colonel_mortis

But this behaviour is intentional, so reporting it to them without an explanation of why it should behave differently is pointless.

I think this is working as designed - if a blog post isn't published, it doesn't really exist yet, and shouldn't be discoverable. (Even if we did want to change it, it wouldn't be easy for us to do because of how it's architected, but that's because this is the intended behaviour)

We used to embed the plain parts list when you paste a link, but then I think the forum triggered their cloudflare rules, so now the forum server can no longer reach the PCPP server, which means we are quite limited in what we can do. The layout of the list in your screenshot looks fine to me though - it's not as pretty as the one on their own site, but it doesn't seem bad.

If you're seeing it on posts, that sounds like a different issue, and a screenshot would be helpful (though if relaunching helps, that sounds like a browser issue to me).

Is this editing a post, or editing a status update that you opened from the sidebar widget on the homepage? If it's a status update, it's surprisingly hard to fix on that page, but you should be able to edit it no problem if you go to your profile (or click the date to open the status in it's own page).

What would you imagine people discussing in the WAN Show subforum, and how much of a problem is that content in its current locations?

We have contacted Invision Community, but they have indicated that they're no longer doing non-critical/security fixes for the version that we're using, and it will be a while before we're able to move to the next major version.

No, it looks like the forum software devs don't plan to fix this. We may look at fixing it in our own version at some point, but that depends whether we want to keep supporting this software ourselves. I have removed your signature for you though.

The architecture of the forum software's theme system makes this fairly difficult unfortunately.

We should make it so anyone who complains about it gets Comic Sans permanently, even after the rest of the forum goes back to Inter.

Do you have any strict privacy extensions or settings, or anything like that? That error means that the token passed in the request doesn't match the token linked to your cookie, which if you're seeing it reliably is most likely because the cookies are getting blocked or deleted or something like that.

This is generally all true (although I don't think the petabyte claim is true, or at least useful in the context of observable effects), but agentic frameworks built around the models have increasingly had success with having models track some state in files and essentially running RAG (or agentic equivalents, eg tools for reading files) at inference time. That is obviously meaningfully different to how we think about memory, and has significant downsides, but it evidently works well enough to produce systems that some people find convincingly authentic. There's a lot more progress to be made before it meets the bar that you or I would hold here, but I think it's a mistake to downplay what progress has already been made. If you showed the models we have today to someone in 2020, I think most experts would say that we had achieved AGI. The consensus goalposts have moved since then, but at this point diferent people have different, often vaguely defined, thresholds for AGI. Your definitions of AGI and sentience seem to exclude LLMs by construction. My claim (which doesn't contradict yours) is that even if those definitions have deep meaning, there are other definitions that one could meaningfully use for those concepts that may turn out to be achievable by current methods, and those definitions could still be useful. For sentience in particular, at a very hand wavey level, is a model that produces reasoning tokens and talks about having inner desires that much different from a human that has an internal monologue and talks about desires? Quite possibly, but I don't think it's a given, and I think there's something interesting about the class of things that includes both of those, even if we don't call it sentience.

Yes, but this is much less practically true than it used to be - automatic context compaction has been getting better, and in this article they were talking about their agent having a persistent memory, probably (though I didn't read into this in detail) similar to openclaw where the model is recording memories and state to be read after a compaction. I've not used anything like that, but my understanding is that while they aren't great yet, they're still good enough to be convincing for some people, which is a big step forwards.

I think there is a genuine question that nobody has managed to answer yet about what sentience actually means. If you have something that can reason with language, an internal train of thought, and persistent memories, is that actually meaningfully different to human consciousness? It doesn't feel the same, and it certainly doesn't feel right, but when we dig into it, I don't actually have a justification for why computers wouldn't be able to emulate the thing our brains do that we call consciousness. It definitely doesn't feel right to call the models we have today conscious, but I can't construct a coherent explanation why they can't be that doesn't just devolve into just trying to point at things that are different about LLMs. To be clear, I am not claiming that LLMs in general, nor this specific setup, are sentient. I am saying that I don't think we know what sentience even means, but I think there is an emergent behaviour of the models that has started feeling close enough that it would be helpful to actually understand what sentience means. I hate to break it to you, but that was in 2022, a few months before chatgpt first launched.