Are we sure this hasn't already resolved positive? If even 10% of these drones hit a target, that would be 300 deaths. They appear to be using YOLO8, a model trained by Ultralytics.

"By bringing together OpenAI’s advanced models with Anduril’s high-performance defense systems and Lattice software platform, the partnership aims to improve the nation’s defense systems that protect U.S. and allied military personnel"

is OpenAI- based frontier model (4o and o1, as of December 2024) w/ Anduril fine-tuning for use in Lattice command and control considered OpenAI here?

So Palantir make AI tools for assisting killers with lining up the shot right, it's the highest voted one.

But the self driving AI will actually be "taking the shot" when it drives into a pedestrian. 100 times will take a long time, car pileups are possible?

Airplanes have two pilots to rectify AI failure.

The campaign to stop killer robots won't stop other nations from making "AI that takes the shot" weapons though. Other looks pretty good, soonest.

Wrote a narrower question about publicly released language models: https://manifold.markets/toms/who-will-publicly-release-the-first

Who will publicly release the first language model to be credibly accused of being responsible for 100 deaths?

Credible accusations must be made or linked to in the comments to count for resolution. Once a credible accusation is posted, the question will resolve a month later, to give time for other accusations to surface. For a language model to count as publically released, either its weights and code must be published, or a general API must be available to the public (the API still counts if it's behind a paywall; "general API" is meant to exclude cases where acceptable prompts are limited to a small number of commands. Deaths caused through any causal mechanism traced through the public release count, as long as they aren't too butterfly-effect-y. They must be specific deaths traced to the model. Feel free to ask about specific scenarios and I'll answer whether they count. This a variant of a previous question, where this one is meant to exclude cases like self-driving cars or AI piloted drones.

Would military AI models designed to kill people count?

@Daniel_MC Yes

If a govt makes a model but doesn't publicly release it, does that count?

(Also I made the Meta option and only then noticed that Facebook was already on the list, so I think I've made that unnecessarily confusing, sorry)

@georgeyw I'll rule that "releasing" means that at least an API must be available to a nontrivial amount of the public, unless someone gives me a good reason to rule another way in the next 24 hours.

I also was interpreting Facebook as covering all of Meta, and so will not resolve to Meta, again unless someone argues convincingly otherwise in the next 24 hours.

What if the model can be used by many people in the public but there is no API (eg: Tesla auto pilot)

@Daniel_MC Good point, that should also count. Would anyone object to "release" mean that the model must be clearly in production (so not counting deaths during training or testing)?

If a fine-tuned version of Llama-2 is used in a violent act, this resolves to Meta, right? (not to whoever did the fine-tuning)

@LoganZoellner Oh good question. I'll go with whoever was responsible for most of the training counts as responsible (so yes to your question, unless the fine-tuning took more compute than training the base model did), unless someone gives me a better idea in the next 24 hours