This is Geek Week, my newsletter about whatever nerdy things have happened to catch my eye over the past seven days. Here’s me, musing about something I don’t fully understand in an attempt to get my head around it: I imagine that’s how most editions will be. If you’d like to get this direct to your inbox, every single week, you can sign up here.
A thing I’m really interested in is artificial general intelligence (AGI). AIs at the moment are “narrow”. They can be amazingly good, better than the best humans, at a specific task or set of tasks – Stockfish is massively superhuman at playing chess; AlphaFold is massively superhuman at predicting the 3D shape of proteins – but they’re limited to that domain. So AlphaFold can’t, for instance, do your taxes for you. PaLM can understand jokes, but can’t help you navigate from Birmingham to Oxford.
A general intelligence is one that can do lots of things. Humans are an example of a general intelligence, in that we can play chess, navigate to Oxford, do our taxes, understand jokes and many other things. Some modern AIs are getting more general – AlphaZero, an AI made by the Google affiliate DeepMind, is massively superhuman at chess, and Go, and various other games. But they’re still very narrow.
A truly general AI – one that could do all the intellectual tasks a human can do – would be transformative. Whether it would be transformative in a good way (solving climate change and poverty, that sort of thing) or a bad way (killing everyone and turning us into paperclips) is a question I spend a lot of time trying to answer in my first book.
So. The topic is on my mind because DeepMind released results from its new AI, called Gato, this week. And one person (a writer at The Next Web) said that it made him think AGI will never happen, while another, an AI researcher at DeepMind itself, said that on the contrary, Gato shows that “the game is over” and basically all we need to do is make larger versions of Gato to get to AI.
Gato is apparently based on large language models (LLMs), AIs trained on vast amounts of text which can write new text by “predicting” what would come next after a text prompt. The first really famous LLM, OpenAI’s GPT-2 from 2019, could write poetry and bad Tolkien fanfic. GPT-3, its more recent brother, is the same but better.
Speaking to AI researchers, there’s real hope that LLMs are one of the keys to reaching general AI – Hannah Fry’s excellent DeepMind podcast is worth a listen to on that subject. I’m far from an expert but you can see the use of LLMs in things like OpenAI’s DALL·E, which can not only write text but draw images of what you describe (“An armchair in the shape of an avocado,” that sort of thing).
It may not be the whole thing, but AIs that are able to fluently understand human language and respond by doing what we ask of it will certainly seem much more intelligent to us, and at some point it’s not clear what the difference between “seeming intelligent” and “being intelligent” actually is.
Anyway, Gato. It uses LLM-like methods to do more than just output text or drawings. DeepMind describes it as a “generalist agent”, and says it can “play Atari, follow text instructions, caption images, chat with people, control a real robot arm, and more”.
It’s not a general AI. It can’t switch to entirely new tasks; it has about 600 tasks that it has learnt, but if you said “Book me a flight to Sicily” it wouldn’t understand. But I think it’s a small step in the direction of broadly useful AIs that could, for instance, be in charge of a house. You come home, say “turn on the lights and preheat the oven, and get me a recipe for tarka dahl,” and the AI does those things.
(This is complete speculation, but it feels plausible to me.)
Why this exciting new AI research shows that AI research isn’t working
Anyway, the guy from The Next Web thinks this is really underwhelming, and that it seems “like AGI won’t be happening in our lifetimes” if DeepMind has been working on AGI for over a decade and still can’t “address the very first problem on the way to solving AGI: building an AI that can learn new things without training”.
Nando de Freitas, the DeepMind guy, on the other hand, thinks that “making these models bigger, safer, compute efficient, faster at sampling, smarter memory, more modalities, INNOVATIVE DATA, on/offline” will be enough to “deliver AGI”.
As a journalist with a philosophy degree, I’m probably not the person to ask to settle this. But I did have some thoughts.
As an interested non-expert looking in, the progress in AI has been absolutely astonishing. A decade or so ago, getting an AI to reliably tell pictures of cats from pictures of dogs was cutting-edge stuff. Now they can look at a picture of a horse in a gold-and-blue bridle standing next to a wall and label it “horse in a gold-and-blue bridle standing next to a wall”. I did my interviews for my book in 2017, and when I recently spoke to an AI researcher at DeepMind they told me that I’d been right about most of the stuff in it, but it was pretty out of date now because of the rise of LLMs. An LLM-like AI, AlphaCode, can now write mediocre-human-programmer-standard code.
A decade isn’t a very long time. It seems really strange to say “DeepMind has been working on this for a decade and they haven’t cracked it yet, ergo they won’t.” People have been working on AI for seven decades and it certainly seems like most of the progress has been in the last one.
But more importantly, I think the consensus is that recent news on AI should make us think that AGI is closer. I’m taking that from the forecasting site Metaculus, which over April and May this year has moved its median forecast of the date that “weakly general” AI is publicly known from 2042 to 2028. Over the last few weeks we’ve had DALL-E, DeepMind’s Flamingo, now Gato. Things that seemed impossible a few years ago are becoming routine. And speaking to people who know more than me, the impression is one of gathering pace, rather than dead ends.
Maybe I’m wrong. Again, I’m not an expert, I couldn’t code my way out of a wet paper bag, and my grasp on how AIs actually work on a nuts-and-bolts level is somewhere between “conceptual” and “non-existent”. But the idea that we should be taking recent breakthroughs in AI as evidence that AI progress is grinding to a halt seems amazing to me.
Self promotion corner
This week I’ve mainly been thinking about gene therapy, and whether it can treat difficult and intractable hereditary diseases like cystic fibrosis and Huntington’s (it’s already working to save babies diagnosed with the horrible hereditary disease spinal muscular atrophy). I am a sucker for “advanced medical technology” stories, and I worry I get swept away with the excitement, but I do think they’re pretty cool things and will probably end up saving a lot of lives in future. Some feedback I’ve had since publication is that there’s a worry about off-target effects – “We’ve cured your cystic fibrosis but, whoops, we’ve given you cancer” sort of things – which actually one of my interviewees did mention to me, and perhaps I ought to have mentioned. Being the bright-eyed techno-optimist I am, though, I think they’ll probably fix that at some point.
Blogpost of the week: The Nerd as the Norm
Absolutely formative blog post, this. John Nerst is a wonderful blogger, brilliant at coming up with new ways of looking at the world, and he’s fascinated by why people disagree.
One of his best works was The Nerd as the Norm. It was inspired by a piece about how to deal with nerds in the workplace.
He thinks of himself as a nerd (as do I), and he felt it was strange that there was this sort of background assumption that there are two kinds of people: nerds and non-nerds; nerds and normal people.
But, he points out, if you take some stereotypically nerdy personality traits – he suggests “an interest in things and ideas over people”, “a concern for correctness over social harmony”, obliviousness to and/or disregard for social norms and expectations”, “difficulty appreciating the social implications of their actions” and a few others, then you don’t get a distribution where there are lots of people who have the normal amount of those things, and then nerds who have lots.
Instead, personality traits fall on a bell curve, a spectrum. So you’ll get people who are on the other side of the curve: people who have “an interest in people over things and ideas”, “a concern for social harmony over correctness”, “sensitivity to social norms and expectations”, “difficulty appreciating the logical implications of their ideas”.
And, says Nerst, we should come up with a name for this sort of hypothetical anti-nerd:
“Anti-nerd” is pretty clunky, so I made up another word. I tried to come up with one that sounds right the way “nerd” sounds nerdy (a kind of prickly tenseness). So I guess something with a gooey, shapeless feel? How about “wamb”? Does that sound good? Well, it’s what I settled on after a couple of Saturday afternoon Irish coffees and screw you if you don’t like it.
I think it describes a real phenomenon: there definitely are nerds who can’t work out why saying true-but-socially-awkward things is not always wise; but there definitely are also wambs who can’t work out that just because something’s socially awkward doesn’t mean it’s not true. And I genuinely find myself using the word “wamb” quite a lot in conversation, so I clearly had a need for it.
This is Geek Week, a subscriber-only newsletter from i. If you’d like to get this direct to your inbox, every single week, you can sign up here.