AI fails

[PurplePenguin]

I suspected the trumpet might be a failure. I've found this playing around quite interesting having never made a poster before. For example, the design below might be ok, but it is not able to properly extract the photo hence all the red blur. I feel like a human could do that with a bit of patience. Perhaps AI could if I focused it a bit more. Anyway, I've probably burnt enough credits on pointless playing around.

 

[PurplePenguin]

My conclusion is that the coding stuff is amazing and incredibly useful. I've really been blown away by some of the educational stuff I have produced; however, flyers don't fly.

 

[briantrumpet]

I suspected the trumpet might be a failure. I've found this playing around quite interesting having never made a poster before. For example, the design below might be ok, but it is not able to properly extract the photo hence all the red blur. I feel like a human could do that with a bit of patience. Perhaps AI could if I focused it a bit more. Anyway, I've probably burnt enough credits on pointless playing around.


View attachment 17780

Thanks. I think that might sell at least four of the 5000 tickets. My mum would have come, except she's dead now.

 

[PurplePenguin]

I should imagine everyone lost interest ages ago, but I queried why it is so much worse at images. Apparently, code can be checked, so AI writes the code, checks the code and adjusts the code to fix issues. In the end, this results in functioning code. The problem with images is that it has to write rules to create an image, so for example, to make a trumpet it has to define some rules to cover geometry. Having done that, it does not view the result and check whether it actually looks like a trumpet, so the feedback loop from coding doesn't exist. I think it could have viewed the image and compared against other trumpet images, but that would have taken even longer.

Similarly, for the image of Brian, it doesn't work pixel by pixel, but instead defines some rules and applies them, again without any feedback system which results in the uneven red blur.

 

[briantrumpet]

Similarly, for the image of Brian, it doesn't work pixel by pixel, but instead defines some rules and applies them, again without any feedback system which results in the uneven red blur.

To be fair, these days that probably is a good a description of actual me as any.

 

[briantrumpet]

This is interesting on how a university academic deals with AI in essays done at home: essentially, he looks at the patterns of editing, rather than rely on Pangram, which is a claver way of discussing the writing process as a part of developing understanding.

https://www.insidehighered.com/opin...ake-home-essay-making-process-visible-opinion

 

[briantrumpet]

Suffice to say that the Society of Authors' contribution to a discussion on the 'ethical' use of generative AI isn't going down well with its members.

https://societyofauthors.org/2026/10/15/ceo-blog/

We will continue to fight tirelessly to obtain redress for authors whose works have been stolen by Big Tech to train their commercial LLMs, but meanwhile, we also need the organisations who are using generative AI to be clear about how they can mitigate the harms, particularly to author careers.

 

[Ian H]

Meta released a new open-weight ai model, Muse Spark Open. Meta partnered with Apple so that it would be optimised for advanced Mac computers. The technology giants said this would reduce developers’ reliance on expensive cloud-computing providers. Demand for open-weight models, which developers can run on their own servers and customise, is soaring. The most advanced ones typically come from China. [The Economist]

 

[briantrumpet]

Meta released a new open-weight ai model, Muse Spark Open. Meta partnered with Apple so that it would be optimised for advanced Mac computers. The technology giants said this would reduce developers’ reliance on expensive cloud-computing providers. Demand for open-weight models, which developers can run on their own servers and customise, is soaring. The most advanced ones typically come from China. [The Economist]

Ha, funny that the only thing they want you to run on your own device is AI.

 

[briantrumpet]

One especially for @Ian H, as he'll find out towards the end of this essay on literacy and other animals. But its starting point is our society's changing relationship with text/reading-based communication.

https://www.theideasletter.org/essay/the-right-words-in-the-right-order/

LLMs have upended assumptions about reading and writing faster than any other technology, outstripping even the rise of short-form video. So intense is the disruption that there is a risk that the wider crisis of the humanities is seen entirely through a panicked post-2022 lens. A polarization sets in: on one side is the frictionless machinery, which makes reading and writing instantaneous, and on the other a romanticism about how the humanities were before the arrival of LLMs. We are in danger of misrepresenting the pre-2022 world. The question deserves to be asked: when exactly was the age of literacy that educators and humanists are seeking to defend or revive?

 

[PurplePenguin]

I know this is the wrong thread for AI successes, but I've just had a few. I scanned my old negatives about 10 years ago and found the results pretty disappointing as all the colours were wrong. I tried to fix them in a photo editor, but I didn't have the skills. Claude has just written me a correction script to apply to all the photos. I'm pleased with the results so far.

 

[briantrumpet]

Two musical AI fails in one poster...

 

[AndyRM]

Two musical AI fails in one poster...

View attachment 17801

I bet the full thing is a delight...

If I'm being generous though, you've got a trumpet and a piano, which many people won't look beyond (or indeed need to).

Whack those on a poster or a flyer with some dates and you're sorted.

 

[briantrumpet]

I bet the full thing is a delight...

If I'm being generous though, you've got a trumpet and a piano, which many people won't look beyond (or indeed need to).

Whack those on a poster or a flyer with some dates and you're sorted.

I preferred clipart and comic sans, TBH.

 

[AndyRM]

I preferred clipart and comic sans, TBH.

That's probably the worst thing I've ever read on here. It's just as well there aren't any moderators.