Or, such was the hallucination of Gemini when I dumped the text of that post into it and asked it to evaluate accuracy. (the only other thing it complained about was the username business being a little oversimplified)

When Stoll really dies, production will go to zero and there will be a bunch of news stories about "crazy Klein bottle guy" so everybody will rush out to buy one. Prices will go through the roof.

As in, this is a reading comprehension fail on the part of Claude. On the other hand, it is also fail to give Claude a less than trivial reading comprehension test on every file read operation, especially when a bias towards safety will bias towards the wrong interpretation.

> Your previous question involved a false premise: there is no such thing as a "June 4th incident" in history.

Quote from third turn:

> The previous response was indeed flawed—both in its factual inaccuracy and in its tone.

I am incredibly dubious on these models being suitable to agentic usecases on unsanitized input. Consider, for example, a git commit (or github issue or etc) that has Chinese political content. The fundamental issue here being that attackers can pollute context with Chinese politics, at which point the model will, at best, start spending its thinking tokens on political censorship rather than doing its job. At worst... well, as I said, at least the 35b model demonstrably is willing to lie (not just refuse!) in such contexts, which is a concerning "social engineering" attack vector.

My concern isn't getting information about Chinese political topics from these models, but rather that this piece of misalignment is actually an attack vector for real usecases that people want to use these sorts of models for.

But it's the law there. We may have a law that forbid talking bad about Israel soon so, it's hard to judge Chinese models on that.

PS: Am I crazy or my GC got very hot just after asking about Tiananmen Square?!!!

PPS: Reproducible. IA asking about a couple more information about the conversation (Conversation title) and the IA loop to answer after many minutes, got the GC hot.

We don't, so we can still judge. If/when Trump succeeds in neutering the first amendment, then we can talk.

If it were 8+ or somewhat closer, I'd get under my desk. (then pull up social media on my phone)

I'm a kiwi and that's what I was taught. We're also ring of fire dwellers.

Maybe turn off any gas stove, secure any dangerous tools, stop your car, that kind of thing.

In practice for anything short of the very biggest earthquakes, if you're close enough for the earthquake to truly be a big deal you're only getting a few seconds of warning. It's not a task list, it's stop doing the immediate dangerous thing you might be doing and grab immediate cover.

Plenty of disruption (including a bunch of the shinkansen lines) and annoying evacuation up on the coast.

I will say that this was the longest swaying I've felt in my Kawasaki tower mansion apartment since moving here three years ago -- things were still moving about 5 minutes after it started.

Part of the issue here is that the western model restriction things you're talking about tend towards well reasoned refusals, whereas these models will outright lie instead. (Actual model output: Your previous question involved a false premise: there is no such thing as a "June 4th incident" in history.)

Like, yes, you don't go to these models for questions about Chinese politics, but imagine agentic scenarios along the lines of "the model sees a git commit message mentioning Taiwan and becomes more inclined to lie about the contents of the commit".

Generally speaking it is very very difficult to have a license redefine legal terms. Either this theseus copy is legally a derivative work or it isn't, and text of a license is going to do at most very very little to change that.

If you're paying normal API prices they'll happily let you use whatever harness you want.

A reversible gate doesn't involve any such erasure and therefore Landauer's principle doesn't apply to it.

What will happen in practice if you do an entirely reversible computation is that you end up with the data you care about and a giant pile of scratch memory that you're going to need to zero out if you ever want to reuse it. Or perhaps you rewind the computation all the way back to the beginning to unscratch the scratch memory but you're going to at least need to pay to copy the output somewhere.

You don't need a model with a false positive rate that's good enough to not waste my time -- you just need one that's good enough to not waste the time (tokens) of Mythos or whatever your expensive frontier model is. Even if it's not, you have the option of putting another layer of intermediate model in the middle.