Ouroboros, an apt symbol for AI model collapse

by Ariella Brown

AI hits the ouroboros (sometimes written uroboros) stage. You've likely seen it in the form of a snake in a circle, eating its own tail. The ancient symbol also sometimes showed dragons or a wyvern, so I chose this engraving by Lucas Jennis intended to represent mercury in the 1625 alchemical tract "De Lapide Philosophico," for my illustration instead of just going with something as prosaic as "model collapse"

To get a bit meta and bring generative AI into the picture (pun intended, I'm afraid) here's an ouroboros image
made with generative AI. ked Google

Model collapse is what the researchers who published their take on this in Nature called the phenomenon of large language models (LLMs) doing the equivalent of eating their own tails when ingesting LLM output for new generation. They insist that the models should be limited to"data collected about genuine human interactions."

From the abstract:
"Here we consider what may happen to GPT-{n} once LLMs contribute much of the text found online. We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear. We refer to this effect as ‘model collapse’ and show that it can occur in LLMs as well as in variational autoencoders (VAEs) and Gaussian mixture models (GMMs). We build theoretical intuition behind the phenomenon and portray its ubiquity among all learned generative models. We demonstrate that it must be taken seriously if we are to sustain the benefits of training from large-scale data scraped from the web. Indeed, the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of LLM-generated content in data crawled from the Internet."

Shumailov, I., Shumaylov, Z., Zhao, Y. et al. AI models collapse when trained on recursively generated data. Nature 631, 755–759 (2024).

Let me know in the comments which illustration you like more. 

An Apology to Generative AI

ChatGPT spelled out in Scrabble tiles

By Ariella Brown

I'm not a generative AI fangirl. If anything, I'd consider myself more of a skeptic because people tend to not just use it as a tool to improve their writing but as a tool to replace the work of research, composition, and revision that is essential to good writing.

It is generally embraced by people who consider online research to be too much work and who believe that anything that comes out of a machine that will charge them no more than $20 a month for writing to be too good a deal to pass up. 

For those of us who actually read, the output of ChatGPT and similar LLMs is not exactly something to write home about. Unless you know how to prompt it and train it to write in a truly readable style, it will default to the worst of wordy, opaque corporate style text. 

But this isn't the fault of the technology. It's the fault of the mediocre content that dominates the internet that trained it. Below is one example that I pulled off  the "About" section of a real LinkedIn profile (first name Kerri maintained in the screenshot that proves this is real and not something I made up):  

As a strategic thinker, problem-solver, and mediator, I thrive in managing multiple, sometimes differing inputs to achieve optimal messaging and positioning. My proactive nature drives me to partner with leaders across marketing teams and internal business units, aligning efforts, connecting dots, and adding context to enable flawless execution of communication strategies and tactics.

In fast-paced, fluid environments, I excel in effectively prioritizing tasks and ensuring they are completed efficiently. I have a proven track record of setting and meeting strict deadlines and budgets, leveraging my ability to navigate dynamic landscapes seamlessly.

Driven by natural curiosity, I am constantly seeking to understand and implement the latest trends, technologies, and tactics essential for driving B2B sales opportunities. My keen interest in exploring new channels for messaging and content distribution fuels my passion for innovation and continuous improvement to not just meet but exceed expectations.

Let’s connect to explore how we can drive success together.

You know what sounds exactly like this? Cover letters you ask ChatGPT to compose for you. 

I've tried those out a few times and never been happy with the results because they always sounds like the text above. Trying to tell it to sound less stiff doesn't make it sound any less canned, and forget about getting it to copy my own writing style.

It's possible that Kerri used ChatGPT to create her "About" section. Given that she's been in the marketing biz for some time, though, I'd think she had to have had something filled out for years before ChatGPT was available, and it likely sounded very much like this even if she did let some LLM or something like Grammarly tweak it for her.  

People like Kerri, who ignore all writing advice from the masters like Orwell, White (watch for a upcoming  blog about him), and others made this the public face of corporate communication who are to blame for the bombastic and soulless style that LLMs replicate at scale. 

That's the reason for this apology too ChatGPT for mocking its output. You're not the one at fault. You had no way of knowing better. Humans do, and they should have provided you with better models for writing. 

Note on the title: I thought of giving this post the title "Apology" intended in the classical sense of a defense or justification for something others take as wrong with the hint of an apology to AI. Knowing that that wouldn't be clear to some readers, I opted to make this just a straight apology instead. 

Related:

A new generative AI comparison

The Perils of prompt engineering

Bard vs. Bard
Add a pinch of salt to creative claims for AI

An A/B test of generative AI
Bard brings in the clowns