Tell that to the humans working on matrix multiplication who spent years of their lives getting it from n^2.3728596 to n^2.371866, only for openai to blow it away at n^2.25
Have you learned something from every Fields medalist's research? If so you are a member of the extreme mathematical elite and you should probably just dig into the results yourself.
But the "mistranslation" is of the procedure that arrives at the final statement. The final statement, the thing that the lean code proves, itself has been well vetted by humans. So the lean proof correctly proves the NS blowup, it's just that the natural language paper has some mistakes and doesn't exactly follow the route the lean proof takes.
No, this isn't what this is about. From the abstract:
> In this process, an AI system translates the text from a natural language (NL) into a formal language such as Lean. Once this translation is done, the argument expressed in the formal language can easily be mechanically verified. The purpose of this article is to demonstrate why this process may offer no confidence in the original NL argument, owing to the various difficulties in performing the translation semantically faithfully.
At my recent physical I had to sign a form acknowledging that if I mention specific concerns they are allowed to bill my insurance twice. The example given said the physical is only for general health and if I mention a concern like 'ear pain' they will bill it as a second assessment. So double the money for the same amount of time because the AI voice assistant said so.
If the government wanted to put iPhones in missiles, and apple said "we won't sell you iPhones to put in missiles" would the correct government response be to say "no government contractor can ever use an iPhone".
The DoD and US military does not allow contractors to use iPhones or anything else that is not backed by government-hosted or compliant infrastructure, such as GovCloud. The fact that the Pentagon and Anthropic could not come to an agreement on the usage and deployment to GenAI.mil means that it can't be used.
The reasons why are incidental to that requirement for defense contractors.
Go into a random defense contractor office and (assuming you can get in to do this) ask them how many of their iPhones are managed by the company. The percentage who say theirs are will be close to 0%. Most are personal devices, and they use them to call into meetings with DOD officials all the time with no special precautions taken.
Exercising its power as a purchaser is "not buying it" or "choosing a competitor." Instead, this is arbitrarily and capriciously using an administrative mechanism in a novel way to punish a company.
Probably not, no private company has the ability to tell all its suppliers "thou shalt not do business with ". They lack the pull to compel them to do so. The closest would be things like what MS did back in the 90s, and got them in hot water with the anti-trust case against them.
That's not what happened here and you know this. It was, literal quote, "“[e]ffective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.”
Fair, and still not something most companies can pull off. Again, without risking running into anti-trust laws like MS did.
Do you seriously think that you can go to all your vendors (if you were a business owner) and tell them, "You can't use Atlassian when working on anything you sell us." How will your business actually compel all those vendors to drop their use of Atlassian for that work? How will you verify compliance and why should any of them sign a contract with you with these restriction? And if they do, how will you avoid anti-trust violations because almost certainly the only way you could talk them into signing such a contract would be to be a massive company throwing your weight around (like MS did with hardware OEMs back in the 90s, among other things).
They are not "just a purchaser" though. It's the world's most powerful organization that is funded by the people and is supposed to be for the people. I'm all for businesses making FAFO decisions, but expect a lot more here.
The fact that it’s a publicly funded entity is more reason for it to adopt hardball tactics, not less. It’s similar to FDR’s comments on public sector unions. When the government is the buyer, it’s even more important to get the best deal and terms possible, because it’s the public’s money’s you’re spending and the public’s interest you’re serving.
Exactly, setting policy for Americans and American companies is extremely important and should not be ego driven. These deals are to improve America, not to look strong.
The U.S. military’s policy has always been to secure the absolute best technology for killing people. So the DOD quite reasonably believes that it’s in America’s interest to have the best technology for building AI killing machines with no restrictions from non-military people on how it’s used.
Somehow people who think open models will lead to infinite abundance and people who think open models will kill AI development are both liking this post.
Your statement is a logical puzzle about responsibility.
It implies that a law prevents responsibility. The opposite is true. Preventing harm is a concern that might form laws. But for many forms of harm, existing laws suffice and should lead to responsibility.
Shifting responsibility can be done in many ways, such as calling for a regulator. Like the law isn’t a regulator in the first place.