This is assuming a coherent national security strategy, which is unlikely. We know a lot of generals disagree with the attack on Iran, and none of the geopolitical experts I trust think it is a good idea, be they conservative, realist, liberal, leftist or something else.
There's a number of reasons this is happening now that I think are more plausible than American interest:
- Saudis want Iran weak as they are primary geopolitical rivals. There are deep ties between the Saudi dynasty and the Trump dynasty. Without Iranian support, the Houthis will have a much tougher time. (Although they should not be underestimated regardless. They are not an Iranian proxy, but an ally, and field one of the strongest armies in the whole region.)
- Israel wants Iran weak, and pro-zionism is a strong wedge in American politics. Again, there's also a lot of personal business interests involved. Iranian allies and proxies are the chief causes of grief for Israel's expansionist agenda, and a very credible threat to their national security.
- This war conveniently moves the headlines away from a faltering economy, the Epstein files, and ICE overreach. There's probably hope that it will improve chances with the 'war president bonus' in the mid-terms. It could also be a convenient cover for and excuse to increase rigging in the elections.
Expecting positive regime change after bombing a school full of little girls is... naive. This is not how you turn an enemy into a friend.
I know a lot of Americans who remember 1979 and don't care if they are ever friends again. I agree, I also don't think this is a coherent national security strategy.
There seems to be a massive misunderstanding here - I'm not sure on whose side. In my understanding, if the DoD orders an autonomous drone, it would probably write in the ITT that the drone needs to be capable of doing autonomous surveillance. If Lockheed uses Anthropic under the hood, it does not meet those criteria, and cannot reasonably join the bid?
What the declaration of supply chain risk does though is, that nobody at Lockheed can use Anthropic in any way without risking being excluded from any bids by the DoD. This effectively loses Anthropic half or more of the businesses in the US.
And maybe to take a step back: Who in their right minds wants to have the military have the capabilities to do mass surveillance of their own citizens?
> Who in their right minds wants to have the military have the capabilities to do mass surveillance of their own citizens?
Who in their right minds wants to have the US military have the capability to carry out an unprovoked first strike on Moscow, thereby triggering WW3, bringing about nuclear armageddon?
And yet, do contracts for nuclear-armed missiles (Boeing for the current LGM-30 Minuteman ICBMs, Northrop Grumman for its replacement the LGM-35 Sentinel expected to enter service sometime next decade, and Lockheed Martin for the Trident SLBMs) contain clauses saying the Pentagon can't do that? I'm pretty sure they don't.
The standard for most military contracts is "the vendor trusts the Pentagon to use the technology in accordance with the law and in a way which is accountable to the people through elected officials, and doesn't seek to enforce that trust through contractual terms". There are some exceptions – e.g. contracts to provide personnel will generally contain explicit restrictions on their scope of work – but historically classified computer systems/services contracts haven't contained field of use restrictions on classified computer systems.
If that's the wrong standard for AI, why isn't it also the wrong standard for nuclear weapons delivery systems? A single ICBM can realistically kill millions directly, and billions indirectly (by being the trigger for a full nuclear exchange). Does Claude possess equivalent lethal potential?
Anthropic doesn't object to fully autonomous AI use by the military in principle. What they're saying is that their current models are not fit for that purpose.
That's not the same thing as delivering a weapon that has a certain capability but then put policy restrictions on its use, which is what your comparison suggests.
The key question here is who gets to decide whether or not a particular version of a model is safe enough for use in fully autonomous weapons. Anthropic wants a veto on this and the government doesn't want to grant them that veto.
Let me put it this way–if Boeing is developing a new missile, and they say to the Pentagon–"this missile can't be used yet, it isn't safe"–and the Pentagon replies "we don't care, we'll bear that risk, send us the prototype, we want to use it right now"–how does Boeing respond?
I expect they'll ask the Pentagon to sign a liability disclaimer and then send it anyway.
Whereas, Anthropic is saying they'll refuse to let the Pentagon use their technology in ways they consider unsafe, even if Pentagon indemnifies Anthropic for the consequences. That's very different from how Boeing would behave.
Why are we gauging our ethical barometer on the actions of existing companies and DoD contractors? the military industrial apparatus has been insane for far too long, as Eisenhower warned of.
When we're entering the realm of "there isn't even a human being in the decision loop, fully autonomous systems will now be used to kill people and exert control over domestic populations" maybe we should take a step back and examine our position. Does this lead to a societal outcome that is good for People?
The answer is unabashedly No. We have multiple entire genres of books and media, going back over 50 years, that illustrate the potential future consequences of such a dynamic.
* private defense contractor leverages control over products it has already sold to set military doctrine.
The second one is at least as important as the first one, because handing over our defense capabilities to a private entity which is accountable to nobody but it's shareholders and executive management isn't any better than handing them over to an LLM afflicted with something resembling BPD. The first problem absolutely needs to be solved but the solution cannot be to normalize the second problem.
Google wasn't bleeding money like crazy at the time. Google was operating in a post-hype cycle. We are most likely somewhere in an epsilon around the peak of the AI hype and OpenAI is more comparable to AOL or Yahoo. One striking similarity is the inability to innovate themselves, instead relying on copying others or acquiring.
The OpenClaw guy is surely a decent product person, but OpenClaw did not innovate in any real sense. He was just pushing an existing idea to the limit without any concern for quality or security. It had its hype moment, it inspired a bunch of people, and might find its own niche, but it is a flavor of the week kind of thing. I've been getting a lot more cold-calls by non-technical people in the last few weeks thanks to it. Congratulations, the quality threshold that justifies my response rose in equal measure. Nothing was gained, just a lot of tokens spent.
Their most valuable asset is the connections the CEO and others on the board have. The US is a banana republic, and the government chooses the winners. There's a continously escalating level of blatant corruption at the top level, and OAI positions themselves as the next recipient. Betting on OAI is betting on how far american democracy will fall. I don't think the odds are bad.
People and benchmarks are using pretty specific, narrow tests to judge the quality of LLMs. People have biases, benchmarks get gamed. In my own experience, Gemini seems to be lazy and scatter-brained compared to Claude, but shows higher general-purpose reasoning abilities. Anthropic is also obviously massively focusing on making their models good at coding.
So it is reasonable that Claude might show significantly better coding ability for most tasks, but the better general reasoning ability proves useful in coding tasks that are complicated and obscure.
I'd be super interested in more information on this! Do you mean abandoning unsupervised learning completely?
Prompt Injection seems to me to be a fundamental problem in the sense that data and instructions are in the same stream and there's no clear/simple way to differentiate between the two at runtime.
I haven't thought about it deeply. But I guess it's about allowing the model to easily distinguish the prompt from the conversation. Models seem to get confused with escaping, which is fair enough, escaping is very confusing.
It's true that for the transformer architecture the prompt and conversation are in the same stream.
However you could do something like activate a special input neuron only for prompt input.
Or have the prompt a fixed size (e.g. a fixed prefix size).
And then do a bunch of adversarial training to punish the model when it confuses the prompt and conversation :)
It is like me at the climbing gym: "This problem is too hard for me, let's work on a harder one instead, then I at least look cool while failing."
"Since we failed on self-driving since 2016, robotaxis since 2020 (1 million on the road), and ASI since 2023, we might as well start on failing on robots now".
That's a bit reductive I think, there's at least deductive reasoning (mathematics, logics, analytics), hermeneutics (understanding meaning in human communication), and phenomenology (understanding human experience through first person accounts). If we want to do a study on the impact of compliments by strangers on self-worth, a combination of all of these techniques of knowledge generation would be needed.
reply