AI For Fun Or Profit ?

The Czech Academy of Sciences, the research council which funds my own employment, recently put on a five-part webinar giving detailed guidance on how to use LLMs in research. It was a good course with some useful tips and tricks and a few tools I'll try to check out eventually. The presenter seems like someone who really uses AI a lot, like for absolutely everything, but as you'd expect in a decent course, it was full of caveats : don't use it to do X, always check its citations, don't take its output for granted. 

The best line was to "treat it like a skilled researcher who's on drugs". You wouldn't discount everything they say, but you wouldn't trust them either.

It's all very common sense really. The course had about 80 attendees, and pretty much matches my real-world experience with colleagues. In every discussion, every single one, everyone gets in a bit of a circle-jerk about how useful AI is but how it can't be trusted. There's basically no-one who isn't using AI to some degree, and similarly nobody who's trusting its output without question.

After this course, I wonder if perhaps I'm not using AI enough. Would I be more productive if I did ? Possibly, to some degree. But I think not a great deal. Personally I simply see no point at all in using AI to replace my own voice : when I want to express myself, in any medium, if it's not me doing it then I might as well not bother at all. Okay, for the final polish here and there, or checking if I'd got the basics correct, or follow-ups... sure. But the basic gist of the text has got to be my own, even if it's imperfect. 

So I just don't see how it can be any help in preparing Power Point files* or writing whole paragraphs in a grant application, let alone anything in a publication**. For writing code I'm perfectly happy to let it go nuts, so long as it's doing grunt work and/or I just want something quick that works. But even then, if it's something I'm going to want to maintain, or want to understand what's going on, I find it far more valuable as an assistant, someone who can teach me and simplify things where necessary, not do the whole thing for me.

* No, we're not calling them "slide decks", thankyouverymuch. WTAF is wrong with people ? 
** The author has prompt instructions for just about everything. If nothing else, maybe some of these will at least be useful guidelines for people to follow when doing the tasks the old-fashioned way.

The use of AI chatbots is very possibly where my real-world and online lives are most dramatically at odds. Offline, AI is already normal. Like totally normal. Online, there's a far bigger fraction who are still clinging to the idea that it doesn't work, won't work, can't work, is innately immoral, etc. etc. These are sentiments I've barely encountered at all in everyday life, which veers much more towards thinking that people not using AI are either a) old or b) weirdos.

What concerns me for today's post is why we don't appear to have seen any transformation in the economy as a result of AI. It seems abundantly clear to me that AI does work, so where are the productivity gains ?

Now I'm not expecting any instant revolution. The hype train that AI will lead to FTL travel and immortality and a utopian world by next Tuesday is not worth considering. Nor do I think it's capable of fully automating any significant number of jobs anytime soon. But it is, unarguably, an extraordinarily useful tool for people using it properly. It's not unreasonable to expect that we see some measurable effects of this, so here's a round-up of some recent articles giving some different perspectives.

---

One widely-reported study found that 95% of companies had seen no measurable impact from generative AI. I asked the lecturer about this : she said she didn't know, but speculated that maybe this was using earlier models, particularly unrepresentative samples etc. This post presents some plausible rebuttals : most crucially, the question is "95% of what ?". Apparently it's 95% of all companies, not companies that actually tried using AI ! So much for that.

This Nature piece leans pretty much in my direction that AI has tangible benefits and will, like the internet, restructure things to such an extent that it's difficult to know which metric to use for judgements. It makes the perfectly reasonable point that AI is advancing so rapidly that it's already difficult to know what we should be measuring; related to this is that adoption does not necessarily keep pace with AI capabilities. All very reasonable, but still... where are the gains ? Where's the money ?

A much more bullish piece* on "Noahpinion" (I never heard of it before) looks more at the different attitudes to AI. This at least partly explains the discrepancy in my real and online worlds : Americans are among the most "AI-concerned" people on the planet. Which fits with typical American bipolarism : let's invent a thing we spend crazy amounts of money on which we really hate. And in fairness, it seems to me that Americans are vastly more likely to be shafted by their employer than Europeans, so this attitude is not at all without foundation.

* Interestingly, the author is convinced that data center water usage is unimportant but that their electricity consumption is extremely high. I've seen other articles claiming the exact opposite. I don't know. To me it feels like this is all a massive distraction on the environmentalism front : what we need is to switch generation methods to renewables and nuclear and invest heavily in storage. Bitching about AI is pointless, and usually when I look at the claims and counter-claims, it seems to me that the impact of AI is heavily overstated at best.

To go off on a slight tangent, the author also notes that complete omniscience is a myth. Yes, AI makes mistakes that humans don't, and its error fraction is higher than that of true experts. I would also note that real experts are generally more self-aware of their own limitations and vastly more likely to say "I don't know" when asked about things outside of their own domains. But still, the problem is fundamentally the same : here is a claim, how do we know to trust it ?

The answer is simple. Everyone's worried about AI fakes and manipulation, but ultimately we have to treat it just like any other source. We literally do not have access to perfection. Everything requires a degree of trust and verification; we should apply the same standards to AI as for anything else. That is, when things aren't critical, we go ahead and provisionally accept its claims. When there are consequences we need to double-check what it comes up with. That's it.

Though, a BBC piece presents and intriguing example of poisoning the well : deliberately writing a credible-sounding blog post to fool LLMs which rely on web searches. For me, ChatGPT wasn't fooled, but it's for sure an important point. Seeking out independent sources of evidence will become more important than ever.

To return to my main theme, Noahpinion claims that the effects of the AI bubble bursting are exaggerated and AI will lead to more jobs rather than less. This is all getting very murky.

For balance, a couple of more negative pieces. "Marcus on AI" is convinced that while AGI is achievable (I am not), ChatGPT will never live up to its promises (I think it's already doing better than I expected it would when GPT-3.5 was released). I think it's a strawman argument to say that because it didn't reach the absurd standards promised by the same techbros who initially claimed it was too dangerous to release (a bit of marketing genius, that) that it hasn't massively improved. Similarly I think his acceptance of the famous "95%" claim is clearly flawed as he doesn't explain his own reasoning : sure, another study finds that not many companies are using AI intensely, but this says nothing about what stage of adoption they're at or their long-terms plans. Maybe the original study he cites does say this, I don't know, but this needs to be included in any analysis.

More interesting is the claim that AI use at work is flatling or declining. But the timeline here is rather muddled, and from my own direct experience, I too became tired of GPT-4 for work use as it just offered nothing of substance. It could proofread for language and make very basic comments on the substance of the text, but it was pretty shite for discussing actual science. It would have some hits, true, but they were buried in a mountain of faff that was often just not worth the effort of digging through to get to the good stuff. And of course, it would hallucinate like nobody's business. 

ChatGPT-5, by contrast, was a massive, game-changing improvement on all counts, and that was only released six months ago : surely we should wait and see to determine what its effects are. Now this is of course is not to say GPT-5 is perfect. But I absolutely maintain my initial excited stance that this is a breakthrough which crosses important thresholds for making it an actually useful tool. 

Perhaps Marcus's most interesting claim :

If GPT-5 had solved these problems, as many people imagined it would, it would in fact of enormous economic value. But it hasn’t.

To be fair he's consistent in that he says it hasn't solved the old problems of hallucinations, lack of common sense etc. But I deny this vehemently. I think it's made massive, demonstrable, in-your-face progress, and saying that it hasn't solved these issues completely is a totally pointless claim. Of course it hasn't ! I never expected that it would. If you were actually expecting AGI, then more fool you. The whole perception of LLMs seems like a massive case study in the old quote that the perfect is the enemy of the good.

But this of course still leaves me with the dilemma : where then is this massive economic value ? We at least agree that a good AI would be of economic benefit. Marcus has the obvious "out" in that he doesn't believe the AI actually is any good, but I don't. Where, then, are the benefits I expect to see ?

One possible answer comes from Business Insider*. Perhaps, it suggests, the answer is ironically in that very demand for increased productivity. Software developers are now working not on one task at a time but on many at once, waiting in between prompts while the AI writes the code and then they clean it up. This is not a natural way of working that understandably causes a great deal of fatigue. In essence, then, AI is good enough to help, but still needs constant supervision and cannot fully automate much. It might be like self-driving cars which are actually more like driving assistants : the worst sort of grey zone, a necessary step towards something transformational but in some ways actually counterproductive in and of itself.

* I'd started skipping their pieces because they tended to be shallow and dull, but from what I've seen lately, they seem to have improved considerably. 

Much the most cynical of all the pieces here is from the I assume ironically-named "Pivot to AI". The author's case is basically that the economy is screwed, that CEOs are firing people left right and centre not because AI is actually capable but because they just love firing people and enshittification and all that. In direct contrast to Noahpinion, he says that the inevitable bursting AI bubble will be worse than the Great Depression and we'll all die, or something. Righty-ho then.

---

What are we to make of all this ?

It's very hard to say. We have two opposing hype trains : AI will transform the economy; AI will wreck the economy. So far it seems to have done neither.

Falling back on my own direct experience, AI is undeniably helping. It's allowing me to tackle things I wouldn't have done otherwise and understand things very much more quickly than I would otherwise. It has not yet had a measurable effect on my actual productive output as typical metrics would indicate (papers and the like), but since only GPT-5+ looks capable of influencing this, and this has only been out for six months, it's probably too early to judge on that score. 

By my own internal metrics it's definitely had a measurable output, allowing me to generate quite a lot of code I would otherwise have liked to have but never have gotten around to writing. Last year I even spent quite a long time using it to write a 75+ page introduction to radio astronomy that helped me enormously... if I ever have time to finish it, I'll put it online somewhere.

But shouldn't AI be solving the "if I ever have time" problem ? Yes and no. The thing is, the bottlenecks in my productivity lie elsewhere : primarily, meetings. In a busy period I might have one or even two meetings a day, which all told take up a full working day each week. Not all of these are useless (although some of them are), but in terms of actually getting stuff done, almost all of them have a negative impact. 

Likewise, not everything I work on is directly tied to productive output. I need to experiment and understand and pursue blind alleys. AI can help with some of this, but not all : in essence, it can alleviate a small amount of my workload to a very large degree. It can't help at all where my code tests are limited by other factors such as download speeds. I don't even want it to automate everything, because if I can't understand the science, what's the point ? Yes, it can help me understand things, but I'm the bottleneck here, not the AI.

From my perspective the answer is clear : AI has only very recently reached the point of being seriously useful, we're all still adjusting to how to best to use it, and there are many things it either can't do or we don't want it to do. This would suggest that we ought to see more substantial improvements in productivity on a timescale of a year or two, allowing for human adjustment, but those will be gains at the level of "nice to have" and won't herald a scientific revolution. Of course, "a year or two" is a crazy long time in AI-circles, and it's anyone's guess if it will finally have hit a wall by then or if it will continue to make radical gains (there are other avenues for improvement besides data volume).

But what of everyone else ? My only answer is for that we'd need a detailed study of the different working practises across multiple sectors. It is not enough to simply say "well it can do task X, your job revolves around task X, so you'll be a million times more productive now". All jobs require a lot of secondary tasks to facilitate the sharp end of productive output and not all of them can be automated. This is perhaps naivety on behalf of the techbros, thinking that because some key component can be done by robots that people will automatically adopt this practise and/or that productivity will be impacted in a linear, predictable fashion : as per the Nature article, it's far more likely that this will lead to more complex, systemic change.

Which means the answers are plausibly a combination of :

• Seriously capable AI has only just arrived, with earlier models massively overrated.
• We don't have good metrics for judging the efficacy of AI outside the lab.
• Poor management strategies can mean that AI can make things worse, not better, even if it's ostensibly extremely powerful.
• AI can improve some tasks enormously, but even when these are the most important part of a job, they are often far from the whole story. 

In short, AI has crossed a threshold for usefulness. It certainly hasn't crossed all such thresholds, and it's far from clear it's anywhere near doing so. Understanding the impact is a sociological and business problem every bit as much as it is a technological one. The good news for the AI enthusiasts is that it definitely does work; the bad news is that implementing this in a profit-generating way is anything but straightforward.