Introduction

On the topic of safety, I suppose one key ethical consideration is when we're in these spaces like Mindstone and other techno-optimist spaces, should we feel guilty about the fact that we're not addressing all the world's ethical concerns all at the same time? And I'd like to make the case, no, we shouldn't.

because one thing that converts a fear and anxiety into some level of reassurance is knowing who the goodies in the space are. We see enormous media coverage of who the baddies in AI might be or who the media might like to portray them to be. but we don't get nearly enough coverage of the goodies.

The Goodies in AI Spaces

So this is an amazing resource put together by the AISafety.com of a landscape of lots of different AI organizations together have, you know, it's the tune of hundreds of millions of dollars worth of funding

And I just thought I'd talk through some of them I'm familiar with, some of them are based here in Cambridge, and then we'll do a prompting exercise to show how you might put together a literature review of some of the best outputs, one or another. And since I've been advertising as an ethicist, I sort of bring up ethical considerations as to where they arrive.

AI Organizations and Their Impact

So if we had Mind Stone on here, it would be in Training Town next to Blue Dot Impact.

Blue Dot Impact is a spin out from Effective Altruism. But Effective Altruism is a community of people committed to doing good better.

Raise your hand if you've heard of Effective Altruism. Yeah, a couple, not a huge amount.

They get very, very interested in preventing the world from ending from artificial general intelligence systems. And Blue Dot Impact is their spin out that does governance courses for people for free.

It's 12-week programs.

Lots of these organizations do great work.

Fellowships and Influential People

The ERA fellowship was founded by someone in my year at Trinity in college called Nandini Sarkar. She raised half a million pounds in her first year at the end of summer to create this fellowship for AI safety projects based in Cambridge as an equivalent to CERI. At the time it was CERI, the Cambridge Existential Risk Initiative in equivalent to CERI, the Stanford Existential Risk Initiative, and since become ERA, the Existential Risk Alliance.

You've got the Center for Long-Term Research, similarly does Summer Fellowships.

If we go across, well, sorry, I suppose I'll show you the whole thing as well.

Advocacy and Research Institutions

Podcast Port, Advocacy Average, where you've got some activist organizations there as well.

EA Forum.

Obviously, we can't talk about all these things all at once.

There's AISafety.com.

Research Institutions

In Strategy Summit, we have more of the research-based organizations at university. So here in Cambridge, we have CFI, where I did my master's in the ethics of AI, the Lev Hume Center, Future of Intelligence, has some of the world's premier AI ethicists. And we also have the Center for the Study of Existential Risk. They're in the same building.

The biggest name they have at the minute is Yuval Noah Harari, who wrote Sapiens and Deus, most recently Nexus. But lots of other amazing ethicists at CFI.

I have to shout out my course director, Dr. Claire Ben, who is doing some amazing work on how we navigate truth in a world full of noise and AI generated content. And what level of skepticism is appropriate, just blanket. As well as lots of other really interesting stuff.

What else do we have? CSET, Center for Security and Emergency Technology, very influential think tank in the US.

The UK AI Security Institute, formerly the UK AI Safety Institute, one of the obviously trends we've had the last two years or so is that AI safety research has deflated in the big labs. We see OpenAI unfortunately disbanding their alignment team. I think Google is disbanding their super alignment team.

But we've seen the blossoming of these safety institutes in the UK, in the US, in Singapore. Last I checked, there were 11 of them worldwide.

And what's interesting is they have access to the premier model, so they can do things like Donato just showed you, try to jailbreak their models and test them back. So there's this offloading and outsourcing of safety testing, which we can debate whether it's good or bad, but is a trend.

Utilizing Data for Better AI Output

OK, so we've got this enormous resource.

And what this resource constitutes ultimately is good data. You see each of them have a little description for what's relevant to them.

So we have a question. How can we use this good data to create a better AI output?

So if I just, yes, if I were to just blanket, I don't know, that's the finished one. But if I were to just blanket, say, to perplexity, and raise your hand if you use perplexity. Great, lots and lots of people, lots and lots of people.

So I wanna benchmark, compare these different results. What are the most influential, AI safety papers coming out of Cambridge recently. And I want to filter for academic papers.

Challenges in Benchmarking AI Research

And we're just using search, not deep research in this feature, because I want to do a like for like comparison. So first problem was we've got a trend, not a paper. Similar broad, we've got themes, we've got summary, it's a summary of themes, conclusion, underwhelming.

Now we could use deep research to try to get a better sample. And what's gonna happen there is it will do its first round of searches, then review them for how good it is, then have another go, then have another go, then have another go.

a little bit compute-intensive. It's nowhere near as compute-intensive as generating an image with like a stable diffusion model, and nowhere, nowhere near as intensive as generating video, but a little bit compute-intensive.

Leveraging Notebook LM

So if we were persnickety environmental conscious, or if we just wanted to leverage this amazing asset from the mathaisafety.com, then we want to somehow convert this asset into a prompt for perplexity to go and then use that amazing filter feature, okay? How are we gonna do that? We are gonna use another amazing tool, Notebook LM.

Who's heard of Notebook LM? Amazing, amazing.

I've actually already done this. So since it's up, I'll just say.

Oh no, sorry, this is a different example. Ah, yeah, no, I have.

I've uploaded the website to this link and one of the awesome features of Notebook LM is it doesn't just read from a document you've given it, it will take information from a web page. And not just any information from a web page, literally the descriptions of the web page that appear when you hover the mouse over it, which shows some impressive computer use features.

So if I ask, could you produce a list of the Cambridge organisations on this map and in the source guide, You should see I've given that, but if you can't see here and would like to know how, there's a way to add sources to it. So you go add, and it looks like this.

So in link, maybe I'll just do it again just to prove a point. Control this. And I'll just add a second time. And now this has two copies of the same source, but that's how simple it is.

And then now that we've got this conversation, we've got this list. Now, what this list contributes, and what's also about this list as well is Notebook Client's amazing feature is it will take you to the relevant piece of the document that has the data, which in this case is literally the description attached to the map when you hold the cursor over it, like really impressive visual processing.

Producing AI Research Summaries

What we're now going to do is we're going to copy and paste this. as a prompt for perplexity, again, using a different, that one's already done. Yeah, we'll try this one.

Please produce a summary of the most influential papers from these labs, which are all Cambridge-based on AI safety recently.

We're going to, again, filter for academic papers. Where's that gone? I think it's because it's a second use.

This is an existing conversation. So I need to copy this and start a new chat. So we'll try that again.

And now we're going to filter for academic papers because that previous chat already had its own filters on. I couldn't change it once it started.

And now we should get, without putting on deep research, without putting on that extra reasoning cost, we have used Notebook LM in the meantime, but Notebook LM's default is Gemini, which is a base model, not a reasoning model. We have... created references to a handful of academic papers from Cambridge.

That one's from Boston. That's not the best. So we want one from the Kruger Labs.

So we're still doing themes. OK, this one's Caesar. This one's Caesar.

This one is mixed paper. from Nanyang and Singapore University. Okay.

Benchmarking and Performance Comparison

Then I tell you what, for relevance sake, let's just benchmark the same prompt, which I think is still a good prompt. I think these are interesting organizations we should try to use, but for the sake of performance, let's use a deep research token.

So again, we're going to a new chat. We're going to put on research query.

Maybe I ran out of them. No? No, I haven't. No, I did. I ran out of them today. Okay, that's sad.

Conclusion

But hopefully you get the idea. Thank you very much.

I think we all enjoyed some pizza.