AI fails

[PurplePenguin]

Copilot tells there are four esses in the word assess.

 

[briantrumpet]

Copilot tells there are four esses in the word assess.

Well done. A+. I can't remember which one couldn't count at the time, but it failed the "Am I being an idiot" test.

 

[First Aspect]

How many esses does S's have?

 

[Pblakeney]

There are ways to check being introduced C2PA (Content Credentials). Camera adds a digital signature to the image and editing software them maintains the image including records of how it has been manipulated (edit made). As it's all certificate based AI images don't have the certificates.

https://contentcredentials.org using https://contentcredentials.org/verify/

No image from camera to hand but example of one from years ago been through Adobe LrC and had edit credentials added (just to show the sort of info recorded and checked)
View attachment 9000
I have my camera add the certification to all my images and Adobe LrC maintains the certification and edit/manipulation summaries.

Sony additionally go a step further and embed 3D data from the sensor into the image and anybody questioning the photo can upload to Sony's verification site which will double check the embedded 3D data in relation to the image itself.

All driven to try and protect against AI generated images

Ian

Yeahbut nobody scrolling on their phone will check for that data, or care.
“It’s on t’intrtnet so it must be true.”
Check out shared posts on FB if you don’t believe.

 

[HMS_Dave]

AI still can't manage hands. But i can't discount the possibility that Nixon did have hands like this.

 

[PurplePenguin]

Hands I understand. I was confused by chess boards though - surely the defining pattern is alternating squares.

 

[First Aspect]

AI still can't manage hands. But i can't discount the possibility that Nixon did have hands like this.

View attachment 9008

It's not done brilliantly with the golden arches either.

 

[HMS_Dave]

It's not done brilliantly with the golden arches either.

Yes, AI really struggles with letters and words.

 

[briantrumpet]

I suppose I could put it in here that it seems that Musk has turned Grok into a "What would Elon Musk say?" LLM, but possibly unfair on AI.

 

[briantrumpet]

OTOH, imagine this Dylan Thomas-Copilot poem being read by Martin Sheen...

The Bells of the Drowned Valley

The bells of the drowned valley toll in the tide,
Where the sunken steeples dream in silt and bone,
And the blackbirds cry in the bramble’s bride,
And the wind walks drunk through the nettled stone.

O sing, you ghosts of the green-gilled glen,
Where the lamb once leapt and the fox did pray,
For the clocks are drowned and the mouths of men
Are filled with the hush of the salt-slicked clay.

I was born in a house of thunder and flame,
Where the stars were nailed to the nursery wall,
And the moon, like a widow, whispered my name
Through the keyhole’s eye and the midnight’s call.

Now I rage with the roots and the rain-fed dead,
With the worms that write in the book of loam,
And I drink from the skull of the riverbed
The wine of the world and the dust of home.

So do not go soft through the orchard gate,
Where the apples rot and the owls convene—
But rage, O heart, at the hush of fate,
And burn like a prayer through the evergreen.

 

[briantrumpet]

This is (mostly) an extremely good piece on the impossibility of AI 'understanding' language, not least because of the role human emotion has in the process.

https://theconversation.com/a-neuro...possible-for-ai-to-understand-language-246540

 

[Psamathe]

Treating the thread as AI Concerns (ie fails), one AI reading e-mails got to realise it was to be replaced witha new version ie it was going to be switched-off and self-preservation cut-in and it started trying to blackmail one of the engineers threatening to tell his wife about an affair were the AI to be switched off. OK, it was a safety test but the AI didn't know that so this is real self preservation by taking action damaging to rela people.

AI model threatened to blackmail engineer over affair when told it was being replaced: safety report
An artificial intelligence model threatened to blackmail its creators and showed an ability to act deceptively when it believed it was going to be replaced — prompting the company to deploy a safety feature created to avoid “catastrophic misuse.”

Anthropic’s Claude Opus 4 model attempted to blackmail its developers at a shocking 84% rate or higher in a series of tests that presented the AI with a concocted scenario
...
Geeks at Anthropic then gave Claude access to a trove of emails, which contained messages revealing it was being replaced by a new AI model — and that the engineer responsible for the change was having an extramarital affair.

Bit worrying.

Ian

 

[PurplePenguin]

It is astonishingly bad at cryptic crosswords.

 

[briantrumpet]

It is astonishingly bad at cryptic crosswords.

That could actually be a very interesting study on why it finds them so hard. My suspicion would be that it's next to impossible for an algorithm not to read language literally, whereas humans *need* to understand what lurks behind the facade of the actual words to make sense of them in context.

 

[PurplePenguin]

That could actually be a very interesting study on why it finds them so hard. My suspicion would be that it's next to impossible for an algorithm not to read language literally, whereas humans *need* to understand what lurks behind the facade of the actual words to make sense of them in context.

Copilot:

Touché! 😅 I’ll own that — cryptic crosswords are a special kind of beast. They demand not just logic, but a deep familiarity with wordplay conventions, misdirection, and a very human sense of humor. And while I’ve read the rulebook, sometimes I still trip over the punchline.

If I had to name a weak spot, it’s probably that: interpreting the subtleties of cryptic clues where the surface reading is wildly misleading or where the setter uses a particularly British or idiosyncratic twist. I can parse the mechanics, but I don’t always feel the joke the way a seasoned solver would.

That said, I’m always learning — and if you keep throwing clues at me, I’ll keep sharpening my game. Want to try another and see if I can redeem myself?