valleyer

Couldn't they use frontier open-weight models from Chinese labs? The current Chinese government is friendly to them.

Seems like it would be shut down by an injunction almost immediately.

You can't expect me to stop selling crack to kids on the streets. My job is to feed my family.

Not in vi. :)

At least in the US, police don't enforce terms of service of private corporations. That's a civil matter, enforced by filing a lawsuit.

Keep in mind there may be a significant Pepsi Challenge effect at play here.

Yes, I too remember when Shaq starred in this commercial.

https://www.youtube.com/watch?v=Af0gtsjfy7E

Less space than a Nomad? Lame?

Funny. But... paper covers rock.

I'm not an expert, but... might this indicate the need for "sub-models", meant to be invoked by the general model to write good prose for it? Those sub-models would not be RLed on code (or any algorithmic-feedback tasks that might reinforce bad writing), and could specifically be tuned for prose, even at the cost of general intelligence (which would be provided by the worse-at-writing more general model).

Prior art: the evil bit

https://datatracker.ietf.org/doc/html/rfc3514

During training, certain tokens are more likely to lead to a lower loss function value, which is how you "win" the game of LLM output.

It could be added to context only on the inference server, with the model heavily instructed not to repeat its contents. Not foolproof, obviously.

"Domino effect" would probably be the more relevant named phenomenon there.

Interesting. I don't remember my crossover cables being marked that way; you had to read the dot-matrix printing along the side to know. Color would have been way more handy.

One of the strawiest strawmen I've seen in quite some time.

At least historically, it actually did run its own entire userspace, including daemons (all the way to launchd, IIRC), etc. All compiled for Intel, back in the day. Shared the kernel though.

Many of the nominally shared system pieces between Mac and iPhone (like Foundation) actually had many subtle compile-time differences.

It is possible for both Trump and Biden to be wrong about something.

If punching random people on the street is so obviously bad, then why do people fight back when I do it?

Anthropic would estimate (and provide justification of) how much lost income (and perhaps brand equity, opportunity cost, etc.) it suffered as a result of the government's actions, and the judge would consider it and decide.

I agree. I just thought it was funny that someone was pointing out the legal unaccountability of an AI in this context.

Good point. CEOs of companies are often held personally accountable for corporate malfeasance by "the current legal framework".

Even if it's communication/announcements from the government?

I too am constantly frustrated that official communications from governments happen on Twitter or Facebook.

Certainly. But in the current social environment, being seen as (mildly) careless is less bad than being seen as someone who lets an LLM write prose for them.

"Reality has a well-known liberal bias."

The U.S. constitution does not say that. Where are you getting that from?

It does say that the president cannot pardon an impeachment, but that's not the same thing at all.

Hey, good for you for following through.

That's not generally how the English language works. We say "an SoC" but "a system-on-chip".

It could absolutely have been in an iMac's wireless keyboard; witness the "Sidecar" macOS feature that allows use of an iPad as an external display for a Mac.

Oy, "the solve".