OpenAI’s ChatGPT: Time To Take Stock Of Things (regardless who the CEO is)

By drawing mainly on the example of ChatGPT—the ‘Bitcoin’ of Large Language Models (LLMs)—this necessarily quirky opinion piece attempts to broach important macro issues pertaining to human language and to the originality of creative human thought.

Check all AI results for accuracy, fairness, and potential harm. Ultimately, it’s human oversight that safeguards responsible use.

Opinion Piece. Written by Dr Chris Alexander, Head of TELC

So what might the future of LLMs be?

The reason why this paper intentionally uses expressive 'outbursts' of written direct speech is to emphasise a key takeaway, which is that no amount of AI prompting can replicate the unpredictability and originality of creative-language use by humans. Moreover, the author even goes as far as to state that language, with all its linguistic intricacies, fundamentally is a subservient tool for human expression—or to put it more succinctly, language has been created by humans for humans. Anyway, I hope some readers do not accuse me of being anti-AI: that is not the impression I wanted to give at all in this rather unorthodox paper. I am an AI fan and religiously follow its developments—it will without doubt, at some point, have a useful place in all our lives. But currently, with all this round-the-clock news and with all this round-the-clock hype, one might find it more helpful to consider approaching this AI field with a strong sense of optimism tainted with a strong sense of pragmatic scepticism.

The Bitcoin price from its implementation release in 2009 to its all-time high in November 2021 went from about zero to over $67,000 in just under 13 years; needless to say, if you’d bought it at the start in large quantities—and resisted any temptation to sell during those frightening years of dramatic price fluctuations—you’d even, I suppose, possibly have been sitting on billions of dollars of returns!

"Cool!" he exclaimed.

Nevertheless, personally, I still can’t get my head around how something that was worth effectively nothing in 2009 could achieve an all-time high market cap of $1.27 trillion[i] or approximately 60% of $2.1 trillion [i], which is what all the property in Manhattan in 2023 is now ostensibly[ii] worth.

"Confused?"

Well, it appears that when certain technological and ideological factors coalesce at certain moments in time, gargantuan upheavals are possible.

In the case of Bitcoin, its technologically decentralized nature, which was then thought to offer ideological autonomy from traditional financial systems and governmental controls, was arguably a key consideration for its early investors; the price explosion was then driven to further frenzied levels by things such as speculation and FOMO (Fear of Missing Out), macroeconomic factors, regulatory clarity, market manipulation, and technological upgrade factors.

Interestingly, at the beginning of 2021, OpenAI was valued at $260 million[iii], and in 2023 the company was, according to The Wall Street Journal on 26 September 2023[iv], ‘talking to investors’ about a possible sale of existing shares at a Bitcoin-reminisce (over)valuation of between $80 billion to $90 billion!

So why is a 201-500-employee[v] ‘AI research and deployment company’ with a (very) lofty mission ‘to ensure that artificial general intelligence benefits all of humanity’ being valued at over 3.28 million Bitcoins (priced at the time of writing this paper)?

In short, in a world of 23400-and-growing[vi] cryptocurrencies, should we always assume anything is possible?

OpenAI’s income, in 2022 was claimed[vii] to have been ‘less than $10 million’, whereas its projected income for 2023 and 2024 is $200 million and $1 billion respectively—maybe it’s more now—who really knows…

Interestingly, it is held[viii] that ChatGPT (the ‘Bitcoin’ of LLM chatbots) managed to garner 100 million users in just 2 months of intense media hype, whereas it took Facebook, for instance, 4.5 years to do the same thing (more organically). Yet, approximately 65% of ChatGPT’s 100-million users are the more technologically savvy, and on the whole, less affluent, millennials, aged between 18 and 34. By way of comparison, only about 8% of these users are, speaking generally, in the more affluent and largely less technologically capable 55 or older age groups. Moreover, ChatGPT has a high bounce rate of 41.80% with the average user spending 4 minutes and 32 seconds during each visit[ix]. So, even though it is not entirely clear from the stats available how many millennial individual-subscription users or corporate-client millennial users are in practice using the paid GPT-4 version with all its affordances, one might wonder, nonetheless, how the above-mentioned income projections for OpenAI might play out in practice should interest wane among the capricious millennials in using ChatGPT bot and any of its developer-related apps.

"But could that indeed happen, could interest in using ChatGPT decline (significantly)?" he speculated whilst slurping his medium-sweet Cyprus coffee.

After all, they keep adding loads of new functions to it to make it supposedly better. I think therefore it’s crucial we take stock of where we are now, as we need to consider soberly how AI’s current capabilities and realistic utility might apply to us individually in our daily lives. By the same token, it is advisable we shouldn’t be quick to make any geeky assumptions about how massive the impact of Large Language Model (LLM) AI might be in the future based on hype about the technology: as a rule, new technological ‘things’ don’t always pan out the way you thought they would. There is a bigger can't-see-the-forest-for-the-trees picture to be seen and understood: for a technology to gain traction, it must be used in a way that aligns with the ideologies of its not-always-predictable humanware users. Consequently, huge fortunes have and will be lost by tech companies that arrogantly ignore, or simply don’t think enough about, the above bigger picture.

CNBC, by way of example, on 10 October[x] even ran an article asserting that ‘Generative AI has ultimately become ‘overhyped’ and smaller developers of the tech will face challenges as it becomes ‘too expensive’ to run ‘. And on the (side) topic of developer plugin over-kill, it should be noted that from about May 2023 to September 2023 the number of ChatGPT plugins apparently increased over 6-fold from 146 to 943 plugins[xi]. Eventually, gargantuan unmanageable numbers of plugins of varying quality, cost, utility, and reliability will more-than-likely overwhelm and frustrate users rather than impress them. Likewise, in Reuters on 11 September 2023[xii], it was stated that ChatGPT plugin ‘Developers whose plugins are in the top 30 or so “popular” category described an initial burst of hype, followed by a steep drop-off in interest’.

Bearing in mind, there are 3.553 million mobile apps available in the Google Play Store[xiii], and bearing in mind all these ChatGPT apps might be suggesting OpenAI wants ChatGPT to become the ‘go-to’ AI platform, should we be expecting a ChatGPT app store pretty soon? If so, will they fix their UX and UI asap because they leave a lot to be desired.

"Maybe GPT-4V(ision) with the right prompt can help?" he suggested tentatively, having already tried doing that.

Anyway, how does having thousands of contrasting and at times, even ‘dubious’ plugins only on its paid GPT-4 service help OpenAI address its core values, which are, as stated earlier, centred around a mission to ensure that artificial general intelligence (AGI) benefits all of humanity?

"And on an important side note, do they know what they genuinely mean by the term AGI?" he asked cynically.

Furthermore, and appertaining to AGI, how on earth can you compare a silicon-based GPT-4 LLM, with its 1.7 trillion (or so, I’m told[xiv]) parameters, to the incredibly complex ‘battle-tested’, carbon-based, and above all, cunning and ruthless (at times) human, who has long figured out how to survive and flourish against any odds? Also, humans, I understand, are biologically astounding with their thousand-trillion-trillion-molecule and thirty-seven-trillion-cell bodies, and with their low-energy 10-to-20-watt ‘carbon-footprint’ brains containing about a billion neurons and 100 billion nerve cells.

What’s more, there is another disconcerting issue with the output from GPT LLMs, ‘stuff’ is made up more often than one might like to think. And, albeit there is a ‘reassuring’ warning sign about ChatGPT possibly producing ‘inaccurate information about people, places, or facts’, and albeit we are supposed to be ‘understanding’ about its so-called hallucinating or confabulating because it is undergoing ‘training’, this is not what I would expect from a company they’re trying to value at about $80-90 billion.

"Have you tried chain of thought and with internal error trapping in the sequence?" he enquired geekily.

'"Yes, and it doesn't help much,'" he replied.

"Well, maybe you're not trying hard enough," he rebuked sharply.

But then again, let’s not assume hallucinating or confabulating is something specific just to ChatGPT, this issue debatably applies to all LLMs. Anthropic's LLM, which powers Claude2 chatbot, or PaLM 2 LLM, which at present powers Bard are both ‘pretty good’ at it. Bard for example, straight up states at the top of every chat ‘I have limitations and won’t always get it right, but your feedback will help me improve’. Bing Chat tries somehow ‘to explain away its hallucinations by asserting in a somewhat woke manner: ‘Let’s learn together. Bing is powered by AI that can understand and generate text and images, so surprises and mistakes are possible. Make sure to check the facts, and share feedback so we can learn and improve!’. A logical follow-up question nevertheless at this point for the averagely-confused ‘punter’ might be—and apologies for the digression—why or how does the act of an AI understanding and generating text or images, necessarily equate to ‘surprises and mistakes are possible’?

And then there are also those tiresome LLM apologies for making mistakes in output discussions—and then there are also those bizarre apologies for when the LLM appears to be ‘knowingly’ hallucinating (see examples below).

Here’s one example of where GPT-4 had been told (by the human user) that it had made factual errors about a famous 20th-Century concert pianist called Colin Horsley.

The above response seems to me staggeringly irresponsible, I mean, how is providing made-up ‘general information’ based on the pattern of works often recorded by classical pianists of his generation and stature’ at all useful to anyone?

Here is another example from GPT-4; in this case the human user stated that GPT-4’s output was not correct and then supplied the asserted correct information. See GPT-4’s response below:

This, parenthetically, is certainly quite a whacky response when you think about it, as it suggests GPT-4 just wasn’t ‘trying hard enough’ to start with but had known the correct facts after all.

Again, this is not what I would expect from a would-be 80-to-90-billion-dollar company.

By the way, Bard chatbot, which is currently based on PaLM2 LLM, offers ‘three factual’ draft versions of the prompt while expecting the (human) user to be able to decide which of the drafts ‘is most accurate’.

"Is this some kind of time-consuming leader-board game?" he asked himself angrily.

Below is Bard’s rather unhelpful explanation of what it expects its human user to be doing with its draft outputs (i.e. 'if they offer different facts')—though, I guess passing the buck over to the unwitting human might be a good way of getting out of being accused of hallucinating.

The winner’s prize for apologising for making mistakes, by the by, goes, in my opinion, to Claude2 for its whinging apologies for its constant hallucinations, which it seemed to know it was doing in the first place. Here it has been asked why it keeps making factual mistakes about that famous 20th-century concert pianist called Colin Horsley, mentioned earlier:

(Below) Claude2 apologising incessantly and pitifully for its constant hallucinations.

Amazon thankfully has recently announced a multi-billion-dollar strategic investment in Anthropic,[xv] so, let’s hope someone figures out how to fix this. Google also stated, and somewhat strangely considering it will, so I've heard, be launching Gemini LLM soon, it wanted to 'help out', when it recently announced it would be investing in OpenAI's rival Anthropic [xvi].

"But surely OpenAI and the rest of them will sooner or later figure out how to stop all this hallucination nonsense? Maybe those new neural-network emerging technologies [xvii] they’re starting to hype now will help?" he mused disconcertedly.

"Or, are all these hallucinations something more serious and fundamental, you know, like a flaw in the way LLMs were designed and programmed?" Then he suddenly went silent for a moment pondering the possible implications.

A lot of damage, notwithstanding, has already been done, and please note, users at large will only have so much patience before they begin to lose complete confidence and trust in LLM output.

Then there is OpenAI’s multimodal ‘mad gold rush’ thing to bring in all those new features as quickly as possible to its paid version GPT-4. DALL-E3 text to generative is pretty cool, but how many of those quick-to-bore millennials will get addicted to it? Or, how many multimedia experts would even be bothered to use it for meaningful work arguing if you know what you want to do, it’s quicker to just ‘do it’ rather than ‘trying’ to get there by possibly exhaustively prompting and mostly hoping for ‘the best’. ‘Browse with Bing’ on GPT-4 is certainly not a deal breaker, Google search engine is plainly much better, and over 91% of users in 2023 agree and prefer to use it[xviii] (Bing has a paltry 3% of the market share for search engines in 2022-2023); moreover, and more revealingly, adding GPT-4-powered ‘Chat’ to Microsoft Bing hasn’t evidently had much or any effect on Bing’s search engine popularity in 2023. I wonder why….?

Another very recent thing is the ability to upload files to ChatGPT and get, for example, summaries.

"But why on earth would I trust using it for something really important? I mean, how accurate are its summaries anyway? Has anyone actually done a proper study on that?" he remarked offhandedly.

However, OpenAI’s addition GPT-4V(ision) to analyse images has many use cases[xix] (including obviously cheating[xx]), though it should be understood GPT-4V(ision) also hallucinates from time to time. Additionally, OpenAI recently announced that ChatGPT ‘can now see, hear, and speak’[xxi] (NB the latter hearing and speaking currently being done using its new phone App). However, in spite of the ‘ground-breaking’ affordances of being more ‘multimodal’, this all feels very ‘rushed’, and ‘haphazard’, and ‘lacking’.

"Something is wrong. And I’m not talking about ChatGPT’s, at times, annoying downtime," he thought to himself while munching into his last miniature Marks and Spencer Jaffa Cake.

"No, something is actually wrong with all these LLM (or Large Multimodal Model—LMM[xxii]) AI platforms," he said pensively.

"What could it be?"

"Well, it’s an ideological quagmire," he remarked smugly.

One part relates to the ‘Where is all this AI stuff going to take us?’, and ‘is all this AI stuff really going to be beneficial for all of mankind?’. Terminator, transhumanist, and next-evolutionary-link jokes aside, I think it is obvious that the mainstream perceptions of where this is all going are not at all positive, ‘fear of the unknown’, and ‘increasing trepidation’ might seem more apt. And, maybe the soundbite idea of (apparently smarter) AIs ‘phasing certain jobs out’ thereby ‘freeing up time’ for the (apparently less smart) humans to do other more ‘creative and productive’ things, really doesn’t sound that appealing too.

"Clearly. they've been bigtime fear-porning it, but then again, that ironically might be really good for AI company valuations, which could explain, I guess, why there's been so much fear-porning to start with," he quipped.

For instance, would AI in reality lead to 300 million people losing their jobs[xxiii] in the United States and Europe? Or, would the CEO of OpenAI, Sam Altman’s prediction (2023)[xxiv] truly come to pass that Universal Basic Income (UBI) could play a ‘vital role in reducing income inequality’ as a result of the ‘rise of AI and automation’?

From a technical perspective however, I feel OpenAI might be running ahead of itself with a lot of development being implemented without ironing out problem issues and without carefully considering the true utility of such development to its principle more ‘fickle’ millennial users. Equally, OpenAI’s ‘virtuous and unrealistic’ soundbite’ mission to benefit all of humanity by using an ill-defined and currently hypothetical (possibly pipedream) AGI doesn’t appear to align with its seemingly haphazard technical development.

"But I can't wait to try out xAI's new AI 'Grok' [xxv].....Musk & company seem to have an interesting approach to this benefiting humanity thing that makes sense to me, well let's see," he thought trying to be hopeful.

From the more entangled ideological perspective, despite the surge in investment in AI, there are already many early warning signs that could be indicating a more underlying rejective human user re-evaluation of AI. To name a few: (i) consider the growing calls to regulate AI or the rising number of lawsuits from artists and writers, (ii) consider the increasingly frequent negative comments on social media pertaining to the ‘garbage in—garbage out’ issue of AI factual inaccuracies or biases, (iii) consider the growing preference for real humans over humanlike AI avatars or bots.

However, there is also something else that might be ideologically wrong: the way ChatGPT and other AIs use language itself.

"Language you say?"

"I understand that even in creative mode it's also good at writing....like an AI, that is," he responded sarcastically.

Naturally, if you’re not good at writing in general, ChatGPT and other platforms of that ilk can be quite ‘handy’—you know, they can write you basic emails, do parts of your self-evaluation yearly reports, and other humdrum stuff like that. But we’re ‘a ‘ways off’ them mimicking in writing the profundity of creative human thought with all its asserted ‘higher human burstiness and higher human perplexity’.

"And what if language is actually profoundly human, created by humans just for humans?" he asked with profundity.

"Well, it arguably is, and that is my point." He then crunched unhappily into his low-calorie organic oat bar lunch replacement.

AI companies might therefore want to be thinking about a philosophical maxim written by an obscure and long-gone Greek-Cypriot poet and philosopher called Erricos Constantinides What imparts meaning to words is the workings of nature, and the feelings and actions of men[xxvi] (1979).

"And, if language is subservient to being human, with its baggage of complexities, does that have any implications for ChatGPT and AI in general?" he remarked rhetorically.

So, we may in fact be kidding ourselves if we think smartish databases with no underlying ephemeral human consciousness will ever be able to ‘mimic their way’ to wide-scale acceptance.

"They most likely won’t." he concluded flippantly.

"Yeah, but ChatGPT, it can still make you a fortune, can't it?" He yawned out loud, and then wondered whether a medium-sized portion of cheese macaroni for supper would be breaking the rules.