Our brains have 100 trillion connections. Large language models have up to half a trillion, a trillion at most. Yet GPT-4 knows hundreds of times more than any one person does.

Geoffrey Hinton, deep learning pioneer

The recent surge in interest in generative AI was sparked by neural networks trained on high quality public, human culture.

Their use of this culture is extremely focussed – they only saw good quality inputs, and only saw each input once (see the paper One Epoch Is All You Need for why). If you show them lots of bad quality stuff, they’re not adaptive enough to tell and ignore it.

So what exactly makes the data they’re trained on “high quality”? I decided to dig into two examples — GPT-3 for text and Stable Diffusion for images — to find out. (I’m not an expert at training networks, so please leave comments about anything you see that I get wrong)

GPT-3 — good articles that Reddit linked to

GPT-3 feels a bit like it saw all the internet as input — actually they started by training another curator network to pick out the “high quality” parts of the internet (see 2.2 Training Dataset and Appendix A in the GPT-3 paper).

What is high quality? Well, the curator network for GPT-3 was taught its concept of high quality by looking at an internet corpus called WebText. That was made by grabbing a copy of all of Reddit, and looking at the outbound links in moderately upvoted posts/comments.

(Details: Read 2.2 Training Dataset in the GPT-2 paper  — that says “at least 3 karma”, which doesn’t make clear sense to me. As far as I can tell it is Reddit users who have karma, not links or posts. OpenWebText2, a recreation of this dataset, used URLs from submissions which have a score of 3 or higher — see their documentation — which seems a good assumption for what GPT-3 did. Possibly they took links from posts by users with a karma greater than 3. Posts are also separate from comments, and users have a separate karma score for each.)

GPT-3 was also given lots of books and Wikipedia — but most (82%) of what it took in was the good pages that Reddit linked to, and other pages that “feel” similarly high quality.

Stable Diffusion — pretty competition-winning images

Once again, this begins with a copy of the whole Internet, filtering out all images that have alt-text attributes in the HTML using them. This is already a filter — well made sites which care about accessibility will have more and better alt-text. A project called LAION then classifies those images
using AI so they can be filtered by language, resolution, chance of having a watermark, and their aesthetic score.

Stable Diffusion’s training is in a complicated series of checkpoints, which starts off with a bit of general LAION data, but ends up mostly trained on highly aesthetic images. This is much the same as for GPT-3 — a curator AI (the LAION-Aesthetic Predictor V2) learns to judge high quality images. Those images from LAION that the predictor scores 5 or higher were used to train Stable Diffusion.

But what does an aeshetic score of 5 mean? For a quick feel, this page shows increasingly aesthetic buckets of images from the full LAION dataset as you go down the page. Digging into the training data used to make LAION, there are two main sources:

1. Simulacra Aesthetic Captions – manually rated images made from an earlier (non-aesthetically trained) checkpoint of Stable Diffusion. Humans, I assume from the Stable Diffusion community, rated a couple of hundred thousand images by hand.

2. Aesthetic Visual Analysis – this is a download of all the images from DPChallenge, a 20 year old digital photography competition site. A few times a week for decades this has competitions like “photograph bananas, any composition, must have at least one banana”. While they only have tens of entries to each competition these days, they used to have hundreds.

There’s a bit more complexity — a small number of logos specially aesthetically rated were thrown in, I think to improve typography and font quality. You can browse all the images used to train Stable Diffusion (browser by Andy and Simon, those links to their blog posts about it).

Conclusion

The notable common properties are:

1. Each piece of training data is only shown once to the AI during training
2. Both have a core dataset with some human-curated metric for “high quality”
3. Both extended that core dataset by training a curator AI to pick out similar high quality items

It’s quite an odd situation, especially given the emergent properties of reasoning, coding and so on that these kinds of models have. The training mechanism isn’t particularly smart, the smart stuff emerges inside the neural networks so trained.

The AIs have learnt to think extremely well about a large chunk of human knowledge in symbolic form. Right now, they are heavily reliant on humans — every upvote you made on Reddit, and an old time niche digital photography competition site.

A three year old draft blog post I just found. Feels worth publishing – the improvements in AI since then if anything make it clearer, and all the “right now” caveats justified.

It’s better to start by thinking of us as pattern matching devices first.

Not simple ones – such as modern deep learning AI that essentially does layered functions to measure correlation. Complex ones, that model causality in a sophisticated way we don’t remotely understand yet.

That’s intuition. Which is an odd word for it, that you’d only invent if you were overly focussed on language, not what we actually are.

So to language. One function of language is to attempt to describe what we pattern matched, our model of the world, to others. To influence them, to train them, to explain yourself to them.

If you’ve learnt a foreign language, or even just lots of varied words for similar concepts in your native language, you’ll know this is hazardous and inaccurate at best. No surprise – a few thousand words even in combination can’t cover conveying the exact logic of hundreds of trillions of synapses.

Rationality and science are attempts to improve our thoroughness and willpower to agree truth, and make explicit our working.

Sometimes this goes reasonably well – academic Maths rarely ends up with persistent mistakes. But it only does this by intense training of people with a very specific ability, and by picking the easy use cases. By definition, Maths is about the cases where logic prevails – and even then there’s Gödel’s theorem to confound that simplistic view.

But generally, it is going to be inaccurate.

You can’t work out everything with data or current AI. You can check you have made mistakes with it. You can feed your brain insights with it. If you’ve got loads of data and lots of resources and you do it really carefully, you can do random controlled trials (A/B testing) and at least be sure about the causal direction of what you learnt.

However hard you try, this won’t ever be a model as sophisticated as only a human mind’s intuition is right now. It’s difficult to use intuition though, because our minds can just as likely be wrong. Good cultural practices of training and validation can hone our minds and insight.

Once we’ve gained our truth, there is only the limited bandwidth of language to try to help others gain it too.

This is a recipe from Phil the Dhal, a friend who has been experimenting with pressure cooking Indian food. I’m posting it here so it doesn’t get lost. Ask if you want this just to be the first of a series!

First pressure cook of split urad dhal was a success.

Needs some understanding and refinement of what you can put in with the dhal in the pressure cooker. i.e. salt and turmeric are uncontroversial, but stuff like chilli powder or garlic and ginger paste or chopped chillies I don’t fully understand.

This is using a Duromatic Inox Frying Pan Pressure Cooker 24cm / 2.5L.

Part 1 – Pressure cooking

1. Wash the dhal about 3 times… no need to soak overnight. Fudco products always take less washing but are slight more expensive. 1 cup of dhal and 3.5 cups of water.

Put it in the pressure cooker.

2. Make garlic and ginger paste. I’ve used a thumb of ginger and 4 plump garlic cloves.

3. Add garlic and ginger paste, 1 tsp of haldi (turmeric) and 1 tsp of pink salt to the dhal.

4. Put the lid on and turn heat to maximum

5. When you see the red line appear thus, turn down the heat to very low.

12 minutes and the pressure cooking part will be done. Turn off the heat.

Part 2 – The tarka

1. Pretty much anything goes here according to taste.

I’m going to use cumin seeds, mustard seeds, medium onion, some red chillies and a large pinch of asifoteda.

I’ve decided to chuck in a few black cardamom pods to help me understand them

Make roasted curry powder as described in this Sri Lankan recipe – or you can substitute with twice as much garam masala.

So, that’s my prep done.

2. I have to wait until the pressure drops naturally in the pressure cooker

Most recipes say to let the pressure drop naturally. No idea why, perhaps the steam is still useful for the cooking / sauce. You can press the valve and release the steam.

Pressure is fully released:

3. I added a bit of extra water and frozen peas. Put on a low heat and bring to a simmer.

4. Cook the whole spices in hot oil until they pop and splutter.

5. Then reduce heat and add the onions. Cook for about 10 mins.

6. In with 1 tsp of roasted curry powder and asifoteda and cook for 3 mins or so.

7. Add the spice mixture to the dhal and stir. Leave it for 15 mins stirring occasionally.

8. Grate in a bit of palm sugar. And stir.

9. Add some fresh coriander (do use the stalks as they are tasty).

Squeeze in a bit of lime and stir again and you’re done!

You could do the recipe very simply, just using cumin seeds, cumin and coriander powder, salt, turmeric, chilli garlic and ginger if you wanted

Urad dhal is very warming.

So, there you go… an easy dhal recipe for the autumn winter months!

This year, the public narrative around Facebook has switched – the company feels on the defensive in lots of ways. I think it deserves to be – with billions of users, it is long past time for them to spend their energy on reducing harm, rather than on more growth.

There’s a bit less talk about YouTube (owned by Google), and the problems with its recommendation algorithm.

The problems

Here are an article and a video which show the span of problems – from causing political radicalisation in every direction, to creating vast farms of weird, abusive videos targetted at children:

1. YouTube, The Great Radicalizer by the excellent Zeynep Tufekci.
2. The nightmare videos of childrens’ YouTube by the great James Bridle.

This is causing pain for content creators too. For example, a board games reviewer I like called Actulol gives some idea of the mental health issues caused by the algorithm in Why Actualol Went Quiet. You can find other examples – ask YouTubers what they think of the recommendation algorithm.

Why it happens

I wrote a Quora answer last year on How does YouTube’s recommenation algorithm work. If you’re technical, definitely read the full paper from Google.

I find two basic problems with the algorithm:

1. Populist. It first of all uses a crude criteria to find videos watched by people similar to you. This means it is pre-filtering for the popular. This is the opposite of what has made Google search a success – where some users mine dozens of pages into search results, and Google uses that signal to slowly increase the rank of good sites. (We watched it find PDFTables and rank it highly by those means).
2. Short-term. After getting those few hundred candidates, it then uses a proper super smart neural network algorithm with thousands of factors fed in to rank the videos based on which you’re most likely to watch for longest. This is extremely simplistic – the idea that the best videos for you to see are the ones you’ll watch the most. Naturally, things that appal or deceive, or that make you unhappy in the long term, will bubble to the top.

Other ideas

It seems that even at a basic level, Deep Mind (the division of Google’s parent company which made the algorithm, and whose offices are but a mile from where I sit now in London), could come up with a better criteria for the algorithm.

As well as a better criteria, Google would really have to want to use it – YouTube is vast in both size and speed of change, making it hard to run whatever algorithms the company wants over the corpus. That feels like an excuse though – they manage to do similar scale work in search. It feels possible, it just needs the budget allocating, so in turn needs the pressure from us.

At Newspeak House the other evening, the topic drifted to ideas for improving the recommendation algorithm. We were just coming up with criteria to train it – I’m sure the boffins at Deep Mind can come up with better ideas, and more interesting technical implementations of them. Here are two of them:

• Feel happiest in 6 months time. It could explicitly ask you – do you feel happy / does using YouTube make you feel good or bad? And train the algorithm on whether that signal improves months later. As well as improving videos, this would be good for the long term brand of YouTube.
• Become higher value to advertisers in the next year. This is interesting, as it sounds like it could make more money for Google. It would naturally tend to push aspirational videos – or videos that lead to aspirational videos – onto people. So they are more likely to get promoted, get a better job, want to buy more expensive things that are advertised on YouTube. It’s not clear this would be good overall but it would be interesting.

Some more radical ideas, going beyond just tweaking the training criteria:

• Get rid of automated recommendations. Instead, they could be curated, perhaps by the person who makes each video. For a more AI based version of this, something like the Spotify model based on the crowd-curation of playlists could help. We used to watch TV stations curated by amazing people like David Attenborough (director of BBC2 in the 1960s). Could YouTube help me do that in a more modern way? If things get bad enough, we could regulate to simply ban robot recommendations outright, and see what innovation human curation leads to.
• Agent on own phone I own and tweak to make recommendations. This is perhaps too demanding on the user, it feels like the way things are going though. The work done on TensorFlow Light to get the clipboard AI to run on Android phones shows that with the right engineering this kind of solution is possible. The end game is like the movie Her. If the customisation is too hard I could follow someone else’s meta-ruleset – maybe a famous brain training coach.
• Split up YouTube. Create competition. Right now I have no choice of recommendation algorithm. All the videos are on YouTube, so I have to go there. I guess I could learn Chinese and move to China and find a different system, but that’s the limit of my choice. How to split up, or regulate the new generation of big tech companies isn’t clear yet – we could find a smart way to do it well.

What do you think, how would you improve YouTube’s recommendation algorithm? What would a smarter criteria be?

A tasty early morning breakfast with friends in town, and time to cycle home. A sunny day, so take the route along the river!

Crossing the road by the docks. The sky impeccable. Delicious light of being away from the tropics, strong and clear, but not overly harsh.

The dome of one of the three graces is just visible, and at the right hand edge the tower which has sucked car pollution out of the tunnel under the river for over 100 years.

I quite like the black lumpine jag into the skyline of the block of flats a friend lives in – it’s an interesting shape, and its thick darkness avoids the rough textures of variety in curtains and decoration of some of the other residential blocks nearby.

I never lived the other side of this road because it is irritating to cross when you’re in a hurry. But on a day like this, you never mind.

Cycling through the old docks of what was the busiest port in the world, and along the front past the gig and convention centres. On my first weekend in Liverpool we blagged some Eddie Izzard tickets here from some students whose aunt had given them, and they just wanted to go to the pub – whisked straight into the show.

The ferry across the Mersey is just visible at its dock. Once on a warm summer’s evening I cycled down after work and crossed on the ferry, to New Brighton, stripped to my boxer shorts and swimming in the sea, building a castle with a bucket and spade. Liverpool often has a brightness you don’t expect.

I love going along this part in varying weather conditions – you can often see and hear the waves. The shipping is always different, and the skies and water changing in colour.

Today the tide is quite far out, and it feels more like the Mediterranean. Here you can see the other tunnel defuming tower over the waters.

Whoever built this path, I’ll always be grateful. The kind of gesture a city does that everyone then takes for granted, but changes so many lives.

The one awkward part, rarely, is here by the marina. On a busy Saturday I’ve had to wait half an hour with sail boats going in and out, before being able to cross the swing bridge!

A couple sit enjoying the view. Maritime parts converted into sculptures litter the path.

I’ve never been to any of the flats down here. The area is more attractive now with the Baltic Triangle flowering with gin bars, rave clubs and hip cafes.

The light speckles, scitter-scattering brightly across the wash. You can just see the high hills across the river.

From slightly further up you can see even more clearly the mountains of a foreign country, exotic names like Moel Famau. Street signs in a celtic tongue.

On journeys home from work through Toxeth and Dingle, zagging along streets with river views, I’d watch them. Happy to know that with a drive of an hour and a bit, I could be up in them. And from there on a clear day I could see further glories – Snowdon and its friends, giants to someone like me from this island country!

A feeling about the wild, the high peaks, the grasses and rocks and views that draws.

Even if you don’t go to something, you like to know it’s there.

Nearly home. Otterspool Promenade – named as there really were once otters here. Maybe one day the dolphins will come back too as we clean the oceans!

(Fact checking just now, it turns out they came back a couple of years ago)

Up and through the festival gardens – too beautiful to have thought to stop and photograph them, so you just get the sign looking back.

There’s a hill in there with astounding views of the countryside. And over further south a rough place to run. Muddy, verdant, uncontrolled, like you’re in the countryside.

I pass up the hill through Priory Wood, a cute place where I’ve had two bike accidents, been attacked by fire crackers and fixed a tent.

My local station has a lovely Victorian building which is also a bridge spanning the rail. Trains every 15 minutes into town, almost always can get a seat.

Commuters flock in and out at busy times, congesting the ramp up from the platform.

Amazing to think that this settlement started with a church, one of the first iron framed buildings in the world, plonked in a field. Expandable because of that architecture, to meet the growing needs of its community.

Years later the station appeared, flood filling terraced houses up to the main road.

I sneak through the alleyway by the station, and arrive at my home street.

Brash in the sunshine. The details of the cast guttering, the shape of the arc of bricks above the doors… Almost 130 years old.

Cars angled on the curbs smashing the pavements. Children born to new families. Political movements grown. Lives lived – harrowing and joyful.

And the best place in the world to have my lunch today.

In the last couple of weeks I’ve:

• Deactivated my Facebook account.
• Logged out of Twitter on my desktop.
• Stopped using Google Contacts, Calendar and Photos

I’m not really sure why, but there it is, I have. I think over the last five years in excitement at the good user experience and high user adoption, I forgot why I was negative about such centralised, out of control services in the first place.

There’s something dehumanising about being subject to them. The behaviour patterns they induce feel harmful to me. They disengage me from the world.

We need to think about and improve our information systems to be good for us, empowering, productive. Just using the default commoditised, VC-funded options leaves no room for that. We’re stuck with whatever we’re given. There’s no choice, no competition, no diversity. It’s dangerously too early to have no diversity in information systems.

It used to be that there was constant innovation and public discussion and experimentation about how social groups are formed and how communication happened on the network. The nearest you see now is discussion of varying moderation policies of subreddits, which is at least healthy. More of that please.

So instead…

• I’m posting on this blog, which uses software (WordPress) I’ve been using for 15 years, on a server run by Bytemark, a northern UK small business.
• I use Riot for some discussions, such as #redecentralize.
• My contacts and calendar have gone over to Fastmail which I was using anyway.
• My photos I keep myself, on the same Bytemark server, synced to my laptop and phone with bits of string and Resilio.

In the main this costs more in cash and time for me to do. Is it worth it? That’s not really the question, it isn’t really a choice.

If you want to keep in touch, perhaps email me or write me a letter or phone me up. Or write a blog post on your own blog, and leave the link in a comment here. Join a channel I’m in in Riot and say hello to me. Invite me to join some new blockchain-based chat system I haven’t heard of yet. Invite me to the pub by SMS. Come round to my house to see me.

Off to not Tweet this. Maybe you could give me a pingback, or leave a comment!

Culture is important.

Particularly, it’s important for technology which is about how we as humans talk to one another.

Telephones, computers, mobile phones, internet… They’re all about how our particular species of primate communicates. And now they’re becoming one powerful industry newly reformed.

We’re each a neural network, thrown abruptly into a chaotic world, which by sheer force and pain and amazement has cohered into the awareness that we are.

I’m amazed when my kitten, now a young cat, first goes out the cat flap[1], climbs on the garden wall, wanders off so I can’t see him. Finds his way back with a dragon fly, tortures it in the kitchen. Emotionally manipulates me with a purr and a rub of his fur on my face.

Yet in this café… Just this one small place in this one corner of Liverpool. In this café, bursting with people with life. Eating brunch, discussing business details, chatting with their loves…

Each of us 50 times more complex[2] than my cat, bound in shared experience to me. Our own thoughts, fears, illness, hope, love, astonishment, as many, as different as me and you.

The user interface we text our family member with.

The algorithm that shows us news.

The affordance of the tool we use to invite friends to our birthday.

The etiquette of when we can phone and whether we can knock unplanned at a home’s door.

The pressure of politeness, or not, in a vital thread.

Tangled bonds of trust, centralising and decentralising in a constant game[3] of tribalism evolved over a span of ages.

I’m reading a book about ants. We’re more like them than we know.

I notice that while I think I use analysis and research (say a new product to buy), most of the time I just do what a friend does (buy what they have). Which is much like how ants decide where to join in the foraging.

A powerful communications industry, newly reforming.

Do we go with the flow, let oligopolistic capital and secret states run their course?

Or do we concern ourselves with how it forms?

We’re ants.

The scents we leave on the trail, the smells we spread when we greet, the frequencies of other workers we count as we pass…

We do all the same things and so much more.

The changing tools we choose to use to manipulate information alter the rules of our herding.

50 kittens worth of cerebral cortex per organism formed into a global colony a thousand times the population of a leafcutter’s.

Humans.

[1] Itself a computer now, only letting his surgically embedded key open it. More on that another time.

[2] Is that linear? Who knows! What does that even mean!

[3] Thanks David for the article link.