@garius.bsky.social "The millennial 18-24 cohort" err www.londonreconnections.com/2025/diana-q...
AI writer at the Economist. I write about it, that is. I’m still human. One of literally dozens of people online who is not American.
38,702 followers 435 following 2,203 posts
view profile on Bluesky
@garius.bsky.social "The millennial 18-24 cohort" err www.londonreconnections.com/2025/diana-q...
www.economist.com/science-and-... come work with me
quite aside from the normal complaints about AI art, it sucks because the wispy lines and indistinct shapes in the background are just terrible for colouring in!
took me a while to clock it but the colouring-in page my daughter got at lunch yesterday is AI generated
Elden Ring is 3D hollow knight
We should get more data on that
But Salome there’s a train now!!!!
Taking a deep breath as I shove a knife in, reminding myself it’s not a toaster
Anyone know how to get cheese out of a Yoto slot
“Oh neat I’ll head to Uniqlo to pick it up then”
@kierongillen.bsky.social
I haven’t even seen the original post lol, I just logged on this afternoon and everyone was back on my favourite topic
The US left knows in its heart of hearts that “being American” is a privileged position and cannot abide it
the reason there is so much arguing between brits and americans on here is because the english are the only people on the world that leftwing americans feel absolutely no shameful compunctions against being xenophobic about, and unfortunately, as a people, americans /long/ to punch collectively
If I were really brave I’d have listed some examples
Aside from anything else has he looked at a map. Most of England is on the outer side of Epping from the sea
I think you should have to choose between “the government shouldn’t cater to the manifestly false beliefs of voters” and “here is a manifestly false belief I have which the government should cater to”
This is also why I think porn/erotica/smut could be an exception. The market for those things is largely people who never discuss their consumption with their peers.
I don’t like to say “never” but this is one of the predictions of the AI sector that seems most off to me. People like experiencing the same culture as their friends; “personalised”, beyond the level of a game, seems like a dead end
You couldn’t make life of Brian today. You’d be sued for copyright infringement by Michael Palin. Hypocrisy.
a lot of jargon and complexity and stuff about compliance here... but what it really boils down to is that the US can lean on Microsoft to get access to UK police data, and UK forces don't see an alternative
Nitrous oxide driving, with drivers using balloons behind the wheel, feels like it’s on the rise in London - backed up by what medical experts told us about referrals. But the scale of the issue is obscured because of how crashes are recorded in police stats. www.londoncentric.media/p/nitrous-ox...
I remember going to Blizzard for a Hearthstone event and deeply wanting to nerd out with the designers about the card game space and they wouldn’t even acknowledge Magic the Gathering exists
Just wind the sea backwards
Downsides: • Criminals can steal your car with a laptop • You can drive off without your keys Upsides: • don't have to put your hand in your pocket
keyless ignition is one of the all time worst technological trade-offs ever
_this_ feels like the key one, yeah, and goes back to a problem that _is_ about the intelligence rather than the tokenisation: how do you teach the systems to reliably recognise their own inabilities
Yeah, I think "different class" rather than "very narrow" is a better way of talking about it. I haven't actually looked at whether LLMs are good at Regex, and how they deal with tokenisation when they're doing it
i'm not an ai
"London, Paris, New York and Tokyo are the four best holiday destinations in the world"
this is the closest i can get to getting my head around why "count the Rs in strawberry" is a fundamentally different class of problem from almost every other LLM weakness, because it is bound up in tokenisation rather than the neural network itself
If I give you a katakana keyboard and a katakana dictionary and use the syllabary to write phonetic English to communicate with you, you would be able to perform most knowledge work to a high quality. You would also be unable to count the number of latin characters in an arbitrary word.
I think it’s perfectly possible to run something like Character.AI at a profit on a $5/m subscription, particularly if post-bust there’s oodles of spare GPUs being let out to cover the cost of electricity!
It’s tempting to read stories about ChatGPT-driven psychosis and think that OpenAI is as irresponsible as it’s possible to be in this space and oh boy that would be nice
A bubble popping probably would get rid of lots of the annoying “do this in AI” buttons that have metastasised across tech. But it would also mean the only people left doing AI research are smaller firms without resources or inclination to do a modicum of responsible development
Think a lot of people assume that if the AI bubble bursts, the things they hate will go away, and I worry that’s almost exactly the wrong way around
Stared at this for far far too long trying to work out what @gshowitt.bsky.social had rebranded to
It’s not the norm for unincorporated sole trades, no. I cannot buy a new laptop on credit and then offset the payments against my freelance income
I was in New York magazine as a Bro Who Likes Taylor Swift before the kids next to me at the Eras Tour were BORN
A lot to say about this but one thing to note is that spacex cannot in fact put a man on the moon
Ah see podcasts are useful because sometimes I can’t read, like when I’m cutting food or hoovering
One of my longer-running bits right now is to insist that I don’t need to watch TikTok or YouTube “because I can read”
Oooh hadn’t considered that now’s a really good time to trawl eBay for old Game Boy cartridges
Right
i thought there was No Money Left
www.bbc.co.uk/news/article... why aren't we slapping a £3,750 tax on all the other cars instead
they should simply give ed miliband, the only effective cabinet minister, all the other jobs
yeah but i forgot
Unzoomed #588 1/6 🟢⚪️⚪️⚪️⚪️⚪️ unzoomed.com
I thought the hard thing about giving up Amazon would be paying more but actually the hard thing about giving up Amazon is not getting a little message at the top of the screen saying “you already own this dummy you bought it last year”
I found the guy and I’m mad
"Comrade! The English software engineers do not even earn enough to pay for a table! They eat on the grass with only a single blanket to shield their food from the soil!"
sorry i know it's very serious but i am so charmed by the north korean hacker* using claude to understand what a picnic is www.anthropic.com/news/detecti... * fraudulently-employed-outsourced-software-engineer-with-uncertain-intentions
invented a guy to get mad at: person who calls LLMs plagiarism machines and posts links to my work through archive.ph
saw this retweeted three or four times before I realised it wasn't encouraging me to get down and dirty with a hunk on savile row
Local democracy reporting service you’ve done it again
I think I’ve negatively polarised myself into hope
How the fuck is this the way I hear the news
yo
spotify launches a new "spam" feature www.theverge.com/news/765771/...
I think the one bit that pumps people's intuition is linking it to brain scans/one to one simulation of a specific person. cf Lena
oh i'm not talking about now. but. I think people who are eager to reject the conversation now are occasionally a little hasty to say "never" when the downside of doing that wrongly would be bad
One of the things I want to do if I'm given like… a week to make an underwhelming video, is do a single pass through an LLM by hand on pen and paper.
and the awkward thing is i'm not entirely confident you can't simulate consciousness in silicon?
one fun corollary about machine sentience is that if it is true that you can simulate consciousness in silicon it follows that you can simulate consciousness in a sufficiently large stack of paper
should you care more about the suffering of 10^100 instances of GPT-5 Pro or one shrimp
main issue with it is it's called X now
x.com/ESYudkowsky/... tweet's held up huh
Like, there are exact cognates of this problem in human language. If I tokenise “how many Ls in ‘Lyla’” into katakana, it’s basically impossible to answer
Thanks David
But it doesn’t! It’s very very specifically a problem with _tokenisers_, in a way that is true of almost no other general problem with LLMs/transformers
I’m nitpicking: the general point (reasoning models are going to be the main product here and that’s bad for the economics) is fine. But I think there’s a desire to generalise a _very_ specific problem in ways that tortures it
Definitely important! I’m not sure that the strawberry problem suggests it makes that confusion (honestly, I’m pretty sure it does not generalise in that direction at all)
I don’t understand, sorry. When is counting the Rs in strawberry while vibe coding?
Yes? So you use code to solve them not an LLM
Ehhh idk, if you give it a specific type of question that is structured specifically to be hard for a tokeniser it can’t answer it efficiently but why are you turning to a tokeniser to solve that question
I don’t think this party trick has any real impact on the estimates for cost per usable token
No matter how many times I am told it I still cannot believe Americans have screwed up saying Bologna this badly
I do this on BA every time and it works 75-80% of the time
The converse of this is why I get aisle every time. I WILL chug a beer and I WILL pee seven times
Used to be you had to travel to Berlin for easy access to cool industrial loads
You have two children
It’s the equivalent flex to a live action kids show just randomly having a massive crowd scene periodically
One of the mad things about Bluey is how often the show will do a single shot from an angle that it never uses again, an incredibly lavish directorial quirk for an animated show
Imagine having one daily. No thank u
Was thinking bout pando. What were that all about eh. Mad. Wouldn’t do it again personally
Judging by my inbox, “actually emailing patrons” is fine dining-coded? I get emails from Luca, the Clove Club, and JKS but not, like, any of my beloved neighbourhood joints where I would actually want to know about popups and guest chefs
If I follow a restaurant on instagram they definitely have my email address from when I booked with them. They could use it! I wouldn’t mind!
Begging restaurants to stop announcing events through a single post on their instagram that gets 6 likes and then squashed by the algorithm so I only learn it happened 2 week’s later when I’m directly visiting their page to try and guess their bank holiday hours cos they don’t update their website
The compete by bidding for the contract at renewal, just like with any other government tender
Here’s my defence of this: it’s abundantly clear the old model of engaging with big tech as a mid-sized economy doesn’t work. Beginning the relationship as a customer is effectively the only way the UK govt would have any meaningful power over OpenAI beyond the thread to eject it from the country.
(Both would be quite good)
BBQ culture is inherently anti-urbanist and as such must be viewed as an ally of the enemy, car culture. It brings me not joy to declare this
Extremely ADHD-coded advice [approvingly]
Please tell me this is not good
I’m glad you’re here, but yeah it’s work
Make sure it's not an emu dressed up as an ostrich or you'll be charged an extra 20%
The "ostriches" exemption is real by the way, it's the specific example they give of an animal that didn't used to be food but is now food. Remember that next time you're buying an ostrich
