addag

The whole history still goes through their servers, though.

And even if they are not doing it right now, they probably keep the history available for future training, "just in case".

The weird similarity between the engineers make it strangely look like it is AI generated (although it is not).

My physics teacher takes the spherical roast beef as an example for heat convection.

I'm wondering if it can be used as a low-discrepancy sequence

From the abstract "Further, our symbolic approximation allows us to modify an LLM's behavior in targeted ways via precise interventions on its internal representations [...]".

If this is true and easily computable, this might have big impact in AI safety, as it seems to be really lacking today.

While I agree with this statement, one could argue that what's important is not much the support but the emergent properties - in the same way that a wave is still a wave whether it is in the water or in the air.

Thus putting things similar to neurons in a network and making them able to learn could create behaviors similar to the brain. The fact that that biology used chemistry + electrical signal and computers use ReLU-like activation could be merely choosing the most efficient way to enable training.

In any cases, it is a strong evidence that the freezer can be remotely controlled globally, and this is in itself a source of concerns (what other electrical equipment can be controlled remotely, especially on such critical features?)

Until recently, I would have Occam's razor'd this and thought of a misconfiguration error from the military.

Now, considering how the OpenAI/HF hack emerged out of nowhere, it might be possible that groups of AI agents might have done this even without any close human supervision.

Isn't sliding-window a part of linear attention methods?

Tbh, I won't be surprised that 1 is true; it does not look that far fetched. Organizing scarcity is already a well-established strategy in the marketing playbook; there are examples of other big organization playing a way longer game to conquer a market (see for instance how Japan became a coffee nation https://www.thetakeout.com/how-nestle-brought-coffee-to-japa...)

This behavior seems to happen only when there is no more words in the list that can complete the current string. Otherwise, it shows only existing domains.

I guess this behavior helps to autocomplete the domain extension for less-known domains that are not in the search list.

For the nonexisting domains, it seems like it only autocomplete with the possible domain extensions, (e.g .com, .org, ....) as the search list is non-exhaustive. But it could indeed be improved by not sending autocomplete requests anymore.

Peak Home Alone behavior right here.

Cutting hair.

I think barber/hairdresser will be one of the last jobs to fall.

I'm wondering if there is a way to watermark those data to retrieve if they were used as training for generative models (obviously this is not a trivial task, and would need statistical instead of deterministic proofs).

That specific example might be relevant to the case "I don't see any acquisition being good for the users", but not for the specific case of the HuggingFace buyout. Workflow was already operating within the bounds of the closed Apple ecosystem. Huggingface is providing models not only for Nvidia chips but for their competitors hardware (Apple Silicon, AMD GPUs, ...).

Interesting ideas, maybe the real open alternative will indeed be torrents.