IanCal

49 ms·

841.
by IanCal 2y ago
Are you going to quibble on the definition of purposefully? Is your issue that you think "faking" is a term only usable for intelligent systems and the llms don't meet that level for you? I'm not sure I understand your i
842.
by IanCal 2y ago
I encounter "spicy auto complete" style comments far more often than techbro AI-everything comments and its frankly getting boring. I've been doing AI things for about 20+ years and llms are wild. We've gone from special
843.
by IanCal 2y ago
It doesn't matter what the reason is though, that was rather my point. They act differently in a test setup, such that they appear more aligned than they are.
844.
by IanCal 2y ago
What it means is that the complaints about language models being "just maths" or "just code" are not sufficient.
845.
by IanCal 2y ago
The problem I have with this is it seems to rapidly assume that humans and animals are somehow magic. If you think we're physical things obeying physical laws, then unless you think there's something truly uncomputable about us th
846.
by IanCal 2y ago
> But that alone is not very scary. So what could justify a term like "alignment faking"? Because the model fakes alignment. It responds during training by giving an answer rather than refusing (showing alignment) but it does
847.
by IanCal 2y ago
Ah yeah sorry I should have mentioned it wasn't the same, I used it as it has a nice number that shows the bias to a pretty extreme degree.
848.
by IanCal 2y ago
3.10 so I redid it on 3.13.1, same results.
849.
by IanCal 2y ago
Yes but if you want to do things in a less obvious way you should be aware of the downsides, such as bias in your random numbers. Also making sure you watch out for off by one errors. Stolen the number to show this off well from a bug repor
850.
by IanCal 2y ago
Yeah I'm not sure what I'd use this for . I tend to know if I've had a bad night's sleep, and the interventions I'd try are all things I should do anyway - don't drink, less caffeine, lower stress.
851.
by IanCal 2y ago
That is the use of the term technically in the UK yes. You don't have a home. You're staying somewhere or sleeping somewhere but you don't have a home. If you're sleeping in a homeless shelter, isn't that a place yo
852.
by IanCal 2y ago
People still use rape but there's a thing there around understanding exactly what that might mean. I've been "sexually assaulted" but not raped, legally, because of UK laws. Unhoused may be because homeless covers people
853.
by IanCal 2y ago
To be fair on this one, most of the times I complain people haven't got to the end of some short story before asking. This is an enormous and dense resource. Great, but I started scrolling and got surprised.
854.
by IanCal 2y ago
Does this come with the page/code to store, or is there a good way of doing this? I looked a while back for properly archiving pages and their code but things were all in the works, maybe that's more solved now. This kind of thing
855.
by IanCal 2y ago
Ahh that makes sense.
856.
by IanCal 2y ago
> A forum that isn't proactively monitored (approval before publishing) is in the "Multi-Risk service" category (see page 77 of that link), That's not true, you'd need to conclude you're at a medium or high
857.
by IanCal 2y ago
> The subsection headed ‘User numbers’ (which begins at paragraph 5.7) explains when a service is to be treated as having more than a particular number of monthly active United Kingdom users for those measures which apply in relation to
858.
by IanCal 2y ago
edit - moved across now the comment is alive
859.
by IanCal 2y ago
Their examples of it being right to say you're low risk have a key part about it not happening before. Medium risk example had "has had warnings of csam being shared before from international organisations and has no way of spotti
860.
by IanCal 2y ago
https://www.ofcom.org.uk/online-safety/illegal-and-harmful-c... You need to do a risk assessment and keep a copy. Depending on how risky things are, you need to put more mitigations in place. If you have a neighbourhoo
861.
by IanCal 2y ago
> Looks like it now requires an online community to have its own bureaucracy in place What do you mean by bureaucracy in this case? Doing the risk assessment?
862.
by IanCal 2y ago
Can you explain what you mean by "this attack"? They'll submit a complaint to the regulator that you've not done a risk assessment? I've tried submitting issues to the ICO before but didn't have enough for them
863.
by IanCal 2y ago
> the real risk I see is that as it's written, and as Ofcom are communicating, there is now a digital version of a SWATing for disgruntled individuals. I'm sorry, what precisely do you mean by this? The rules don't punish
864.
by IanCal 2y ago
> You just have to make individual value judgements every day on thousands of pieces of content That's simply not true.
865.
by IanCal 2y ago
> It's insane that they never carved out any provisions for "non big-tech". That would be insane, and it's not true. You have to consider the risks and impacts of your service, and scale is a key part of that. I think
866.
by IanCal 2y ago
The rules aren't "never have bad things on your site" though. The example of "medium risk" for CSAM urls is a site with 8M users that has actively had CSAM shared on it before multiple times, been told this by multi
867.
by IanCal 2y ago
I doubt it. While it's always a bit of a gray area, the example for "medium risk" is a site with 8M monthly users who share images, doesn't have proactive scanning and has been warned by multiple major organisations that
868.
by IanCal 2y ago
And excellently usable. I took it out of the box, tightened a few things where required and I think that was about it. Excellent bit of kit.
869.
by IanCal 2y ago
That's a different thing. This system doesn't do that, but that's one use case they have for it.
870.
by IanCal 2y ago
I don't think that really helps if the company doesn't have any money.

More ›