Since the last edition of That Space Cadet Glow the world of AI has completely exploded, ignited by the launch of ChatGPT and further fuelled by panic-released versions from competitors. Apart from just trying to keep up with it all (imagine your favourite band releasing an album every week for 3 months), I’ve held off commenting too much until things have settled down a bit. And whilst things are far-from-settled, let’s try and make some sense of it together, especially without all the (mostly deserved) hype.
The State of the GPT Nation
The Beginning of the Beginning
One thing is for certain - the world will look back at November 2022 as one of those pivotal moments, on a par with the release of the iPhone, where super-advanced technology suddenly becomes available and usable for all. The release of ChatGPT reached public consciousness so quickly that it even took its creators, OpenAI, by surprise. Suddenly everyone was able to create stories, generate code, make jokes, write strategies, etc., using one of the most advanced AI models ever built. Hype built upon hype as every media outlet regurgitated ChatGPT stories, usually starting with the line “ChatGPT wrote this article…”. It looked like we were finally witnessing the democratisation of AI.
Google panicked and released their own attempt, called Bard, much too soon, Microsoft (one of OpenAIs major investors) integrated ChatGPT into Bing, and every Big Tech firm promised their own version of a Large Language Model (or LLM, the current collective noun for these models). Meanwhile, OpenAI released GPT4, then a subscription model, then an API (to connect directly to the model, bypassing the browser), and then third-party plug-ins (to connect to pre-configured services such as Expedia). All of this within 6 months. It’s been breathless.
In the wider universe of ‘Generational AI’ other firms ramped up their image-creating models to compete with OpenAI’s DALL-E, including Stable Diffusion and MidJourney. Music has been generated by AI to sound like Drake, The Weeknd and Oasis. How we make, consume and appreciate creative content has changed forever.
The Beginning
But is any of this change actually any good? If you have used ChatGPT, or any of the other models, yourself you know that you will be genuinely amazed at its ability to create decent answers to widely-diverse questions. Long-standing readers of this newsletter will already have been familiar with generative AI (back in February 2019 I wrote about the launch of GPT2) but ChatGPT’s ability to follow a dialogue, and GPT4’s sheer creative capabilities are beyond anything we could imagine just a few months back.
During the first few months of 2023 though, most of the content that has been created by most of the people could best be described as ‘whimsical’. It was showing what the technology could do rather than actually doing something with the technology. The limitations, especially with the fact that ChatGPT was trained in September 2021 (so it has no knowledge of anything that happened after that date) and that it was so popular that it often crashed under the weight of demand, meant that it could not be considered for ‘serious’ enterprise use.
Those more recent developments (and, remember, we are talking in months here, not years) have shifted the dynamic a little. The paid subscription ($20 per month for guaranteed access), Bing’s ability to layer recent search results on top of GPT’s answers, and the direct access with the API give glimpses of how it could be used with more seriousness.
I’ve accessed all of these capabilities and have used them in my day-to-day work to great effect (IMHO!). I have used it to help me with the code to fine-tune an algorithm, to write layman summaries of technical terms, and to generate copyright-free images (such as the one in the masthead of this very email). I’ve used the API to get GPT to ask questions of, and summarise, specific business documents. (To see some of this in action, you can watch a recent webinar I did in collaboration with Everest Research). I’ve tested applications that have started to embed GPT capabilities, in some cases acting as additional functionality to the software or wrapping bells-and-whistles around a GPT core.
But all of these are just the tip of the iceberg of what will be possible and how these technologies will eventually be used. Over the next few months (again, months not years) we will transition into the next phase of generative AI which will signal serious, productive use in the enterprise.
The End of the Beginning
What do you do with an AI that can do anything? This is the rather bizarre question that many enterprises are currently asking themselves (and me). What do you do when the ‘killer app’ for the technology is ‘everything’? Here’s where I think enterprises need to focus when they are thinking about using generative AI.
Firstly, although every business is different, there will be some very general use cases that will apply to most of them. These include the writing of sales plans, generating research, creating marketing copy, formulating FAQs, writing computer code, etc. In each of these cases, the tool will be in the hands of individuals, and it will do something like 80% of the work for them - they will only need to tweak the content using their own experience and skills. Conversely, the AI may be used just for the initial 20% - generating ideas that can be used as inspiration for creative content, be those stories, characters, product designs, or images. Giving these people a $20 per month subscription to ChatGPT may be the best money you could spend right now (but bearing in mind all the potentially major caveats mentioned in the next section).
And, rather than people going and getting generative AI, for some it will come to them. Microsoft’s Copilot, although not formally launched at the time of writing, will be an integral part of Microsoft 365 (né Office), therefore it may actually be difficult to avoid. But from what we have been shown so far, it looks like it could be really useful, doing the job of creating ‘adequate’ content that will be enough for some and a jumping-off point for others. Furthermore, I think it will also provide an ‘air gap’ between content that has been (fully or partly) generated by AI and that which is genuinely creative, whether it be writing, music, images or even presentations. In other words, it will denote the difference between craft and art (which is one reason we should avoid calling the LLMs’ output ‘generative art’).
LLMs will also fundamentally change the way we interact with software. Gone will be drop-down lists and tickboxes, replaced by a natural language interface driven by a ChatGPT-like AI. (The soon-to-be-released ChatGPT interface for Microsoft’s Virtual Assistants is a good example of this). And let’s not limit ourselves to typing the queries in - there is no reason why we can’t use our voice to generate the words that the LLM will process. (This will starkly show how basic the current ‘voice assistants’ like Alexa actually are and will, I predict, precipitate their slow death).
To use this new way of interfacing with technology we will need to learn new techniques. Just as being able to do a good Google search is now a life skill, the same will happen to the prompts and questions we will be asking the LLMs. Already this has become a market in itself, akin to the rise of the SEO businesses, where people will train you to engineer the ‘best prompts ever’.
But I think the true, long-term value from Large Language Models will come once businesses start to point them at their own corpora of data. At the moment, the models, such as GPT4, have been trained on (pretty much) the whole internet, meaning they are generalists and only knowledgeable about publicly-available information (created before September 2021 in the case of ChatGPT). But imagine a GPT4 trained on all the data in your company, so you could ask it anything about anything to do with your business. This is actually the ‘killer app’.
Microsoft is just starting to roll out OpenAI tools as part of their enterprise cloud platform, Azure. Amazon Web Services recently announced ‘Bedrock’, a platform that gives developers access to a range of LLMs. Both of these (and, I am sure, subsequent offerings from other Big Tech companies) will allow enterprises to train the models on their own data, opening it up to a completely different way to interact with business information. And that means both employees and customers. Employees will be able to find out information simply by asking a natural language question into one single interface, rather than working out where the information is stored and then searching through nested folder after nested folder. Customers can ask questions about products or services, or the status of their orders all through one interface. Forget chatbots that have been configured for specific, limited purposes - in this new world, there will be one chatbot to rule them all. I should be clear here though - do not confuse LLMs with the ‘all-powerful brain’ trope beloved of sci-fi, such as JARVIS, Deep Thought or HAL (examples chosen deliberately): LLMs are not sentient in any way - they are some very clever maths running on very powerful computers that are able to predict what the next word will be in a passage of words to give an uncanny semblance of sentience. I repeat, they are not sentient.
A few years down the line, I imagine we will have our own personal LLM, trained on all of our data, wrapped up in everything else from the internet. We will be looking back at pre-LLM days in the same way that we now look at landline telephones.
But…there is always a but…and in the case of LLMs there are a whole barrel of buts…
The End?
I’ve deliberately avoided talking about the risks and issues with generative AI so far in order to keep the focus on how transformative the technology is. But there is a dark side to this whole LLM story, with genuine risks that threaten its own existence, that have the ability to cause harm to whole groups of people, and that potentially represent existential threats to society. Combined, these risks could outweigh all of the benefits I’ve described above. Let’s remind ourselves of what the biggest risks from LLMs are.
One of the biggest risks is the models’ tendency to make stuff up. These ‘hallucinations’, as they are whimsically called, are done with such seeming authority that they can easily be believed, especially if you are unaware of the issue and not an expert in the domain you are enquiring about. Our world, which is already suffering from deliberate misinformation, will soon be suffering from naive misinformation, which is arguably more dangerous as the perpetrators actually believe their statements to be true. You might think that this hallucination problem will eventually go away, but right now, none of the developers of LLMs even knows why or how their models hallucinate, let alone how to go about fixing it.
The democratisation of AI, putting it in the hands of ‘the general public’, was always going to cause problems. This naivety of the hallucinatory risk, combined with the seeming authority of the output, will mean that many people will become overly dependent on the models to do everything for them. People won’t question the outputs enough, and won’t take the time to alter, tweak or enhance them so that they are actually meaningful and, well, correct. The world will be awash with inaccurate and mediocre content all sourced from the same place.
Then, at some point, the models will get retrained and will start to ingest all that content that has been their own hallucinated output. Just like photocopying a photocopy over and over again, we will eventually lose any semblance of what was originally there (i.e. the truth).
And it’s not just factual correctness that we need to worry about. As the models have been trained on the most biased content available (the internet) then many of those historical biases have been baked in. OpenAI and the other model developers are desperately trying to block the bad content getting through, but we still see answers where the professor is assumed to be a man and the student a woman. Most worryingly, it is the Big Tech companies themselves who are deciding what is ethically correct. And these are the same Big Tech companies that have recently disbanded their ethics committees. Who watches the Watchmen?
Then there are the issues with the copyrighted content that was used to train the model (which, in many cases, equates to stealing the livelihoods of content creators) and the ease with which the output can be passed off as original content in academia (a weird sort of plagiarism where there is no human that is directly being plagiarised). We are already seeing some countries, such as Italy, banning the service altogether, whilst others, like the UK, are looking very closely at that copyright question. I can’t see outright bans being put in place permanently, but there may be some very tight regulation introduced (such as the EU’s upcoming AI Act) that severely limits how the models are trained and run. The honeymoon period we are in right now may turn quite acrimonious pretty soon.
Finally, we need to consider the sheer computing power and data that these models require to be trained and run. Not only is there a huge environmental impact from this, but it also centralises control of these systems in the hands of a very small number of profit-led companies.
And that’s just the risks we know about now. There are bound to be emergent risks that we simply haven’t encountered yet or haven’t been able to imagine. These are Donald Rumsfeld’s infamous ‘unknown unknowns’.
Worried yet?
Fortunately, there is plenty of discourse around the (known) risks. But even this has developed its own level of hype, verging on the comical. An open letter from a large number of renowned, not-so-renowned and made-up members of the AI community proposed a 6-month pause in the development of LLMs for fear they would develop into ‘powerful digital minds’ with ‘human-competitive intelligence’. Maybe it was just to get the conversation started, but the suggestion is completely impractical and ungovernable, and would simply allow those who have been left behind (China?) to catch up. Elon Musk (an early investor in OpenAI who then fell out with them) signed the letter and then, just a few weeks later, started to build his own LLM.
The genie is certainly out of the bottle, so rather than trying in vain to stuff it back in, even temporarily, we need to work out practical ways to control it without getting carried away with the alarmist (and incorrect) ‘God machine’ narrative. The best proposals I have seen were written by the authors of the seminal ‘Stochastic Parrots’ paper, including Timnit Gebru, who was fired from her ethics role at Google for speaking out about some of their practices. The focus of these recommendations is on the near-term, genuine risks that the models present rather than trying to prevent some ‘long-termist’ hypothetical risks. The core recommendation is for transparency - for when we are using these systems (through provenance declarations and watermarking, for example) and in how they were trained and built. That means the onus is very much on the models’ developers, who must involve those who are most impacted by its risks (the artists and creators who have had their IP stolen, the content moderators suffering from PTSD, etc) and independent academics and industry experts. The Big Tech developers can’t cede their responsibility by simply saying their technology should be regulated on a case-by-case basis when there are an almost infinite number of cases to consider.
As the paper says, “we do not agree that our role is to adjust to the priorities of a few privileged individuals and what they decide to build and proliferate. We should be building machines that work for us, instead of ‘adapting society to be machine-readable and writable… The actions and choices of corporations must be shaped by regulation which protects the rights and interests of people”.
It’s a fine line between making the most out of these transformative technologies whilst making sure they are safe and fair for everyone, especially when the benefits and the risks have the ability to impact so many people, groups and society both positively and negatively. What is certain is that the technologists will move faster than the regulators, therefore there is no time to lose in starting to discuss and develop those controls and regulations. But as consumers of the technology, let’s not get carried away with all the hype and believe everything the Big Tech firms are telling us. Push back, ask questions, call them out if necessary. The technology is indeed amazing, but let’s not run blindly into a world that, as a result, would become more unfair, more unequal and more dangerous.
The new Space Cadet Glow
Regular readers will have noticed the different format of this edition of That Space Cadet Glow. One reason is that I had a lot to get off my chest regarding ChatGPT (sorry!) so this has kinda dominated this issue (like it has most other discourse), but I promise it will be back to a more diverse mix of stuff in the next edition.
But another reason for the change is that I’ve also switched authoring platforms and now use Substack. It’s somewhat easier to use, and has better web capabilities, but it has also recently introduced ‘Notes’. You can think of these as mini-updates - more substantial than a tweet but less comprehensive than a full newsletter. I’m going to be working out the best way to use these, but I’d love you to subscribe to them as I roll them out. Think of them like Twitter but without the algorithm (or megalomaniac owner).
I’d love to hear what think you think about the new format, and any views you have on my ChatGPT rant.
Take care everyone,
Andrew
That Space Cadet Glow
Greenhouse Intelligence Ltd
11 Vernon Road, London, SW14 8NH