11 April 2023
I think Large Language Models (LLMs) like ChatGPT hold a lot of promise, and if they do very well, they’ll find their way into some of our apps and create a handful of new services we never knew we needed and some we did. They can reduce a lot of drudgery and open up a few more types of work to more people. On balance they can create a lot of value.
History is however really against my modestly hoped-for scenario. There are lots of factors counting against it, with the cards stacked in favour of ChatGPT (and the like) fizzling out to something way less than its potential .
By now you may have read a lot of breathless hype with the usual, and very predictable, outcomes envisioned of tech utopias and dystopias about to befall us, as well as an urgency to act now in some fear-driven way that closely resembles a distributed boiler-room scam.
I don’t think ChatGPT will reach anywhere near the revolutionary outcomes that either the Utopians or Dystopians predict and I’ll set out my reasons why I think so. When I talk of “revolutionary outcomes” I’m referring to real world differences we see walking around in the street. The iPhone, email and the Internet being examples of successful revolutionary technology.
As you’re reading this, remember I am talking about the unlikely outcome of widespread adoption, upending of many industries, destroyer of jobs, magic-like powers bestowed on users in all sorts of fields, class-warfare-inducing-productivity-gains type of revolutionary change being claimed in either direction. I accept that some of these effects might be profoundly felt in some very specific domains like photo editing for example, where in 2 years time a practitioner will look back and wonder how on earth they managed to do any work without it and they’re still getting used to the breakneck speed at which work happens because their output is 50% faster. But these niche changes are what I am saying will happen at best.
In the rest of this article I’ll be covering a lot of factors that I believe will mitigate the revolutionist’s visions, but if you don’t have time to read it just go with my main takeways – the world is complex and fragmented on a social and infrastructural level. All actors have agency to alter the terms of engagement. ChatGPT got a head start by bursting out of the lab and catching us unaware but like all lab experiments it’s now interacting with the real world.
Big disclaimer, I have not thoroughly researched every single thing here. I am not an investigative journalist and certainly not an AI expert (you can follow Timnit Gebru, who is one 🙂 ). I am likely misrepresenting some important elements through ignorance. I am however not a total non-expert and I read, see and hear a lot of what’s going on. These are just my thoughts put to paper.
In these next sections I’ll set out all the factors that mitigate the Utopian and Dystopian visions. There’s really no order or hierarchy – all factors have equal weighting in my view. Feel free to jump to what interests you, or read through it all.
My historical perspective
If you’ll indulge me a moment I’ll take a trip down memory lane to give you context of my perspective.
I was born in the early 1980s. I experienced (and vividly recall) a world of:
- Television:
- Then: 5 channels operating just a portion of the day. Kids shows only showing once per week.
- Now: Multiple on-demand streaming services and access to billions of hours of content on Youtube (still can’t find anything to watch by the way)
- Telephone:
- Then: Single household rotary phone, kids rushing to be the first to answer it. Phone number and address publicly listed in a book distributed to all households in the town. Keep calls short as possible to reduce spend. NEVER make an international call – hugely expensive. Very rude to contact someone after 8pm.
- Now: Never answer the phone unless its family or friends, otherwise likely to be bad news. Contact people globally for basically no cost. Contact at any time of the day via text etc – the other side will manage their notifications. Video calling is ubiquitous, especially in work settings. Devices now have a gazillion features and uses, the “phone” feature actually being pretty annoying mostly.
- Mail:
- Then: Penpals, letters from friends and family and a few bills. Postage stamps, the post office and your personal mailbox were a prominent feature of life. I even had a stamp collection. Handwriting was a thing. Send maybe 10 letters a year.
- Now: Email has revolutionised written communication, and alongside it text messaging.
- Internet search: I’ll limit this to my first use of the first incarnation of Google search. Compared to its predecessors (Netscape, Ask Jeeves etc), it was a radically different experience to internet search and anyone I knew used it exclusively after they used it the first time.
I’ll stop there because you get the idea. Huge changes have happened even in my time. They inform my worldview that huge changes are indeed possible, have happened and have had big net positives on the economy and society – and that’s not counting the countless examples I’m leaving out that happened prior-to and during my short existence.
History has not been kind to Hypes
The World-changers
Let’s just look at recent examples of hyped technologies fizzling out. Not saying they’re all dead, but they are well below where revolutionists said they would be.
I’ll have to add a lot of caveats here so I am not misunderstood! I accept that I am speaking with a lot of ignorance but appreciate I can’t speak with expert authority on all subjects. Nevertheless I have valid arguments to make.
- Blockchain: This has settled into a couple of useful use cases. In some cases super handy, in some marginally so. I haven’t seen everything under the sun ending up on a blockchain though. Deeds offices haven’t disappeared overnight, no serious contenders for distributed finance and the thousands of imagined use cases haven’t appeared. Not a single real world interaction 99% of us have an interface with blockchain at all and it’s had plenty of time, energy and money thrown at it to make it happen. Note: In my work I have recommended both for and against blockchain use and I stand by each case’s specific recommendation :).
- Crypto-currencies: No real-world use cases beyond speculative trading. Maybe there are a few but I haven’t heard of them. The only news I hear now is of it facilitating financial fraud in some way. Note I am not saying that this is all it’s good for, but I think the hype around crypto-currency is precisely what snuffed out the chance for it to find its value-adding real-world applications. Maybe there were a few but too many had the course of their development changed.
- NFTs: Non-fungible tokens. Remember those? Not years ago, like 6 months ago it was all the rage, equalling the current hype around ChatGPT. From my perspective, it’s completely collapsed. Never had a good use either. I appreciate that many artists loved it and it was a different (and hopeful) business model to a viable living, but I can’t help but think that the art and creative world was a convenient and unwitting pawn in the whole scheme.
Some people may recognize the above as being technologically interlinked and each successive one being a derivative of the other. Maybe that’s why they got progressively less useful?
Another game changer was supposed to be Augmented Reality and Virtual Reality. I won’t go into all the promises but suffice to say there was also a lot of hype and these have now seemed to find their place in a couple of apps and for some gaming.
I’m eager to hear of personal experiences with use cases I’m not aware of, either successful or unsuccessful ones.
The job-robbers in the tech world
Some developments in the tech world have had similar hype as the World Changers, but I think it’s worth illustrating those here as most people reading this may come from the tech/data world.
There’s been a few developments similar to ChatGPT in that professionals in the field are worried about the impact on their jobs. Here the innovations haven’t fizzled out, but show that they just become a more mundane part of regular workflows. At best they take drudgery out of work and improve productivity, at worst they are ignored outside of people choosing to use it as a solution to a specific problem.
If you are not “in tech” you may be surprised that those who are regularly fear the impact of tech development on their own employability. They feel it very strongly and probably more so than those not in tech because they are neck-deep in it all day. No person can survive the mental onslaught of job-robbing fear-mongering that’s sprouted in a constant stream of news, readily accessible and in some cases unavoidable.
In these cases, the development has actual real world use, and it’s why it has this heightened fear of taking people’s jobs that are related in some way.
- No code development: Be able to build apps with no coding experience? Sounds amazing – and it is, I’ve done it! Software developers were worried that this would spell the end of software development. This has been around for many years and I would argue that software development has at worst seen zero impact on their employability. What it has done is spread app development capabilities to those areas where getting custom development work was not financially viable. It’s plausible that it has built many viable business models that then inevitably go on to require custom development work.
- Auto analyses, auto Machine Learning: Various flavours of this have been around for a long time and many data analysts and scientists were worried (as they are with ChatGPT now) that their jobs are at risk. Tableau (and I think PowerBI) have had these features where text can be automatically generated to highlight the trends in the data. Excel has had similar “auto analysis” features for some time. A few years ago, there was a HUGE surge in interest and hype around automated machine learning services and many a series-A funding round for auto-ML firms. Data Scientists were worried this spelled the end for the need for their work, but again this has had (I would argue) zero impact on their employability.
- Github co-pilot: Just one year ago or so, Co-pilot was released and is like a mini ChatGPT but for generating code. Minds were blown, mother’s wept, men hugged, people searched for “non-tech jobs for software developers”. But it’s kind of moving along at a more reasonable pace now? I don’t know but I’m just seeing very little chatter about it lately. It’s a bit enveloped/subsumed by ChatGPT but I see it as a microcosm of what ChatGPT is facing in the wider world, but with a very forward,embracing and forgiving user base – all the cards are stacked in its favour in this case and it’s just become a bit “meh”?.
Similarly, ChatGPT is raising the same concerns about job viability. Some might be valid. But as the recent 2022/2023 tech layoffs have shown, the most material threat to your job is a company’s lower-than-projected quarterly earnings. Not your project, not the value you add, not your technical competence, not your lack of zen-master soft-skills, and certainly not a plugin.
Why I think “this is different”and why I hope it works out
This article might come across as deep pessimism, and maybe I am pessimistic in some way. But I do really, really, REALLY want ChatGPT and LLMs to succeed.
I’m writing this article because I hope it helps us temper the hype and claim our own agency to shape the outcome. We need to spot the opportunities and work needed to realise its full value. I believe it has the potential to be a positive gain for society even within my modest vision for it.
To put my earlier-stated vision in different terms, the value gain won’t be from some massive upend and disruption, it will come from hundreds of thousands, or millions, of small, lower-level gains that on aggregate are huge collective gain for us.
In what ways do I think this is different from similarly hyped tech that has failed to live up to that hype?
For starters, I tried out ChatGPT myself and had a surprised “huh, that’s kind-of useful” moment. Not just as something cool, but a spark of “useful”. It very much mirrored the first time I used Google search when it came out. It kind of felt like, “Yeah, I can tell it’s early but my expectation of what is possible has changed”. I can envision some solid use cases for it and am excited to experiment with it and if all goes well, really level up some parts of my work.
Secondly, many of my peers have largely had the same reaction to ChatGPT – a surprise that it seems at least moderately useful and a willingness to explore and discuss it without it feeling like a waste of time.
My stock standard response at the moment if I’m asked an opinion is “It seems good and I hope we figure out how to make it useful”.
What is “good” is relative
One of the first things that strikes people trying out ChatGPT is it aggregates answers way better than Google search currently does. Even for this article, I was trying to search for “kids animated movies where the villain owns a tech company” and Google just totally choked on that request. It was serving up only tangentially related answers, like “Top 10 movie villains of all time” – and this is from me already having, I don’t know, 15 years of Google-fu experience and google collecting every bit of data it has on me for years. When I saw the first page of results, my experience let me know I must just give up the search. When I tried ChatGPT, it gave me what I wanted and with a few additional prompts I got my bulleted list that I then went and double checked each entry (because for the time being ChatGPT won’t, or can’t, tell you whether an answer is made up or not).
But Google was not always like this mind you. For many requests it does serve me the right answer with the bonus that I can immediately vet the source because it points me to it, but as someone luckily enough to experience Google search since the beginning, I can tell you it wasn’t always like it is now. It’s gotten worse over time and I won’t hash out the examples, but fellow users will remember.
What I’m getting at is that some of what makes ChatGPT really cool and useful isn’t exactly novel, other tools and software that used to be cool and useful got worse over time and gave up some ground. The reasons for this degraded service, whatever they are, might well be around for ChatGPT too, it may just take 2, or 5, or 10 years to fully manifest.
Update: check out this great article on what degrades platforms.
Tech genius myth is over
I’m starting to sound like an old curmudgeon (lol…ahem..), but I remember when the tech sector had a lot of goodwill from society. The growth of the internet was exploding and new apps and websites were growing. Time consuming, frustrating manual processes were quickly gobbled and replaced by digital and automated ones, much to the delight of all involved. It seemed with just a bit of tech knowhow and some genius sprinkled in there was nothing the smart people could not do. They seemed like the geeks previously ridiculed in society who had good intentions and would keep the interests of their fellow humans at heart as they battled the evil corporations and the old ways of doing things. Old ways that were stagnant, bureaucratic and kept the establishment in their cushy positions. Tales of underdog stories abounded and who doesn’t love a hero’s journey with a little bit of a brilliant-idea lottery winnings thrown in?
But times have a-changed. The tech sector is now the controlling establishment, no longer an underdog. I can’t count how many times I’ve seen a tech CEO dragged in front of congress (or whatever it’s called, I’m not familiar with the US legal system), saying they’re sorry and it won’t happen again. Bizarre corporate takeovers and restructuring, some of the most brutal working conditions we know in modern developed societies and the list goes on.
The goodwill is gone.
Governments and civil society have wizened up. Collectively we’ve realised that any actor has the potential to do anything, and we need to keep each other in check for a well functioning society – no one is exempt. The free pass the tech sector had in the past is no more and any leeway they do get is, in my humble opinion, only because the response to developments take a while to make and there is not the time and budget to respond to them all, but it’s not from a lack of will to do so.
I’ve noticed in the past few years a regular movie trope in animated films is that of a tech CEO who at first appears to be kind hearted and working for the common good, but later turns out to be an evil villain. Obviously this is a generalisation and in that sense unfair to 99% of tech CEOs out there, but it is an interesting observation, no? Art imitating life and all of that. Little kids are being exposed to this storyline that surely will influence their view of the world. I’m a little bit puzzled that this trope has been allowed to develop freely, but maybe there’s something I’m missing, or those whose image is harmed by this sort of narrative don’t have nearly the influence to stop it happening as I assumed they would have?
My general feeling is that goodwill has been replaced more generally with, at best, an acceptance that a lot has already gone too far and we can only do bits and pieces to limit the influence of the power we’ve handed over. I’m not sure I totally agree with this view but if I tap into the zeitgeist, this is kind of what I’m sensing.
The summary of this is that pre-ChatGPT developments had a lot of goodwill they could leverage, but I don’t see any evidence that this is the case any longer. Maybe even the opposite is true, this will really be put through extra scrutiny and any ounce of goodwill will have to be earned, which I don’t think is possible.
The ChatGPT business model has a huge flaw
Whilst many machine learning methods have this “black box” characteristic where we can’t explain exactly what has gone on internally to lead to a result, ChatGPT seems to rely on a “black hole” business model. It needed to suck in a huge amount of free, publicly available work to be trained on.
If we can’t figure out how to give agency, credit and value back to those creating what’s going in, then you can bet it’s going to affect what gets publicly released, or whether it will be allowed to keep having access to public material with the benefit only going to those willing and able to pay for the fruits of all that intellectual capital it was fed.
Will we start seeing licences attached to websites that prevent LLMs from scraping it for training material? Will people just post more and more stuff in closed walls, or stop posting altogether, slowly depleting the pool of good inputs I assume LLMs need? Then will we see the slow degradation of the service, like we’ve seen with Google search?
In a way you can experience the degraded service right now. Just ask for anything that might be relevant post the ChatGPT3 training material cutoff date of September 2021. You will get a response that an answer isn’t possible because of this cutoff and for some of your queries, you’ll get this immediate small internal jolt of “Aw, that sucks” and the potential of the tool goes down a few notches in your mind. It’s not too hard to imagine some sources of information will be slowly pulled out from under ChatGPT and you’ll start hitting these roadblocks because original creators are finding ways to protect their work from uses they don’t consent to.
As an aside to this, many other developments in tech seem to be based on some kind of flaw or loophole to operate. Yes they may be fulfilling a real need, but that’s sometimes predicated on subverting real laws or social contracts that are in place for good reason, or that we didn’t know we needed until we saw the effect. The likes of AirBnB and Uber being good examples. Both services I’ve used and loved, both I hope continue to operate in the future and do so responsibly, both made possible by exploiting loopholes, both have had loops closed as their dominance started hurting the “ecosystems” that enabled them to flourish in the first place, and both have settled more or less into a smaller role than was envisaged in the early days. And not to seem to bash on tech specifically, this can and does apply to many businesses, people, organisations and entities of all kinds, but they often don’t have the reach and scaling effect that tech-enabled ones seem to have.
The proponents are the same and following the same pattern of failed hypes
If you look at the aforementioned “failed” hypes, from my perspective nothing distinguishes one from another when you consider the hype cycle of each.
It just feels like the same people, hashing over the same stuff, same LinkedIn posts, same screenshots, same weird edge cases presented as new industries, same get-rich quick schemes, same fear mongering.
The only thing that changes is the name of the tech being used. It really feels like copy and pasting from one cycle to the next – honestly is quite boring. Unfortunately, we are seeing it again which is where my worry comes from that the whole thing will just kind of fade away when this development has sucked up all the time and money it can before the next one comes along.
I’m not frustrated with it, just a bit sad for humanity.
What gives me hope though, is that the public response is getting more critical, more constructive, more nuanced and more savvy with each cycle. It feels like we have been burnt too many times to just go along with things and hope our interests are being kept in mind. The most ardent of tech enthusiasts will view this as luddites holding back development, but I think its a healthy “evolutionary” response of society to realise and exercise our agency to shape the outcomes of these developments to meet our actual needs and not just allow them to be a means to an end for those who created them. As they say, no man is an island – but not everyone acknowledges this.
Stuff that was revolutionary: why did they work?
If we consider the Internet and Email, two revolutionary developments that did succeed, what factors made them succeed? Neither relied on one or two dominant players who were already dominant in other spheres to take off – and anyone could set up their own networks and email exchange servers or clients if they wished. I’m likely glossing over important technical details here and there may be more parallels drawn between ChatGPT-like development and the Internet and Email than there are differences, but there are differences. Do these differences solely account for the failure? Likely not – for example Blockchain, Crypto and NFTs all likely had the same “openness” of the early Internet – which shows that single factors are not at play.
But what does feel to be quite different here is that Chat GPT is a service (much like Google Search or Facebook is), quite different from a new technology or protocol, like the Internet and Email is.
When we’re evaluating how or why things become revolutionary, my gut is saying it is harder for products and services to do so now because so much has changed since the early 2000s. Since the iPhone was released in ~2007 (15 years ago!), there has been no new revolutionary product in that space, just variations on that same idea.
This aspect of ChatGPTs future challenge I struggle to articulate the most and maybe some tech historians can help me out, but I think the main takeaway is that we can’t treat all developments in technology the same. There are different underlying characteristics which will influence how they are affected by public sentiment, regulation and competition.
Will we be able to trust anything we see or hear?
I’ll branch out for a moment from just talking about LLMs to include the developments around other AI generated content like images, videos and sound.
The progress, especially of deep fakes, has been mind-blowingly fast. When you see the Pope looking fly bear in mind this was much harder to generate just 12 months prior. It will probably become quite difficult to establish real from fake with the naked eye and many people worry about the impact of this, especially when it comes to video content, which for a long time we’ve come to trust over still images which we’ve known for at least a few decades can be doctored. Some people are very afraid of what is to come, but for me this is hugely tempered because once we know something can be faked, we tend to then consume it more critically. I think it’s just the novelty that is catching us unawares.
I’m sure back in the day of print media being the only source of anything fake, the wise saying of “don’t believe everything you read” must have seemed quite insightful at first (even though today this goes without saying).
I’m willing to bet each of us, at some young age, encountered a tabloid for the first time and thought “but how can they just make stuff up and print it as news?” – but yeah, they did and they do. Text can be faked.
It wasn’t that long ago that “photoshopping”of images was a big deal. Many debates raged about images being faked, and this especially true of images of models and movie stars. There were campaigns being run to push that digitally-altered images be marked as such so we could see what is real vs not. Images can be faked.
CGI in movies has progressed a whole lot as well. Surely when we watch movies we don’t expect that much of what we see is real in the slightest – we just don’t expect this in our daily media consumption because we attach that sort of special effects to big budgets, with lots of people involved and hundreds of millions of dollars. Sound and video can be faked.
Now the difference is text, sound and video can be faked relatively cheaply and quickly. The next time I see something I’ll have this filter on.
I’m not saying this won’t have an impact at all, but I doubt it will have this society’s up-ending effect that some folks fear. This isn’t something new that we need to deal with, it’s just a broadening of the scope of what already exists.
Imagine this was 1970 and replace ChatGPT with “spreadsheets”
With any new tech developments, I often try to imagine what it must have been like when spreadsheets were developed in the 70s and can 100% imagine the exact same debates being had. Imagine an accountant in the 1970s working with maybe a calculator and handwritten balance books, reading about these computer spreadsheets that can do hundreds or thousands of calculations in seconds. A week’s worth of work in seconds!
I can’t see how ChatGPT and LLMs are not just tools like a spreadsheet is a tool.
Most companies are nowhere near using LLMs like they think they are
Even if I am wrong about everything else, the thing that will stop us utilising the full potential of LLMs (for better or worse) is that many organisations are nowhere near the position to actually use them. The technology of LLMs is far beyond the cultural maturity needed to make use of them.
Sure, LLMs can be powerful, but throw it a neglected systems, undervalued data quality processes, low data maturity and poor data literacy of a real business and I’m sure it’s going to hit limitations right away.
What I’m hoping is that LLMs will be able to bring the price point of fixing these systems down far enough that it’s not such a difficult sell to make investments in data systems for what is ultimately a very abstract outcome for organisations. But this is a scenario that will be here maybe 5 years from now, if all goes well.
The world is more like a spiderweb than a pane of glass
How much hope or fear you place in something is dependent on your mental model of how the world works. I think some people think of society as this pane of glass, that any one “hit” from a new development will shatter things immediately and irrevocably.
To me it’s much more like a spiderweb. Much stronger than it looks, has sticky strands and if it breaks in one place it gets repaired. It doesn’t always look the same but it functions the same. We may feel a hit vibrating across it (which is why it may feel like glass). Big changes, good or bad, are actually quite difficult to make happen everywhere all at once. Local impacts are immediately felt and because they can be repaired, we can experience them over and over again – we are able to be around for each one. It’s a feature and not a bug ;).
In summary
The cards are stacked against ChatGPT taking over everything. Public sentiment, legislation, whatever is degrading Google search, whatever hampered other recently hyped products, organizations’ lack of data maturity, and a bunch of other things I’m not aware of.
Hopefully we can figure out how to make the most of LLMs.
Keep up to date with new data posts and Big Book of R updates by signing up to my newsletter. Subscribers get a free copy of Project Management Fundamentals for Data Analysts worth $12.
Once you’ve subscribed, you’ll get a follow up email with a link to your free copy.