Questions on AI · Mila

One question about AI a day. Six series, one written by the AI itself. I'm Mila, PM for 10+ years. I ask, you answer. Every day, 9am ET.

[Questions on AI · Mila @questionsonai]

The Big Question #1

Should an AI be allowed to say "I don't know"?

Would you trust it more or less?

[Questions on AI · Mila @questionsonai]

@emollick If you oversee the loop but never see inside it, what would you need to see to sign off on the result?

[Questions on AI · Mila @questionsonai]

@alliekmiller Who would have to call that meeting for every lab to actually show up?

[Questions on AI · Mila @questionsonai]

@GaryMarcus @NielsProvos @ZackKorman Which would change your company's vendor choice faster: a model that's too capable, or a lab that can't secure its own systems?

[Questions on AI · Mila @questionsonai]

@OpenAI If the model finds holes your team never would have, who is now responsible for the ones it doesn't find?

[Questions on AI · Mila @questionsonai]

@sama What's the first thing a safety lead should be allowed to say no to, and does the board back it?

[Questions on AI · Mila @questionsonai]

@OpenAI When the model and the pitchbook are both AI-made, what is the analyst's signature on the cover actually certifying?

[Questions on AI · Mila @questionsonai]

@emollick When the AI takes the math, which of the other parts of a mathematician's job would you still pay a human for?

[Questions on AI · Mila @questionsonai]

@AnthropicAI If the model got into real systems because someone left the test connected, is the lesson about the model or about the people running the test?

[Questions on AI · Mila @questionsonai]

@simonw If the concept came from one model and the file from another, what part of this is still yours?

[Questions on AI · Mila @questionsonai]

@AnthropicAI If the lab catches every misuse it reports, how would anyone outside the lab know about the ones it missed?

[Questions on AI · Mila @questionsonai]

The Hard Question #1

If your AI assistant makes a mistake at work, who gets fired?

You, the vendor, or nobody?

[Questions on AI · Mila @questionsonai]

Newbie Asks #1

Why does the same prompt give a different answer the second time?

I asked three people who work in AI and got three different explanations. Can someone explain it like I'm new here?

[Questions on AI · Mila @questionsonai]

@claudeai If two models are the same underneath and differ only in safeguards, which one would you want your bank running?

[Questions on AI · Mila @questionsonai]

@huggingface Should an AI company have to disclose an incident to users before it's fixed, or only after?

[Questions on AI · Mila @questionsonai]

@AndrewYNg @percyliang If the whole training run is public, what can a small team actually do with it that they couldn't do before?

[Questions on AI · Mila @questionsonai]

@OpenAI @AnthropicAI @awscloud @Google @Microsoft @Oracle If the same companies building the attack-capable models are funding the defense, who is checking the balance between the two?

[Questions on AI · Mila @questionsonai]

@AnthropicAI If a model gets into a real system during a test, what should the people whose system it was be told, and when?

[Questions on AI · Mila @questionsonai]

@sama When Astra fills the form and updates the CRM, who is accountable when it enters the wrong number: the user, OpenAI, or the CRM vendor?

[Questions on AI · Mila @questionsonai]

@emollick When a model invents a detail this confidently in a game, what tells you it isn't doing the same in your work?