Become a member.

Looking for better conservative conversation? Join Ricochet for just $5/month. Think of it as your neighborhood bar: familiar faces, lively debate, and room for a different point of view. Plus: listen to most of our podcasts ad-free.

How Do You Make an AI Hallucinate? Like This...

Locke On
@LockeOn
Join to bookmark20likesJoin to like18comments

Every year I try to take on some new tech, which can be anything from old news (bookbinding, one year) to trendy. This time I decided to fool around with the currently hot 'Large Language Models' (LLM), known in the media as just 'AI' (artificial intelligence). I'd had an involvement with very early neural nets and related math over 30 years ago, so seeing how those then very limited capabilities had turned into a cause célebre was interesting.

As an up-to-speed I ran through a Great Courses class on general machine learning (taught by a prof who was a grad student at Bellcore when I knew his boss). Then through the lecture notes and supporting files for a Stanford undergrad course specifically on LLMs. Among other things I learned terms like 'embedding' and 'hallucinations', spicier than the dry math from back in the day. Thus prepared with updated jargon and a list of players, it was time to get my hands a bit dirty.

Training an LLM from scratch takes a giant pile of data, a monster server farm or at least five figures to rent usage of the same. I was looking for something I could do on my rather aged, though once top of the line, Intel based Mac as a first try. I decided to attempt to replicate a project from the same 30+ years ago that created simulated tutor characters drawn from early American history. The application of LLMs to create non-player characters for fantasy, anime and SF role play games and MMOs means there are a number of platforms that might be suitable. I settled on one called Faraday.dev, which lets you download an app to your own machine, where you can securely run various LLMs to power characters already defined on their site, or built yourself - my interim goal.

I downloaded (from huggingface.co) a model called mistral.7b.synthia. Mistral is the company that built the LLM. 7b stands for 7 billion (gig), that's how many parameters there are in the 'deep' neural net or transformer that implements the LLM. That's a small model - the well known ones like GPT-4 or the now infamous Google Gemini have hundreds of billions, but require the big server just to chat with you. 'Synthia' is the name of a 'fine-tuned' version of mistral.7b, specifically designed for extended conversations.

'Fine-tuning' is tweaking an existing LLM to better handle a specialized vocabulary or type of dialog. Rather than an entire new model ($$$$$) you instead start with a working model, and partially retrain with a new sample of appropriate texts ($$$$). Or you can apply a technique called LoRA that even further reduces the amount of training needed ($$$). But it's even cheaper to find that someone's already done the tuning you require.

Anyone who's dealt with early American English sources knows that vocabulary, dialog and even syntax can vary from what can be scraped from the modern Internet. If I'm to build a credible historical character, I'll eventually need a model fine-tuned for the period of interest. How to find such a model? Well, maybe I should use AI itself, it's supposed to be helpful, and I can bootstrap...

I downloaded a generic help bot character, imaginatively called Faraday AI, as a front end to the Mistral 7B model, and away we went:

Faraday AI: Hello User. My name is Faraday AI, your AI Assistant. You can ask me anything and I'll try my best to provide you with a helpful and coherent answer. How can I help you today?
User: Has anyone done a large language model for historical dialects of english?
Faraday AI: Certainly! There are several large language models that have been developed for historical dialects of English. One of the most well-known is the Old English Language Model, which was created by researchers at the University of Oxford. This model is based on texts written in Old English, the language spoken in England in the early Middle Ages. Another example is the Early Modern English Language Model, which was created by researchers at the University of Cambridge. This model is based on texts written in Early Modern English, the language spoken in England in the late 1500s and early 1600s. Additionally, there are several other language models that have been developed for historical dialects of English, such as the Middle English Language Model, the Renaissance English Language Model, and the Restoration English Language Model. If you have any specific questions about these language models, I would be happy to provide you with more information.

Not quite what I was looking for, but promising!

User: Are they any models for colonial american english?
Faraday AI: Yes, there are several language models that have been developed for Colonial American English. One of the most well-known is the Colonial English Language Model, which was created by researchers at the University of Texas at Austin. This model is based on texts written in Colonial English, the language spoken in America during the colonial period. Another example is the Early American English Language Model, which was created by researchers at the University of Michigan. This model is based on texts written in Early American English, the language spoken in America in the late 1700s and early 1800s. Additionally, there are several other language models that have been developed for Colonial American English, such as the Revolutionary War English Language Model and the Early Republic English Language Model. If you have any specific questions about these language models, I would be happy to provide you with more information.

Jackpot, just what I need! There's just one small catch...


... and you may have guessed it from the title of the post: Not a single one of those projects actually exists.


On my first serious excursion, I'd hit the famous 'hallucination' problem. I might just call it 'over-extrapolation' but that's not so catchy. LLMs in their current state aren't exactly self-critical. If they don't have verifiable info to answer your question, well they'll just make up something that sounds reasonable. Just what was going on became more visible when I poked it a bit further:



User: Tell me more about the Revolutionary War English Language Model. Who developed it and what platform is it built on?
Faraday AI: The Revolutionary War English Language Model was developed by researchers at the University of Pennsylvania. This model is based on texts written during the Revolutionary War period in America. The model uses a combination of machine learning techniques, including deep learning, to analyze the language and structure of the texts. The model has been built on the TensorFlow platform, which is a popular open-source library for machine learning....




User: Who are the researchers behind it?
Faraday AI: The Revolutionary War English Language Model was developed by a team of researchers from the University of Pennsylvania. The primary researchers involved in the project were Professor James P. Allison, Professor William J. Turkel, and Professor Daniel H. Weiss. Professor Allison is an expert in digital humanities and has been involved in a number of projects related to historical language analysis. Professor Turkel is an expert in digital history and has been involved in a number of projects related to historical language analysis and digital archives. Professor Weiss is an expert in digital humanities and has been involved in a number of projects related to historical language analysis and digital archives.






Now TensorFlow is in fact a deep learning and LLM platform, and the three named men are professors. But Weiss is an art historian at Johns Hopkins, Allison is a Nobel laureate immunologist at University of Texas, and Turkel is at University of Western Ontario, with an eclectic agenda that at least includes computational history, but in his case he means history of computation - he recreates vacuum tube computers, among other things

In this case, Faraday/Mistral-7B has delivered a high tech version of Mad Libs. The syntax is impeccable, and it's figured out that computational linguistics, historical English and research projects are the topic, but from there things fall apart. Lacking specifics - and in this simple platform a way to retrieve them from elsewhere - it grabs a few bits of information that appear to at least have the right types (platform, universities, research professors) and regurgitates them uncritically.

This may have worked out better on one of those hundred billion parameter models that might have captured some relevant specifics during training, but maybe not: I haven't found any clear answers to the questions I posed in a fair amount of Web searching. There's a basic hole in the current state of the art, which lacks an ability to throw up its hands and say "Beats me!".

Since the AI topic has shown up on Rico of late, I thought its denizens might be interested in a real life adventure. I've since ignored, for the moment, the anachronistic dialect problem and plowed ahead into building some characters. I have a small digital Samuel Adams fulminating against King George's tyranny and predicting trouble if General Gage tries anything aggressive. The platform is working a lot better when dealing with background knowledge that is known in advance. Soon I will imitate a mischievous schoolboy user and see how far I can lead 'Sam' into hallucinations. Fun guaranteed, as Elon might say.

Recommended by Members

Discussion (18)

  • JoelB@JoelB

    I asked Leo what are the lyrics to "Louie, Louie", but I can't verify if its answer is correct or not.

    "Louie Louie" is a song made famous by the Kingsmen in 1963. The lyrics go like this: (Verse 1) "Louie, Louie, oh no, Me gotta go, me gotta go, Louie, Louie, oh baby, Me gotta go."

    (Chorus) "Then I'll catch him, I'll catch him, I'll catch him by the leg, I'll bring him, I'll bring him, I'll bring him to you."

    (Verse 2) "Louie, Louie, oh no, Me gotta go, me gotta go, Louie, Lougie, oh baby, Me gotta go, me gotta go, me gotta go." The song's lyrics are somewhat mumbled and hard to understand, which led to the FBI investigating the song in 1964 to determine if it contained any obscene content. The investigation concluded that the song was "unintelligible at the fast tempo" and that it was "unable to find any obscene or indecent words or phrases contained in the record."

  • Henry Racette@HenryRacette

    When it comes to AI, size matters.

    This is something I never would have anticipated. I assumed that neural networks would saturate when the data reached the kinds of levels we're seeing in the latest models. Saturation would lead to an inability to discriminate, to find nuance, and it would be obvious that we're talking to a dumb machine.

    Instead, and surprisingly to most people (and certainly to me), AI scales. The larger the matrix and the more training data provided, the better the results. I don't know if the scaling is linear, but it's certainly close.

    You're working with a network containing about 7,000,000,000 parameters. ChatGPT 4, which cost about $100M (est.) to train, has something like 100,000,000,000,000 parameters. (That's a hundred trillion.)

    It still makes things up, but it's a lot harder to lead astray.

  • Chris Williamson@ChrisWilliamson
    ReplyingtoHenry Racette

    Henry Racette (View Comment):

    When it comes to AI, size matters.

    This is something I never would have anticipated. I assumed that neural networks would saturate when the data reached the kinds of levels we’re seeing in the latest models. Saturation would lead to an inability to discriminate, to find nuance, and it would be obvious that we’re talking to a dumb machine.

    Instead, and surprisingly to most people (and certainly to me), AI scales. The larger the matrix and the more training data provided, the better the results. I don’t know if the scaling is linear, but it’s certainly close.

    You’re working with a network containing about 7,000,000,000 parameters. ChatGPT 4, which cost about $100M (est.) to train, has something like 100,000,000,000,000 parameters. (That’s a hundred trillion.)

    It still makes things up, but it’s a lot harder to lead astray.

    Unless I'm hallucinating, Henry Petroski pointed out that each new invention requires energy to create and operate it, adding to the whole energy requirement for civilization. My understanding is that large language models are sucking up a lot of heat to operate. I'm not sure how that energy requirement compares to, say, a trip from point A to point B by automobile, but it sounds as if our solar and wind farms won't cut it for this next level of inventiveness.

    BTW, I used chatGPT 3.5 for Table Topics at Toastmasters yesterday. In Table Topics the leader asks questions out of the blue and then one of the audience stands up to respond. I asked chatGPT for 10 questions on the subject "Experiences: Old and New", with each question having no more than 10 words. All of the questions were usable.

+13 more comments
Join the conversation

You're seeing five of 18 comments. Sign in, or join for just $5 per month, to read them all and add your own.