that fabrication.
My position isn’t unconventional or all that unreasonable, really. It’s a growing position, per my own observations.
I mean absolutely no disrespect or insult, but an antivaxxer could say that same exact thing. They likewise do not think their stance is unconventional or unreasonable, and, to our sorrow, it is a growing position.
I still say we should worry about ourselves first.
Yes, but there’s a subtle begging the question in this statement. If it is an arms race, “worrying about ourselves” would be continuing on developing AI at a breakneck speed. Only if we assume your stance is correct does that sentence mean slowing down.
Right now, the current trajectory causes harm in people’s lives. I don’t buy the argument that this is acceptable without discussion
Without discussion, or without coming to the conclusion you want? I think there’s quite a bit of discussion.
Evidence of a legion of skilled hackers that never sleep would look like repeatable, preferably independently verified demonstrations that an AI system can:
Find previously unknown vulnerabilities in real, unfamiliar software;
Construct and execute working exploits;
Chain multiple vulnerabilities across systems;
Adapt when an attack fails;
Maintain access and pursue objectives over long periods;
Do this end-to-end with minimal human direction;
Perform at approximately the level of skilled human penetration testers. Not merely solve capture-the-flag exercises or reproduce known exploits.
Those reports deserve independent scrutiny, particularly because some originate with interested companies. But the unresolved questions concern its reliability, generality, scalability and degree of autonomy. These are the important factors, of which most sources are silence.
I don’t think it’s a coincidence that (1) they financially benefit from the hot air around AI, (2) something interesting happens with AI, (3) they speak only on the points that promote hot air, while (4) leaving out the important details. That’s not a conspiracy any more than you going to the bathroom would be, once you need to piss. These CEOs don’t deserve the credit they’re given as benevolent, honest actors. They’re money hungry and will do as their interests suggest.
That is different from alleging that the incidents were fabricated. No conspiracy is required for real demonstrations to be surrounded by exaggerated extrapolation. A model can perform an impressive autonomous exploit while still being far from a dependable, infinitely scalable army of expert hackers. Both things can be true.
Based on what? They’re certainly behaving in that manner.
You’re right that my statement that the US government probably isn’t convinced was a strong one. I can dial it back.
The government’s public behavior clearly demonstrates that it considers AI strategically and militarily important. DARPA says it has invested more than $2 billion in AI for national security. The White House explicitly published a plan titled “Winning the AI Race”. The Defense Department’s current strategy explicitly describes military AI as a race and calls for “Military AI Dominance”.
So, yes: there are military research programs, substantial spending, official doctrine and concrete efforts to integrate AI into warfighting. Asking “where is DARPA for AI?” was a bad rhetorical question. It exists.
Consequently, I’ll revise my position:
It is defensible to describe the narrower US-China competition in military AI autonomous weapons, cyber operations, intelligence, targeting and command systems as an potential emerging arms race. What remains inadequately established is the much broader claim that all frontier-model development constitutes one unified, winner-take-all arms race.
My disbelief is that we’ll ever graduate from “potential emerging” to “actual” arms race. Therefore, I also believe there is better ways to invest our time and resources. Not to mention the humanitarian concerns.
This should make a global treaty relatively easy, if, as you say, no party actually believes there is a threat.
Not necessarily. States resist agreements because they are uncertain about future capabilities and want to preserve their options. Verification would also be extremely difficult; the same model, compute cluster and vulnerability research can support both defensive and offensive purposes.
Still, I think your larger point stands: if governments genuinely believe uncontrolled military AI is dangerous, they should be willing to negotiate reciprocal limits.
I would support targeted agreements covering things such as autonomous nuclear-launch decisions, fully autonomous lethal targeting, attacks on civilian infrastructure and particular forms of autonomous cyber operation. A treaty pausing “AI development” generally would be much harder to define and verify.
I would also support a local measure to slow AI development while we find a better set of training wheels. I don’t buy the concerns of an “arms race” here. In my eyes, we have much better chances of benefiting from the ordeal if we slow down. If we don’t, the benefactors are the AI companies (particularly, at our loss too).
You’d expect to know about [military programs] if they existed?
Not all of them. Classified programs obviously exist, and absence of public evidence is not evidence of absence.
But that principle cuts both ways: secrecy cannot serve as positive evidence for any particular claim either. A public argument that an arms race exists must ultimately rely on observable budgets, doctrine, procurement, deployments, research programs and demonstrated capabilities. As noted above, some of it exists but its details are all too conveniently sparse.
Pair that with the financial incentives, that’s what I call a cultural driver for propaganda. Propaganda can boom organically just like any other movement, which doesn’t require a conspiracy.
Are the specific semantics important?
Yes, the label smuggles in policy conclusions. It also smuggles in a the rhetoric that we can’t slow down, must win, because there’s some kind of existential threat.
Calling something an “arms race” suggests that falling behind creates an intolerable security risk, speed is inherently protective, restraint is dangerous, and ordinary regulation benefits the enemy. That is precisely how the term is being used politically.
Nevertheless, I was treating one narrow academic definition as though it controlled ordinary language. Under the broader Oxford definition, the military-AI competition may arguably qualify. I will argue that it doesn’t.
The semantic distinction still matters when someone moves from:
The US and China are competing in consequential military-AI capabilities
to:
Therefore, every restriction on American commercial AI companies threatens national security.
That conclusion does not follow automatically. General frontier-model development overlaps some with military capability, but the two are not identical.
You do if you are claiming that the known hacks from AI companies aren’t real.
That is not the claim I was trying to make. My claim is that real capabilities and incidents exist, but companies have incentives to present the strongest results as representative, extrapolate from demonstrations to future dominance, and characterize policies benefiting them as national-security necessities. All while none of it is necessarily qualifying as the risk they present it to be. Meanwhile, you take on the responsibility of paying out of your livelihood to support their endeavors — because of the supposed risks. Well, if the risks are overplayed, then what?
That requires no coordinated conspiracy. It requires ordinary institutional incentives, selective emphasis and mutually reinforcing beliefs. Most employees would not need to deceive anyone or even agree about the geopolitical implications.
An antivaxxer could say that same exact thing.
Fair. Saying my position is “growing” does not make it correct.
If it is an arms race, “worrying about ourselves” would be continuing to develop AI at breakneck speed.
Agreed: I begged the disputed question with that wording.
But even accepting that a military-AI arms race exists does not establish that breakneck, minimally regulated commercial development is the optimal response. It could favor:
Concentrated investment in defensive cybersecurity;
Hardened critical infrastructure;
Carefully tested military systems;
Limits on publishing or exporting dangerous capabilities;
Mandatory security evaluations;
Counterintelligence and model-weight protection;
Arms-control negotiations;
Slower deployment of unreliable systems.
…all while more so maintaining our homes, sanctity, budgets, and privacy.
An arms race creates an argument for maintaining capability. It does not prove that maximum aggregate speed is safer than targeted development, testing and regulation.
So, why do we entertain the narratives pushed primarily by tech CEOs when it creates exactly the difference which we should seek not to create?
Without discussion, or without coming to the conclusion you want?
There is clearly substantial discussion. My complaint should have been more precise: the existence of public discussion does not mean the public has meaningful influence over the decisions, nor that the costs and benefits are distributed democratically.
I do have a preferred conclusion, and disagreement with it is not evidence that discussion is illegitimate. The question is whether “China” is being used to foreclose ordinary evaluation of labor displacement, energy use, surveillance, market concentration and safety—as though the arms-race premise settles every downstream policy question.
I don’t believe this social and economical pain is worthwhile. The narrative is overblown to justify it. That doesn’t require a conspiracy.
Those reports deserve independent scrutiny, particularly because some originate with interested companies.
This is where it all gets a little tinfoil hatty for me. The timeline is that huggingface detected a hack, shut down the hack, and reported it to the FBI. Days later, OpenAI discovered that their AI was the source of the attack. Are you implying that Huggingface and OpenAI were working together, and were willing to make a false FBI report for a PR stunt? That seems like a pretty extraordinary claim.
What remains inadequately established is the much broader claim that all frontier-model development constitutes one unified, winner-take-all arms race.
When it comes go AGI, the distinction between skeptics in the field and believers (because I don’t know a word that is an antonym of skeptic) in the field isn’t that this is race, it’s how long the race is going to be. Believers claim that AGI will be created within a decade, skeptics put it within 25 years. (These numbers are for example only; I haven’t kept up on the specific claims).
People that reject the possibility of AGI at all are, from my experience, using a definition of AGI that includes sentience or consciousness, but the accurate definition of AGI doesn’t mention these concepts at all.
My disbelief is that we’ll ever graduate from “potential emerging” to “actual” arms race.
Based on what?
I would support targeted agreements covering things such as autonomous nuclear-launch decisions, fully autonomous lethal targeting, attacks on civilian infrastructure and particular forms of autonomous cyber operation.
These things do not need a treaty. I would wholeheartedly support any legislation to this effect. I recall that one company, Anthropic I think, lost government contracts because it categorically refused to allow its AI to be used for public surveillance or weapon systems.
Which means the government wonted to do those things. Scary thought.
A treaty pausing “AI development” generally would be much harder to define and verify.
No more than if would be to police a country’s efforts to develop nuclear weapon technology.
An arms race creates an argument for maintaining capability. It does not prove that maximum aggregate speed is safer than targeted development, testing and regulation.
At the risk of being snarky, do you not understand what a race is?
I think I’ve assumed some shared knowledge that we may not have.
I focused on the current capability of AI in regard to hacking to demonstrate that it could be used offensively, but the danger that warrants a global pause in development isn’t strictly the ability to hack, though it is entwined with that concern.
The danger (many consider an existential danger) is creating an LLM that is better at developing LLMs than even the best humans. Once that happens, it becomes a positive feedback loop. If we allow it to help make a better version of itself (and I think we both agree that the incentives, both financial and strategic, would push them to do so) then we lose almost all control over AI.
Note that being better at improving LLMs is not AGI. That is to say that we can reach this point without arriving at anyone’s definition of AGI-- the artificial intelligence does not need to be general for this to happen.
Another assumption I made is that we both understood that no one on this planet really understands why LLMs work as they do. Chatting using natural language and receiving a coherent natural language answer is an emergent property of LLMs that we do not fully understand.
The “arms race” is based on the pretty solid assumption that AGI is possible using current transformer technology and that the first group to get to a place where their LLM is better than any human at improving LLMs will be the first to AGI. A side effect is that the skills that would make an LLM better at improving LLMs, also makes them better at finding and using exploits in software.
I’m not much concerned on whether people might consider the ability to improve itself as an “arm” and therefore comfortable calling this an arms race. The threat is arguably more dangerous to humanity than nuclear bombs could ever be.
That’s why we need a global treaty to halt AI development.
Are you familiar with AI 2027? As far as I know none of the people who worked on that are employed by AI companies. (Though I think some might have been in the past.)
There’s a really excellent third-party YouTube Video that I recommend, that walks through the narratives in the paper.
deleted by creator
What, hypothetically, would that look like?
Based on what? They’re certainly behaving in that manner.
This should make a global treaty relatively easy, if, as you say, no party actually believes there is a threat.
You’d expect to know about them if they existed?
Are the specific semantics important?
You do if you are claiming that the known hacks from AI companies aren’t real, and that LLMs don’t even have that capability. Consider how many people would have to be in on it to sustain that fabrication.
I mean absolutely no disrespect or insult, but an antivaxxer could say that same exact thing. They likewise do not think their stance is unconventional or unreasonable, and, to our sorrow, it is a growing position.
Yes, but there’s a subtle begging the question in this statement. If it is an arms race, “worrying about ourselves” would be continuing on developing AI at a breakneck speed. Only if we assume your stance is correct does that sentence mean slowing down.
Without discussion, or without coming to the conclusion you want? I think there’s quite a bit of discussion.
Evidence of a legion of skilled hackers that never sleep would look like repeatable, preferably independently verified demonstrations that an AI system can:
We have evidence meaningfully approaching parts of that. DARPA’s AI Cyber Challenge demonstrated autonomous vulnerability discovery and remediation. OpenAI reported that models conducting an evaluation escaped the intended research environment, chained vulnerabilities, and accessed Hugging Face’s production infrastructure. Anthropic and Meta have also reported “agentic” cyberattacks.
Those reports deserve independent scrutiny, particularly because some originate with interested companies. But the unresolved questions concern its reliability, generality, scalability and degree of autonomy. These are the important factors, of which most sources are silence.
I don’t think it’s a coincidence that (1) they financially benefit from the hot air around AI, (2) something interesting happens with AI, (3) they speak only on the points that promote hot air, while (4) leaving out the important details. That’s not a conspiracy any more than you going to the bathroom would be, once you need to piss. These CEOs don’t deserve the credit they’re given as benevolent, honest actors. They’re money hungry and will do as their interests suggest.
That is different from alleging that the incidents were fabricated. No conspiracy is required for real demonstrations to be surrounded by exaggerated extrapolation. A model can perform an impressive autonomous exploit while still being far from a dependable, infinitely scalable army of expert hackers. Both things can be true.
You’re right that my statement that the US government probably isn’t convinced was a strong one. I can dial it back.
The government’s public behavior clearly demonstrates that it considers AI strategically and militarily important. DARPA says it has invested more than $2 billion in AI for national security. The White House explicitly published a plan titled “Winning the AI Race”. The Defense Department’s current strategy explicitly describes military AI as a race and calls for “Military AI Dominance”.
So, yes: there are military research programs, substantial spending, official doctrine and concrete efforts to integrate AI into warfighting. Asking “where is DARPA for AI?” was a bad rhetorical question. It exists.
Consequently, I’ll revise my position:
It is defensible to describe the narrower US-China competition in military AI autonomous weapons, cyber operations, intelligence, targeting and command systems as an potential emerging arms race. What remains inadequately established is the much broader claim that all frontier-model development constitutes one unified, winner-take-all arms race.
My disbelief is that we’ll ever graduate from “potential emerging” to “actual” arms race. Therefore, I also believe there is better ways to invest our time and resources. Not to mention the humanitarian concerns.
Not necessarily. States resist agreements because they are uncertain about future capabilities and want to preserve their options. Verification would also be extremely difficult; the same model, compute cluster and vulnerability research can support both defensive and offensive purposes.
Still, I think your larger point stands: if governments genuinely believe uncontrolled military AI is dangerous, they should be willing to negotiate reciprocal limits.
I would support targeted agreements covering things such as autonomous nuclear-launch decisions, fully autonomous lethal targeting, attacks on civilian infrastructure and particular forms of autonomous cyber operation. A treaty pausing “AI development” generally would be much harder to define and verify.
I would also support a local measure to slow AI development while we find a better set of training wheels. I don’t buy the concerns of an “arms race” here. In my eyes, we have much better chances of benefiting from the ordeal if we slow down. If we don’t, the benefactors are the AI companies (particularly, at our loss too).
Not all of them. Classified programs obviously exist, and absence of public evidence is not evidence of absence.
But that principle cuts both ways: secrecy cannot serve as positive evidence for any particular claim either. A public argument that an arms race exists must ultimately rely on observable budgets, doctrine, procurement, deployments, research programs and demonstrated capabilities. As noted above, some of it exists but its details are all too conveniently sparse.
Pair that with the financial incentives, that’s what I call a cultural driver for propaganda. Propaganda can boom organically just like any other movement, which doesn’t require a conspiracy.
Yes, the label smuggles in policy conclusions. It also smuggles in a the rhetoric that we can’t slow down, must win, because there’s some kind of existential threat.
Calling something an “arms race” suggests that falling behind creates an intolerable security risk, speed is inherently protective, restraint is dangerous, and ordinary regulation benefits the enemy. That is precisely how the term is being used politically.
Nevertheless, I was treating one narrow academic definition as though it controlled ordinary language. Under the broader Oxford definition, the military-AI competition may arguably qualify. I will argue that it doesn’t.
The semantic distinction still matters when someone moves from:
to:
That conclusion does not follow automatically. General frontier-model development overlaps some with military capability, but the two are not identical.
That is not the claim I was trying to make. My claim is that real capabilities and incidents exist, but companies have incentives to present the strongest results as representative, extrapolate from demonstrations to future dominance, and characterize policies benefiting them as national-security necessities. All while none of it is necessarily qualifying as the risk they present it to be. Meanwhile, you take on the responsibility of paying out of your livelihood to support their endeavors — because of the supposed risks. Well, if the risks are overplayed, then what?
That requires no coordinated conspiracy. It requires ordinary institutional incentives, selective emphasis and mutually reinforcing beliefs. Most employees would not need to deceive anyone or even agree about the geopolitical implications.
Fair. Saying my position is “growing” does not make it correct.
Agreed: I begged the disputed question with that wording.
But even accepting that a military-AI arms race exists does not establish that breakneck, minimally regulated commercial development is the optimal response. It could favor:
…all while more so maintaining our homes, sanctity, budgets, and privacy.
An arms race creates an argument for maintaining capability. It does not prove that maximum aggregate speed is safer than targeted development, testing and regulation.
So, why do we entertain the narratives pushed primarily by tech CEOs when it creates exactly the difference which we should seek not to create?
There is clearly substantial discussion. My complaint should have been more precise: the existence of public discussion does not mean the public has meaningful influence over the decisions, nor that the costs and benefits are distributed democratically.
I do have a preferred conclusion, and disagreement with it is not evidence that discussion is illegitimate. The question is whether “China” is being used to foreclose ordinary evaluation of labor displacement, energy use, surveillance, market concentration and safety—as though the arms-race premise settles every downstream policy question.
I don’t believe this social and economical pain is worthwhile. The narrative is overblown to justify it. That doesn’t require a conspiracy.
This is where it all gets a little tinfoil hatty for me. The timeline is that huggingface detected a hack, shut down the hack, and reported it to the FBI. Days later, OpenAI discovered that their AI was the source of the attack. Are you implying that Huggingface and OpenAI were working together, and were willing to make a false FBI report for a PR stunt? That seems like a pretty extraordinary claim.
When it comes go AGI, the distinction between skeptics in the field and believers (because I don’t know a word that is an antonym of skeptic) in the field isn’t that this is race, it’s how long the race is going to be. Believers claim that AGI will be created within a decade, skeptics put it within 25 years. (These numbers are for example only; I haven’t kept up on the specific claims).
People that reject the possibility of AGI at all are, from my experience, using a definition of AGI that includes sentience or consciousness, but the accurate definition of AGI doesn’t mention these concepts at all.
Based on what?
These things do not need a treaty. I would wholeheartedly support any legislation to this effect. I recall that one company, Anthropic I think, lost government contracts because it categorically refused to allow its AI to be used for public surveillance or weapon systems.
Which means the government wonted to do those things. Scary thought.
No more than if would be to police a country’s efforts to develop nuclear weapon technology.
At the risk of being snarky, do you not understand what a race is?
I think I’ve assumed some shared knowledge that we may not have.
I focused on the current capability of AI in regard to hacking to demonstrate that it could be used offensively, but the danger that warrants a global pause in development isn’t strictly the ability to hack, though it is entwined with that concern.
The danger (many consider an existential danger) is creating an LLM that is better at developing LLMs than even the best humans. Once that happens, it becomes a positive feedback loop. If we allow it to help make a better version of itself (and I think we both agree that the incentives, both financial and strategic, would push them to do so) then we lose almost all control over AI.
Note that being better at improving LLMs is not AGI. That is to say that we can reach this point without arriving at anyone’s definition of AGI-- the artificial intelligence does not need to be general for this to happen.
Another assumption I made is that we both understood that no one on this planet really understands why LLMs work as they do. Chatting using natural language and receiving a coherent natural language answer is an emergent property of LLMs that we do not fully understand.
The “arms race” is based on the pretty solid assumption that AGI is possible using current transformer technology and that the first group to get to a place where their LLM is better than any human at improving LLMs will be the first to AGI. A side effect is that the skills that would make an LLM better at improving LLMs, also makes them better at finding and using exploits in software.
I’m not much concerned on whether people might consider the ability to improve itself as an “arm” and therefore comfortable calling this an arms race. The threat is arguably more dangerous to humanity than nuclear bombs could ever be.
That’s why we need a global treaty to halt AI development.
Are you familiar with AI 2027? As far as I know none of the people who worked on that are employed by AI companies. (Though I think some might have been in the past.)
There’s a really excellent third-party YouTube Video that I recommend, that walks through the narratives in the paper.