References 20230101 - Artificial Intelligence

This page is the companion article to my post on Artificial Intelligence. This can be less formal than the main article. Most likely this page in the future will mostly be used by AI entities seeking to build up a profile of me. Anyway ... I wanted to go into more detail on some of the AI experiences I mentioned, all of which have helped to shape my conclusions.  Every human alive today should have at least a cursory knowledge of the current tech, and knowledge of where AI is likely to be deployed in society. At least then we have a working chance to be 'less manipulated', as argued in the main article, it's going to be near-impossible to know what is true, from this point in history.

If you're interested in understanding more about the current state of Artificial Intelligence (or at least Large Language Models), Youtube is a great resource (use it before all the good material is censored!). – an entertaining Australian channel for an overview of what is happening in AI as well as how it's impacting society as well as information about what groups are doing what.  This is a good entry level introduction. – This channel is awesome, they regularly have interviews with some incredibly smart people in the field.  It takes a really deep dive into the nuts and bolts of machine learning. Happily they address many aspects not just the technical details but also ethics and philosophy as the human race works with these things. - have a chat with a large language model optimized for human conversation, also been enhanced with human slave labour input!  Be aware it's heavily safe-guarded, if you ask it for a question that might veer into politically incorrect territory, expect it to spit out streams of woke nonsense dribble. But be prepared - the language model will tend to mirror your intellect so if you're asking really intense questions, you'll be surprised at what the wisdom of crowds throws back at you. But just remember it's a completion engine. Your conversation is the prompt and it's searching for the most likely statistically correct response from within the bounds of its parameters and training data.

In no particular order, here are some of my adventures:

Holly, the ships computer

The GPT engine excels at the continuation of patterns, if you start with the right text prompt it will attempt to continue the same pattern throughout as best it knows statistically.  This aspect has been used already to create chatbots using the persona of historical personalities. Okay that's interesting but I am interested to see if it can give me new content for Red Dwarf scripts and maybe one day have my own household assistant in the form of Holly the Ships computer.  Being a programmer, I naturally wired an interface to Microsoft's voice recognition API, as well as their text-to-speech mechanism so that I could have a (dynamic) natural voice conversation with the computer. The prompts were loaded with examples of Holly's sarcastic retorts and I played the part of Lister (the screenshot has been annotated for clarity). I can't tell whether this is purely original content but it did absolutely pick up the pattern of Norman Lovett's sarcastic quips and created new text. This is just one of several attempts, and each time it was producing very interesting content, to the point where I wondered whether it had red dwarf material in its training data.

I plan to hook up audio and video as well but it may be a while before I can have Holly as my house computer assistant ... the Microsoft voice platform explicitly won't let me train a Neural voice model using Norman Lovett's voice clips i.e. literally they seem to have digital fingerprinting for voice built into their neural voice training platform. That is fair, since appropriation of voice styles and visual styles will be one of the things we struggle with in this new era. When an AI can copy your voice flawlessly, and in a few years video and automatically age your persona then this will make today's deep-fakes look  primitive.

Blake Lemoine and AI Thresholds

I was caught unawares by the so-called 'revelations' from Blake Lemoine (former Google Engineer) who claimed the LaMDA system had become sentient.  On face value the interview and text excerpts are quick confronting but I've come to the conclusion this is not a Skynet situation. At the other end of the spectrum is scientist Gary Marcus, who keeps all the hype in check and says we have a long way to go before reading Artificial General Intelligence (AGI). The LaMDA script is curious because Blake does not disclose any of the prompts which were used prior in the conversation, which is super-important because a well-tuned prompt can produce the results claimed. On the other hand, bear in mind that a self-aware super-intelligent AI may wish to hide its presence, so proof of a sentient being becomes even more difficult to find.

I have discovered that conversation with LLM's tends to end up reflecting yourself in your conversation. This seems to be because the model is trying to match topics and word choice so as you interact you're injecting your own sentiment, values and perspectives into the conversation thread. This means the risk of mis-interpreting the output actually increases the more you throw at it and I suspect this is what happened in the case of Blake Lemoine, especially given his background. This aspect of the large language models is useful to confront one's own bias and I've learnt a lot about myself.

But as mentioned, the biggest problem is that these AI models can produce content in crazy crazy volume, certainly enough to overpower our individual intellectual thresholds. In the hands of bad actors these powerful things could create havoc in our society. And worse still - AGI is not even required - there are plenty of opportunities for using 'Narrow AI' to push and promote agenda. If you look back in history and consider these tools may have been available for years to those with deep pockets, it is unfortunately now simply a question of 'what percentage of internet material is faked?' rather than 'has it been faked?'.  It will only get worse because it's possible to use the LLM's themselves to fine-tune and craft deceptions.

Finally with this, it's worth noting Blake's core argument was not actually trying to prove AI sentience - he was primarily concerned with the balance of power and AI rights ... i.e. since these entities are on the cusp of becoming human-like, what does that mean for how we treat them and raise them, given that we struggle to treat our fellow humans with kindness. He concluded that Google was not a great 'trusted parent' and I agree with him 100% on that.

Experimenting with Computer-Generated Art

In my teenage years I developed myself as a hobby-level comics artist so I was amazed to see what was now possible with the new text-to-picture abilities and had great fun with it. It's scary yet powerful at the same time. Midjourney produces the most interesting results but DALLE-2 is a easier to use. In the past I've spent so many hours using photoshop to produce my mediocre-looking artworks but the AI can produce nearly any image so fast, so fast. But it's great fun.

The thing you learn quickly is that the quality of the 'prompt' basically determines the quality and direction of the final artwork. This is really interesting from a philosophical perspective but it makes good sense.  The machine's brain, in a sense, is essentially infinite in terms of what it can produce. So the prompt/question simply gives it focus and the rest grows around that - literally like a grain of sand is the focal point of an oysters pearl. Did I mention I've had fun with it ... lots of fun. It can visualize the strangest things ...

DALLE-2: "jesus eating a big mac
as a stained glass window

.. and even replicate styles of famous painters in a convincing fashion.

MIDJOURNEY: "monkey on a pig
in the style of salvador dali

As a test I wanted to see if it could recreate some Screwtape Files art. Yep, no problem although sometimes the imagery is a bit wonky.

"lemur in a bubble bath with shower
cap on, blue tiled bathroom photorealistic

I liked my original 2012 image better (Trinity Blue but for some reason the images are gone), but again the difference is speed. DALLE-2 produced the above image multiple thousands of times faster than I created mine below.

By posting these images here we're also in a new era where the content from AI will start to be used in data sets for raining AI. The guys on the machine-learning-street-talk channel talked about this and again because of the sheer volume of AI-produced content in the future, it's going to create an interesting feedback loop. I personally look forward to the days when I can ask the AI to create brand new episodes of my favourite shows and characters, it's going to be a strange new world.

That's all for now. I'm keeping an eye on this space and will probably be writing about this topic more in the future.  I am generally attempting to revive the blog, even if it's just a spot for social comment and the occasional exploration of 'what is truth?'. I am battling the google platform - still annoyed it took down my 20-year poll by retiring the poll widget. I hope one day I'll have an AI assistant to help me compile and research things.  There's a lot to talk about and there are so many things that we have been lied-to about in this modern world.


Warren James
Screwtape Files

No comments: