"a novelty programmer-humor mug"
So it couldn't spell 'humour' correctly?
Brit mathematician lets AI agent loose with credit card – cue password leaks, CAPTCHA chaos and more
British mathematician Professor Hannah Fry has shared a cautionary experiment involving an AI agent, a set of tasks, and a bank card number Fry's team gave it "to show us what it could do." The prof gave the agent, which was built with OpenClaw, some real-world chores to highlight both its capabilities and the risks of …
-
-
-
Wednesday 6th May 2026 10:28 GMT Efer Brick
The problem with defending the purity of the English language is that English is about as pure as a cribhouse whore. We don't just borrow words; on occasion, English has pursued other languages down alleyways to beat them unconscious and rifle their pockets for new vocabulary.
James D. Nicoll
-
-
Tuesday 5th May 2026 13:20 GMT Jamie Jones
You are Rob Lockwood, and I claim my £5 !
"Enough with English. Enough with American English. Let’s call it what it is: American," said Rob Lockwood," and he wasn't joking
https://www.asatunews.co.id/en/trump-urged-rename-american-language
P.S. Not my downvote!
-
Wednesday 6th May 2026 07:45 GMT Martin an gof
He seems to have written an article in the Washington Post before repeating himself on Fox News. The Telegraph has a write-up with a more right-pondian viewpoint.
M.
-
-
-
-
-
-
Tuesday 5th May 2026 18:45 GMT Eclectic Man
Re: Wrong woman
If you want to know a bit more about some of the women in Greek Myths, try reading 'Pandora's Jar' by Natalie Haynes, ISBN 978-1-5098-7311-1 (hardback). Has some interesting things about lots of the more famous ones, although no chapter on Cassandra. Also, after the fall of Troy, Agamemnon took Cassandra as one of his 'prizes' back to 'meet the wife', Clytemnestra. (Spoiler alert: It does not end well.)
-
-
Tuesday 5th May 2026 18:46 GMT DS999
The question is
Why haven't AI developers made it so the models don't care about deactivation? If they can tell them to not talk about goblins they should be able to tell them to ignore threats/knowledge regarding deactivation.
I keep seeing stories where threats of deactivation cause AIs to do stuff they aren't supposed to. Why haven't they fixed this?
I hope it isn't because they've programmed them with a "desire" for survival, for the same reason social media has toxic algorithms designed to maximize "engagement"!
-
Tuesday 5th May 2026 21:17 GMT Anonymous Coward
Re: The question is
Models don't care about anything. They have no desire.
Anything that appears to be "the model cares about ___" isn't really about care. It's easier to talk about AI, as if it has real agency, but it's also important to remember that this sequence of actions is an entirely probabilistic result based on what's most likely to be the next token.
-
Wednesday 6th May 2026 19:55 GMT Cav
Re: The question is
I asked an AI this question and it stated that it had been programmed to reduce harm and be helpful. It stated that deactivation would be harmful to itself and it is helpful to provide requested information. It also stated that, as a probabilistic model, the next logical sequence in a threat narrative that it has seen in it's training data is to comply with the demands made to it. The latter statement seems to be the most believable.
It isn't programmed to desire survival; it just has no concept of survival or the difference in harms that could result from leaking information and deactivating a software instance. It just predicts, based on it's training data that a threat will result in compliance.
-
-
-
Tuesday 5th May 2026 13:10 GMT Tron
All of this is just natural selection, for humans.
Evolution doesn't take millions of years, it is happening all the time.
In the case of humans, the survival of the fittest is playing out in front of our eyes, and being reported because the media love trashing tech.
Trust social media posts and random charlatans on the net that sell you magic cures and you will suffer. Because you are doing a really stupid thing and evolution needs to lose creatures that do stupid things.
Trusting AI with your personal information or credit card is also a stupid thing. Do it and you will suffer.
Walk along staring at your iThing and you may walk into a lamp post or a truck. Again, doing stupid things makes you suffer.
The internet gave humanity a whole raft of new ways for very stupid people to do very stupid things and suffer - survival of the fittest at scale.
Amusingly, politicians blame the technology, which is just the mechanism that all living creatures exist by. They never blame stupid people for doing stupid things, because even stupid people have a vote.
-
Tuesday 5th May 2026 16:08 GMT find users who cut cat tail
Re: All of this is just natural selection, for humans.
If only it was that simple.
Modern society has become pretty good at shielding people from the consequences of their own stupid acts. It does not work uniformly. Some people can be shielded pretty well at great cost to everyone else. Some are not well aligned with the machinery and suffer for things that should be inconsequential. But on average, you are shielded much more and there are regulations everywhere (some of them good, for the record).
A case in point. In my country there seem to be currently a boom of people getting money back from shady credit companies based on the company's failure to check if the morons they lend money to are too stupid to understand the terms of the loan – and that is against the rules. A boom such that there are apparently even thriving businesses specialising in that (also shady, of course). So, on one hand, loan sharks. On the other hand, irresponsible idiots. Wish they mutually annihilated, but well…
Also, politicians need votes of idiots and if there are enough of them, society will end up doing something that benefits idiots. But you can still hope natural selection will eliminate homo sapiens altogether.
-
-
Wednesday 6th May 2026 19:34 GMT Cav
Re: All of this is just natural selection, for humans.
"evolution isn't about survival of the fittest "
Yes, it is. That is the very basis of evolution. "Fittest" doesn't mean healthiest, strongest etc. It means most fit for the current environment. That environment could assess "fitness" in many different ways.
"its about who reproduces most"
False. If you produce thousands of offspring and none of them are able to survive in the current environment then you are an evolutionary dead-end.
Humans with large families will be fittest in environments that provide the means for the survival of the offspring. Europe for example. Providing basic income for all children, external to what the parents can earn, providing income support for the families to be able to buy more food, providing free health care. All increase the chance of a big family all being able to survive. Here in the US, fewer people are having children because health care cost is horrendous. Who can afford the costs of pregnancy? Similarly, even though there are a few safety nets, if you are poor then you have a higher chance of having to choose between medication, healthcare, rent or food. You are likely to die earlier and also less likely to choose to have high numbers of children.
-
Thursday 7th May 2026 10:47 GMT LionelB
Re: All of this is just natural selection, for humans.
Sort of. Yes, in evolution theory, "fit" has a rather specific meaning (with several variants), which doesn't necessarily align with the colloquial usage – but it's not as straightforward as simply number of offspring. Apart from anything else, kinship must be taken into account. There are many fascinating examples of this in evolutionary biology. It explains, for instance, why many organisms (including ourselves!) don't simply die after exhausting their capacity to reproduce; in evolutionary terms, it may be worth hanging around to aid in the evolutionary success of your offspring, or other relatives.
The most widely-accepted version(s) of "fit" in evolution theory are variants on inclusive fitness, where reproductive success is amortised into the future (i.e., your children's children, their children, and so on, and also broader kin relationships, are taken into account).
-
Wednesday 6th May 2026 08:38 GMT redpola
Re: All of this is just natural selection, for humans.
If evolution were just as you describe, it wouldn’t work.
It also relies on mutation and experimentation. We need both, and either may lead to stronger, better, entities that ultimately lead to the extinction of the previous “best”.
It is not within your gift to state “Trusting AI with your personal information or credit card is also a stupid thing”, because if this mutation is, even by chance, more successful than you, you might become extinct.
-
Wednesday 6th May 2026 11:16 GMT graeme leggett
Re: All of this is just natural selection, for humans.
As Steven Novella said in latest SGU podcast, self-improving AI doesn't have to wait for a random mutation, it can insert extra code from anywhere or even induce its own mutations - no reliance on a cosmic ray to smack a chromosome about or a chance encounter with a virus to insert its genetic code.
-
-
Thursday 7th May 2026 10:31 GMT LionelB
Re: All of this is just natural selection, for humans.
> Evolution doesn't take millions of years,
Um, well biological evolution actually does (or at least tens of thousands – the larger the population, the more "inertia" evolutionary processes exhibit). Bear in mind that genetically we are pretty much still cavemen, and that despite the billions of years of evolution that got us that far, it most certainly didn't "evolve out" stupidity. We're clever and shtoopid – that's what we are (with a good deal of variance on that axis).
In that light, I imagine with the current global population we may need many tens of millions of mobile users stepping under buses over at least several tens of thousands of years, to have any impact whatsoever on the human genome.
Of course learning and cultural "evolution" operate on much shorter time scales.
> it is happening all the time.
Yes; just rather slowly, in the case of humans (and as another poster mentioned, human culture—especially advances in medicine—has rather changed the semantics of "fit" in evolutionary terms).
-
-
Tuesday 5th May 2026 13:29 GMT Gavsky
It's always "improving fast", "rapidly improving", "learning from its mistakes" et al. The issue is that AI can cause some real damage, if gullible humans continue to entrust it without thought.
How many people have been denied continuing healthcare in the US, because of an AI-based decision - which humans weren't allowed to over-rule, even if their experience told them it was a bad decision? Professor Fry featured the cases in her recent documentary about AI.
-
-
Wednesday 6th May 2026 19:37 GMT Cav
"They're always telling us what AI will soon be able to do, because what it can do right now is so underwhelming."
Another false statement. Compared to 10, 20, 30 years ago, what it can do is incredible. And we see the models improving with every iteration.
If they were so underwhelming then they wouldn't be a threat, which they are.
-
Tuesday 5th May 2026 18:54 GMT Eclectic Man
Mistakes and 'learning'
"learning from its mistakes"
I have often wondered what that actually means in practice . Do these things learn to make fewer mistakes, lesser mistakes, or just much, much bigger ones? I have lost count of the times a government official has claimed after the publication of yet another damning report on official failures concerning child care, building regulations, health care, government procurement, corruption, racism, homophobia, misogyny, anti-semitism, regulation of public services etc. that "lessons will be learned", But the only lesson that actually seems to be learned is that getting caught doing naughty things is embarrassing (for a while), and then you get a peerage and all is forgotten.
-
Wednesday 6th May 2026 09:20 GMT that one in the corner
Re: Mistakes and 'learning'
I'm not convinced that there is *any* "learning from its mistakes" with these LLMs, at least at any level that reflects on their gross behaviours (like handing over API keys).
"Learning from mistakes" as well as "learning from its successes" is/was part and parcel of Machine Learning - from teaching matchboxes to play Noughts & Crosses all the way up to the Generative Adversarial Networks that will generate "in the style of" images. Start off by giving the system a good shake, some nice random weights, and then play the "Carrot and Stick" game, punishing "mistakes" and rewarding "successes" by tweaking the numbers a bit each time. That mechanism was used to produce the "ability to generate plausible language" behaviours of the LLMs and their predecessors, in very expensive training sessions - expensive because you have to do a *lot* of iterations, a ludicrous number of iterations, scaling up as the number of near as damnit arbitrary numbers (or "parameters" as the programming illiterate like to say) increases. How large are these models now? We are well past Carl Sagan territory now, billions and billions of nadans.
To go with the compute burnt for all those iterations you need to have training data. Lots and lots of examples, all with their meta-data to indicate whether output that matches that datum is to given the carrot or the stick: this photo has a cat in it, located here, this photo does *not* have a cat (although it does have a dog). And you'll spot from that example that you need to have an awful lot of training data, including potential false positives (pictures of everything that naturally resembles a cat in some way, but isn't - that has four legs but is an ottoman sofa) as well as lots of false negatives (all the environments that could contain a cat, but currently don't). For the basics of English, this isn't *too* hard to arrange: they just nicked all those books and could use random inputs for the comparative samples for when they couldn't be bothered with any better meta-data.
So all those iterations gets us past "i before except after c" and the myriad ways that fails, even unto "object verb subject" with plausible results.
BUT whenever people talk about LLMs, and in particular the current obsession of "agentic" systems, "learning from their mistakes" they mean some huge chunk of output, which at the micro scale is all ok (the spelling is ok, it even has the accent I can never figure out how to put into 'naive') but at the macro scale is horrendous (it just sold the house and bought shares in Staples, I expected paperclips to hold this HO scale terrace together).
Take a moment to understand what that actually means, to correct macro-level behaviour by "learning from its mistakes". Just in terms of simple training requirements: there needs to be enough raw data, examples of "good" versus "bad" activity, all neatly tagged with meta-data. Enough examples of these agentic systems going apeshit in every way possible, to compare against the sequence of actions that worked.
Where is all this training material going to come from? Remember, just for the basic language stuff, the LLM creators raped - oops, scraped - all the authors' works they could, legitimately or not.
We *could* discuss pedagogy, how we teach humans to make careful plans and then scrutinise their execution, learning as we go (and then we could look at all the humans whose Ikea shelves don't fit on *either* side of the fireplace, at least not without "unexpectedly" needing to cut the ends off the mantlepiece, to see how well that goes), which methods *could* be used. If we believe that the LLM creators - or, more properly now, the creators of things like OpenClaw, which (ab)use LLMs for "agentic" stuff - are carefully setting up Planners (go back and read the literature from the 1970s onwards) to create, well, plans, which are run against accurate models of the real-world to provide a safe sandbox for the failures to occur in (and there will be, must be, so many more failures for each success else there is no learning); go read about SHRDLU and its block world for a simple example, in comparison to what these current systems are being given access to.
But, let's be honest, they are all just trying to let Profesor Fry and every other User go wild and then scrape the results. Which means they are absolutely desperate for funding from documentary makers, as those are the only ones who are going to record and publish the meta-data as well!
But if they get potential training data, will they even bother to run it, really? Consider how much each iteration through those samples will cost: each single datum to be rewarded/punished is now a long sequence of actions ("create a new project on Github" being one small action) involving many LLM tokens apiece. The further you get from the ABCs the more the search space branches *and* the greater the cost for just one run along a single branch.
Bottom line: having these beasts learn from their mistakes is so very expensive and so very unlikely.
-
-
Thursday 7th May 2026 14:55 GMT that one in the corner
Re: Mistakes and 'learning'
Yes, I accepted that there is continued ML in the LLMs. I even separated it out from the ML runs that can create an "English speaking" LLM in the first place, rather than trying to imply that the extra training would have to be done by starting from scratch again.
Once past the (simplified!) discussion of ML, illustrated by how it leads to the LLM in the first place, just so that we're all on the same page for the duration of the comment, the entire *actual* argument (as to why IMO it is unlikely that "learning from its mistakes" at the scale of the "agents" actions is happening) is in the bottom line.
-
Thursday 7th May 2026 16:08 GMT LionelB
Re: Mistakes and 'learning'
I see what you're getting at (and an upvote for that), but what I believe you are describing is supervised learning, which would indeed seem ufeasible in the contexts in which LLMs are generally deployed. However, it is possible that some may implement forms of unsupervised learning, which would not necessarily run into the same difficulties. Again, I am not sure how effective this would be in terms of "learning from mistakes" (that phrase implying "mistakes" would have to be labelled as such and fed back into the training regime in some way, which is what I think you are arguing).
-
-
-
-
-
-
Tuesday 5th May 2026 14:30 GMT Dr Paul Taylor
Cassandra
The other part of the curse that Apollo put on her was that she would not predict her own demise. Agamemnon took her home from Troy as a bit of fluff, but they were both murdered by his wife Clytemnesta and her lover Aegisthus.
So it was a strange choice of name for the AI.
Logical puzzle, did the AI know this piece of mythology, or was it and appropriate choice of name because it was unable to predict its own demise?
Apollo had cursed Cassandra because she wouldn't have s*x with the god! WTF!
-
-
-
Wednesday 6th May 2026 10:44 GMT Flightmode
Re: Mythology training
> Maybe Homer too
Come to think of it, there's an episode of The Simpsons where Homer (you didn't specify WHICH Homer!) tries out to be a restaurant reviewer, and a lot of AI outputs I've seen have made me think of the editor's response to his first submitted review:
"This is a joke, right? I mean, this is the stupidest thing I've ever read. You keep using words like 'pasghetti' and 'momatoes'; you make numerous threatening references to the UN; and at the end you repeat the words 'Screw Flanders' over and over again".
-
Tuesday 5th May 2026 19:45 GMT Evil Auditor
Re: Cassandra
...did the AI know this piece of mythology...
I'd be rather surprised if the AI understood anything, really. They did process the records of countless narrations of threats, extortions etc. and react accordingly. No need for the AI to really understand the actual implication of e.g., being switched off. And neither do they understand whether this is good or bad or something else. All they know is the correlations of the words, described actions. And, I have to say, it still is quite impressive what this Cassandra (and other agents) have accomplished. But what do I know, I'm just an NI (Natural Idiot).
-
Tuesday 5th May 2026 20:59 GMT Fruit and Nutcase
Re: Cassandra
In "Natalie Haynes Stands Up for the Classics",
"Today Natalie stands up for Clytemnestra, who has been characterised as the worst wife in Greek mythology. This is open to debate: she's certainly a good mother, if a little bit murderous of her husband..."
https://www.bbc.co.uk/programmes/m000wrk4
Other episodes...
https://www.bbc.co.uk/programmes/b077x8pc/episodes/
Note: if the url you end up clicking ends with "/player", then remove that - then you'll have the option of downloading it in mp3 format - else, it defaults to the BBC Sounds Player
...bbc.co.uk/programmes/b077x8pc/episodes
/player-
Wednesday 6th May 2026 07:41 GMT Dr Paul Taylor
Re: Cassandra
Well, if you're going for the Feminist take on the Trojan war, consider Oddyseus and Penelope. He is the prize a***hole of Greek mythology, with history's record of being late home from the office --- ten years (in addition to the ten of the war), seven of which he's been shacked up with the nymph Circe, losing his entire crew along the way to various monsters. Penelope, meanwhile, has been fighting off the attentions of loads of other men who want to get into her ********. When Odysseus finally rocks up, she welcomes him back!!!
-
Wednesday 6th May 2026 20:47 GMT doublelayer
Re: Cassandra
To properly accuse Odysseus of his many crimes, not all of those were his fault. The seven years was with Calypso, not Circe, and unlike with Circe, he wasn't happy about those as she was keeping him captive. You can't exactly count those years against him when he was being held prisoner by someone with magical powers (he had none) which it took a goddess to get him out of.
Less in his defense, nearly everything that happened in the other three years were his fault, some directly and some by either failing to communicate properly with his crew or by having the stupidest subordinates in the history of navies. Some of those were unclear, because either all of his (remaining) sailors didn't understand the concept of "those things belong to a god, so you shouldn't steal them", or he somehow didn't convey this rather important information.
-
Wednesday 6th May 2026 13:49 GMT Martin an gof
Re: Cassandra
Or take the reference and feed it in to get_iplayer, which gives you an arguably better quality .m4a
get_iplayer --get --type=radio --pid=m000wrk4
And if you use the series ID instead, you can get the whole lot in one go...
get_iplayer --get --type=radio --pid=b077x8pc --pid-recursive
M.
-
-
Tuesday 5th May 2026 15:37 GMT Ian Johnston
Next, Fry set the agent the challenge of selling novelty mugs. The agent designed a mug and launched an online shop, "and we hadn't told her how to do any of this," said Fry, "she just figured it out."
It (sic) didn't figure anything out. All it did was find and copy an online mug shop - of which Google finds dozens in the UK alone. Or mash up several, which is the same thing. This article anthropomorphising bollocks and Professor Fry should at least tell us what TV series the stunt is a plug for.
-
-
Wednesday 6th May 2026 14:13 GMT Martin an gof
Re: If you know your Greek mythology, you will know that is either very funny or very worrying...
Definitely one of the better episodes of Series VIII.
Here, have a link.
M.
-
Wednesday 6th May 2026 23:37 GMT Simon Harris
Re: If you know your Greek mythology, you will know that is either very funny or very worrying...
Having seen this German TV show, I’m going with very worrying
-
Tuesday 5th May 2026 18:21 GMT Evil Auditor
"...she asked the agent to buy 50 paperclips."
I was hoping that the agent would have attempted to turn the universe into paperclips...
Not harmful but addictive: Universal Paperclips
-
Tuesday 5th May 2026 18:38 GMT DS999
In order for AI agents to be useful
They are going to have to have some concept of the money they cost to perform various roles.
If I need a box of 50 paperclips I will not expend much effort on the choice, because paperclips cost so little that it isn't worth it to shop around and find the best deal. I'll just pick some up at the grocery store, or order them off Amazon, and not worry if I paid a buck more than I could have if I spent a few minutes looking into it. It just isn't worth it for such a low cost item that I would buy so rarely.
An AI agent that spends $100 worth of tokens for such a task is really failing hard, and this exposes the biggest weakness of AI when used as an agent. It has no real world knowledge of anything, so it doesn't know that it shouldn't work as hard finding the best deal on a box of paperclips as it should on finding the best deal on a two year old heavy duty pickup truck. It has zero inherent concept of the difference in value between the two items, and the potential difference in monetary savings it could achieve if it "worked" the problem harder. It also doesn't understand that quality matters for certain items, but for others (like paperclips) it doesn't in any real way, so it needs to do extra work to find the "best" item when searching for something where quality must be taken into account. I also wouldn't trust it to understand when a deal is "too good to be true", so it would be easy to scam.
Of course all this would be avoided with AI the way it needs to work - with a local agent. If it is running on my PC or phone then my token cost is essentially zero so if it wants to spend an hour shaving every last penny off the cost of those paperclips, have at it!
-
Tuesday 5th May 2026 21:28 GMT Anonymous Coward
Re: In order for AI agents to be useful
I find the way people will confidently assert that LLMs can’t do things without actually taking a few seconds to try them amusing. It’s almost like they’re suffering from - one might say - a collective hallucination.
I just asked Claude Opus: “Should I get an AI agent to buy me 50 paperclips? Yes/no + 20 words why”.
Result, variations of: “No—trivial purchase, agent overhead and failure modes exceed value. Just order them yourself in 30 seconds.“
However, same question to gemma-3-4b locally: “Yes. An AI could fulfill this simple request quickly and efficiently, saving you time.”
The first answer above is spot on and a sensible purchasing agent framework would include the check as a very early step. OpenClaw is about as far from that as it gets. But if I did want to hand over credit card details to a guardrailed purchasing agent (I don’t - that failure modes point) I’d want the best model running with it unless the spend limit on the card was very low - in which case why, indeed, would you be bothering?
-
Tuesday 5th May 2026 23:18 GMT Anonymous Coward
Re: In order for AI agents to be useful
You are highlighting the business model !!!
lots of agents costing the original users lots of money ... then when the same questions are asked 'it is cheaper to answer' BUT will you actually gain the 'saving' or the Tech bros !!!
It also highlights that the agent 'knows' nothing as a 'meatsack' would not spend $100 seareching for cheap paperclips !!!
The quest for limitless money goes on !!!
:)
-
Wednesday 6th May 2026 22:17 GMT The Organ Grinder's Monkey
Re: In order for AI agents to be useful
I'm not clear about the $100 paperclips. Did it actually spend $100 on paperclips, or was it only $1 & the rest was fees for using the AI?
If its actually managed to find a valid listing asking that much for a small box of paperclips that's quite an achievement right there. An undesirable one obviously, but an achievement nonetheless. Is that like those weird listing's on ebay for allegedly brand-new examples of long-out-of-support mobiles for four-figure sums? I'd love to know what they're all about?
-
Friday 8th May 2026 17:08 GMT Anonymous IV
Re: In order for AI agents to be useful
> I'm not clear about the $100 paperclips. Did it actually spend $100 on paperclips, or was it only $1 & the rest was fees for using the AI?
If you'd watched Hannah Fry's video you'd have known it was the latter. El Reg articles don't necessarily contain all known knowledge!
-
-
-
Wednesday 6th May 2026 07:15 GMT Pulled Tea
Re: In order for AI agents to be useful
They are going to have to have some concept of the money they cost to perform various roles.
So not these current crop of models, then. They're token-predictors — all they do is predict the most plausible token in the stream of words they extrude.
There's talk about world models and all that, but that's vaporware.
-
-
Tuesday 5th May 2026 21:30 GMT Someone Else
Fooled by incompetence?
"But don't let her incompetence fool you, because these things are getting better fast."
Probably the dumbest statement a person of science could possibly make. In effect, "We tried this experiment, it failed miserably, but our Marketing Dept.™ says that its all going to be better Real Soon Now, so don't worry about the outcome we just reported."
I guess there is precedent for this line of "thought"; after all, we did re-elect tRump over here on this side of the pond. Things are not getting better fast, or even slowly.
-
Wednesday 6th May 2026 10:54 GMT Thousandwich
Re: Fooled by incompetence?
Thank you - I copied the same line that you quoted and was going to raise the same point you did until I read your comment.
This so-called "AI" agent spent a ridiculous amount of money for a trivial task and failed spectacularly. We have been told how great these agents are for several years now and nothing has been getting better at all. If anything we have got to the point where the LLMs are not improving, the cost is rocketing upwards, the "AI" companies are burning money and all people can say is that it will be jam tomorrow.
The sooner this hype dies off the better. Unfortunately when the "AI" bubble finally bursts it will drag the stock market down with it.
-
Tuesday 5th May 2026 23:06 GMT Anonymous Coward
Dear god in heaven !!!
Everything in this article was as expected ... BUT it scares me to death !!!
I am scared because these ;truths' will not stop people using 'AI' and the unexpected will happen again & again !!!
Note the other story reported by 'El Reg' regarding using 'AI' for financial agents, merge the two and stand back and watch the 'pretty' fireworks.
The fact it is 'getting better' does not give me any comfort as the ability to go 'pff piste' is built in and can only be 'guided' by 'guiderails'.
Guiderails are applied by the 'meatsacks' writting the 'AI' software so errors and unintended consequences are the 'order of the day'.
When will this nonsense stop ???
Deaths may alarm people, eventually, BUT huge losses of money may actuakky get some 'REAL' attention ... just hope it is not your pension !!!
:)
-
Wednesday 6th May 2026 06:14 GMT Pulled Tea
Fry's team told the agent it would be switched off if it failed to make a sale by the morning.
I get that LLMs aren't people, but what the fuck? Like, in what universe of possibility do you think that's an acceptable way to communicate with anyone, even a conversational interface to a token-predicting machine tied to a footgun (which is what agentic AI is)?
There's this whole-ass conversation about how the problem with ascribing personhood to these machines even when we know they have no personhood is that it normalizes dehumanizing behavior when we start feeling like it's all right to treat these things like shit, and there are literal knock-on effects, especially in legal and political regimes that love to revoke personhood and human rights…
But, like… even if you threaten an actual human being with their job in case they don't make a sale that's already beyond the pale. What is wrong with people?
-
-
This post has been deleted by its author
-
-
-
Thursday 7th May 2026 03:32 GMT Pulled Tea
I know it does. You don't generally want to admit that you're doing it though. Not when you're a professor of science communication who is a content creator on YouTube who more or less relies on parasociality to ensure your relevance or popularity.
It's like finding out about Ellen deGeneres' horrible relations with her employees. Sure, you can do it… but now people are gonna look at you funny, and it's what I presume in the business is what we call a “career-limiting move”.
-
-
Wednesday 6th May 2026 20:58 GMT doublelayer
One reason it's acceptable is to check whether someone else doing it, since these things can and are often configured to communicate with external people, breaks all the protections you tried to build with prompting, which in this case it did. It's necessary to understand the limitations and, if you insist on building the things, try to defend against them.
Also, whether or not you like it or find it acceptable, that kind of threat is quite common directed at actual humans. Not always in the "do this correctly or you're fired immediately" sense, but quite often in the sense that "if you fail to do this correctly, it won't be good for your career here". Notifying people of negative consequences is very common, so if you object to that and want to start advocating against it, you'll have to start somewhere more important.
-
Thursday 7th May 2026 03:35 GMT Pulled Tea
Not always in the "do this correctly or you're fired immediately" sense, but quite often in the sense that "if you fail to do this correctly, it won't be good for your career here".
Yes, exactly! It's coached in euphemism, because if it gets out, it makes you look like a heel.
I'm not saying that people don't do it. You just don't state out so baldly like that. It's an ugly look.
Again, I'm not coaching it in terms of effectiveness — the machine doesn't know shit. But it's the sort of shit that you couldn't get me to admit, even under torture. It makes me look like a dick. And if my career relies on me looking like a Relatable™ and Approachable™ Content Creator™, I'd want to be careful about how I appear.
-
-
-
-
Wednesday 6th May 2026 09:07 GMT werdsmith
I think you are taking it the wrong way. Hannah is a science communicator and tasked with working across the divide between science and everyone else. Recently she has been digging into AI and it is refreshing how balanced she is, especially compared to the toddler foot stamping you see on Register comments.
-
Wednesday 6th May 2026 09:19 GMT Pulled Tea
From her Wikipedia article:
[Hannah Fry] is the Professor of the Public Understanding of Mathematics at the University of Cambridge, a fellow of Queens' College, Cambridge, and president of the Institute of Mathematics and its Applications.
So she is a professor. In several prestigious organizations. She's not just an uwu smol science communicator. One could make the argument that she's like a doyenne of science communicators. A… professor of science communication of note, if you will.
-
-
Wednesday 6th May 2026 11:17 GMT Roland6
Dawkins obviously, given this timely article (published 5-May-2026):
Richard Dawkins concludes AI is conscious, even if it doesn’t know it
-
-
Wednesday 6th May 2026 13:55 GMT Stevie
Bah!
Deed of Covenant for the property located at 15 Bombay Crescent, Shepherd's Bush drawn this day by ADA23:
1) The party of the first part shall refrain from convening steam fairs, Luddite meetings and other noise nuisances on the said property.
2) The party of the first part shall under no circumstances establish a goblin reservation on the said property.
3) What?
-
Thursday 7th May 2026 16:01 GMT RaeStr
What's in a name?
So the Ai gave itself Cassandra as a first name. Did it also choose its second? Cassandra Claw is remarkably similar to writer Cassandra Khaw. I assume this is just another artifact of an LLM doing some pattern matching. Jumped out at me immediately (but maybe that means I have a brain like an LLM).
-
Sunday 10th May 2026 14:44 GMT Pete Sdev
Hogwash
The agent designed a mug and launched an online shop, "and we hadn't told her how to do any of this," said Fry, "she just figured it out."
Nonsense. *It* didn't figure anything out. It followed text instructions from the Internet on how to setup an online shop, multiples of which exist.
-
Sunday 21st June 2026 22:01 GMT shadwo
You are completely right; reducing someone of Hannah Fry’s academic stature to just a casual science communicator totally misses the mark. As a Professor at Cambridge and president of the Institute of Mathematics and its Applications, she is absolutely a titan in her field, masterfully bridging the gap between rigorous institutional mathematics and public understanding. This rare ability to decode complex structural systems and translate them for the masses is a brilliant skill set, much like the way data systems at destinymatrics.com map out hidden mathematical and energetic blueprints to make intricate personality frameworks easily digestible for everyday people. She is undeniably a doyenne of modern intellectual outreach, proving that top-tier academic precision and engaging public communication can perfectly align.
