tech
"The Illusion of Thinking" - LLMs face "complete accuracy collapse" beyond certain complexities....
Apple’s Machine Learning Research group has published another damning piece of analysis of the capabilities of Large Reasoning Models (LRMs) and Large Language Models (LLMs). This important piece of work further exposes the limitations of LRMs and LLMs - both types of model experience “complete collapse” when dealigning with high-complexity tasks.
The paper is available here: https://machinelearning.apple.com/research/illusion-of-thinking
The key takeaway from this is that Generative and/or General AI are far far away from being safe and useful. More limited Machine Learning algorithms designed to analyse specific types of data are clearly useful. But we should be very wary of extrapolating from this to more “general” AI models.
AI, the Brogliarchs, and the Surveillance State
Excellent Daily Show discussion with Carole Cadwalladr about the dangers for AI, the Brogliarchs and the ‘techno-authoritarian surveillance state’….
You can read more of her work here:
Generative AI and LLM (large language models) are in serious trouble
There is growing and very tangible evidence that Generative AI and the Large Language Models (LLMs) that underpin them are in serious trouble. We are talking here about OpenAI’s ChatGPT and its underpinning models, Google’s Gemini, and other similar LLM tools.
Not only are they built on the theft of the underlying intellectual property on which they have been trained, but it is increasingly clear that they are generating vast numbers of errors, ‘hallucinations’ and frankly, bullshit.
Now the evidence is clear that the latest, more ‘advanced’ Generative AI models are increasingly prone to errors and bullshit. The newest OpenAI model has an error rate of 40-60% - a huge jump from the 14% in the first version of the model. No one - least of all their designers - actually understands why - but a leading theory is that the model’s ‘reasoning’ is leading to a recursive number of made up errors and moving towards ‘model collapse’.
Its definitely time for anyone interested in the truth, actual work and delivering value, to hit the hard pause on any use of Generative AI and LLMs until it becomes clear what is going on here.
To be clear, this isn’t to suggest that all forms of machine learning are prone to these kinds of errors. It seems that small, less generic and targeted models can be useful tools. But it does seem that the rush to create ‘general’ models is heading for a train wreck using current approaches and tools.
References and links:
Academic Article - ChatGPT is Bullshit (July 2024) https://link.springer.com/article/10.1007/s10676-024-09775-5
AI Hallucinations are getting worse - NewScientist (May 2025) https://www.newscientist.com/article/2479545-ai-hallucinations-are-getting-worse-and-theyre-here-to-stay/
AI Hallucinations worse than ever - Forbes (May 2025) Why AI ‘Hallucinations’ Are Worse Than Ever https://www.forbes.com/sites/conormurray/2025/05/06/why-ai-hallucinations-are-worse-than-ever/
AI is getting ‘more powerful’ but its hallucinations are getting worse - NYT (May 2025) A.I. Is Getting More Powerful, but Its Hallucinations Are Getting Worse A new wave of “reasoning” systems from companies like OpenAI is producing incorrect information more often. Even the companies don’t know why. https://www.nytimes.com/2025/05/05/technology/ai-hallucinations-chatgpt-google.html
AI model collapse - The Register (May 2025) Some signs of AI model collapse begin to reveal themselves Prediction: General-purpose AI could start getting worse https://www.theregister.com/2025/05/27/opinion_column_ai_model_collapse/
AI model collapse - BGR (May 2025) AI model collapse might make current hallucinations seem like a walk in the park bgr.com/tech/ai-m…
What is Model Collapse (Jan 2025) In this episode of the Charlotte Content Marketing Podcast, Andrew Rusnak discusses how AI model collapse threatens the integrity of data on the Internet. Learn how AI data is feeding upon itself and how you can take steps to protect your brand from harm through authentic content. www.charlottecontentmarketing.com/knowledge…
Generative AI is garbage 🗑️
The Chicago Sun-Times used AI to write a Summer Reading list - and it was full of garbage including recommending books that don’t exist …
www.thepopverse.com/literary-…
www.theatlantic.com/technolog…
Stop using Amazon Kindle (before it's too late)
From 26 February 2025, Amazon will stop you downloading your *own* Kindle eBooks. You will be forever stuck in Amazon’s grip.
If you’d like to escape, download all your Kindle ebooks now before it's too late:
https://www.theverge.com/news/612898/amazon-removing-kindle-book-download-transfer-usb
Remove the DRM using Calibre:
Ideally, start using a service that supports local bookstores:
- Libro (audiobooks): https://libro.fm/
- Bookshop.org: https://bookshop.org/ebooks
Or if you prefer another option:
- Apple Books: https://www.apple.com/apple-books/
- Kobo: https://www.kobo.com/nz/en
- Tom's Guide to DRM free ebooks: https://www.tomsguide.com/tablets/e-readers/no-kindle-no-problem-5-places-to-buy-drm-free-e-books
Posting on Bluesky
I’ve set up an account on bsky.app
I love micro.blog - but it’s small and paid.
I love that I can use my own domain (jeremybaker.nz) as my handle!
And lots of other cool stuff about Bluesky:
Alexa, Amazon and loss leading
So it seems that Alexa is making huge losses for Amazon.
It is not that making losses is unusual for Amazon; their entire business model is based on ‘loss leading’. The problem with Alexa appears to be that the ‘loss’ wasn’t ‘leading’ to anything…
Amazon Alexa is a colossal failure - on pace to lose 10 billion this year
This is so wonderful…
literature-clock.jenevoldsen.com
@Miraz - I think I might need to learn how to embed things!!
I have “defederated” (deleted) my cross-posting to Mastodon from Microblog - not because I’m not enjoying Mastodon (especially the paid hosting model at Cloudisland.nz - but because it is becoming clear that, for me, the two spaces will involve quite different types of posts.
Taking it easy on this first day of November 2022. I’m leaving Twitter, and focusing on Micro.blog (with an occasional side of Instagram). Beginning to figure out how to integrate with Mastodon NZ. #mbnov
I have a level of dependence on the new word for Microblogvember being posted before I go to sleep, but sometimes I miss it! Oh well :) #mbnov
So far, I am really, really enjoying macOS Big Sur. Installation went well, and my 2017 iMac is noticeably faster than it was before the upgrade. And the look and feel of the whole OS is much improved.
I guess I’ll give downloading and installing macOS Big Sur a go… or should I wait a few days? Hmm 🤔 #mbnov