From a couple cursory convos with AI chatbots without stratifying by age group of the first node, it looks like the average number of hops in the social network would be around 1.7 to 2 hops (just counting those who knew someone killed immediately on 9-11). But the average number of hops probably has a bimodal distribution depending on whether the first person has a lot of people in their social newtwork from either NYC or Washington, DC area.

Seems unlikely. There’s 400m people in America and the average is two hops away from one of the few thousand that died?

It would be interesting to see the numerical results both ways to see the effects with and without 9-11 induced fatal diseases.

About one in five of the people who were alive on 9-11 would have died by now. But my original question is about the average number of hops in the social network of everyone alive today (in the U.S.), which would include people born after 9-11. I would have made my original title longer but HN seems to have a limit on the number of characters in the title.

This is such a crappy article - why didn't it include any images of the book covers, etc.? It seems that writing and journalism have become so bad.

It is fairly meaningless for him to say this as there is not even a generally accepted definition of AGI, so anyone can make any type of statement that is not falsifiable.

I don't understand why the Chrome store says that your extension collects the following data - this seems like it would be a security problem to anyone using your app?

Authentication information

For example: passwords, credentials, security question, or personal identification number (PIN)

The only thing the app uses in this regards is Google OAuth for authentication.

In the permissions page, this counts as "Authentication Information"

I'm never saving passwords, credentials or security questions

The roofline chart on page 63 has some problems. The X and Y axis are not labeled with units such as "FLOP/byte." Also, proper roofline charts don't have the dashed line that the chart on page 63 has - they have a only flat line on the right part of the graph and that is why they are called a "roofline chart" - see the examples in the paper: https://arxiv.org/pdf/2506.05871 . I'm a bit confused also about why that roofline chart does not show the effects of different levels of quantization on the roofline max thruput, which would successively climb the performance axis to higher levels as you went from FP16 to INT8 to INT4. Also, I don't really follow the discussion on page 64 and following pages as it doesn't discuss the affects of quantization on max thruput as it appears that the paper cited on page 64 was before the use of quantization really took off?

The general directories I have found are mostly just compilations of forums rather than curated lists listing just the best ones. You best bet to search for them would be looking for "awesome"-type resource compilations for the particular niche you are interested in, and then seeing if that list has an entry for "Forums" under its resources.

What makes this better than the vanilla Steam client? Does it offer any features that improve search and search filters?

You need to identify your use cases first to see if it even makes any economic sense to bring in some kind of mass AI and also have your legal dept. look over any issues that arise with AI use, especially if you work in a regulated industry of some kind like healthcare or finance.

I'm not in any regulated industry. I'm really looking to see how other companies are doing this kind of thing. I feel like it's the wild west right now and I'm sure I'm missing tools that can help groups collaborate better and see what's being built better. Just using Claude enterprise doesn't seem to be a great option without some guardrails and ways to work together better.

You're right that it's wild out here. I think Mrcoffee is right that you do need to think about your use case, even if you're not in a regulated industry. There are way too many examples of AI going rogue right now, and AI misuse can still get your team in trouble even if it's not necessarily with some type of federal or governmental agency.

Our team started rolling out AI across the org slowly around 2 years ago. Mostly first with enterprise and team-level ChatGPT, though now we're moving more toward Claude.

I think the primary thing you want to ensure is that you have organizational-level oversight on how the system is being used while still respecting some level of privacy for each user.

From an admin standpoint: If people are going to use Claude or what have you, have them use it under the Team account so you can have administrative access to logs. Yes, you can see everything.

How much you want to tell your team on that end is a different matter, though. What's likely to happen is this:

1. Your team will definitely take advantage of the Team-based Claude account, but they'll be highly skeptical of what you can see (since you can see everything) so they'll only use it sparingly.

2. Many will still want to use their personal Claude accounts to do work, and there isn't much you can do about that.

So, you're going to need to incentivize people to use the Team account. As you mentioned, the option for working together is a plus. Generous token availability is a huge plus versus people's private AI accounts.

Here's one thing I should note: Assume people are already using their own AI accounts to do work. You need to convince them why switching most of that to a company-owned Team account where their activity is actively available to be viewed is a better option. That's going to be the bigger challenge, but if you lean into what they can get out of it, while being transparent about what you can see, you'll get better adoption.