Will Your Manager’s New Favourite Technology Take Over The World? A Handy Guide On Why It Will Not

We’ve all been in this situation: You’re sitting in a meeting and some manager goes "Hey, this thing we’re doing is cool. How about we use (some currently hyped technology) for it?". We’ve all then thought "You have no idea what you’re talking about, and it shows", because usually the technology they suggest doesn’t actually help with what you’re doing.

The specific buzzwords that are currently being hyped change every once a few years - a while ago it was blockchain, at the time of writing it’s generative AI, and I don’t know if I’m looking forward to finding out what comes next. To prepare ourselves for it, let’s take a look at those two technologies and why they aren’t going to significantly change the world. Ideally, we’ll figure out a few questions to help find out whether something new is worthwhile.

Note: Some of the examples I’m using happened in germany, thus most coverage on it is german. Links that lead to german articles are tagged with "(DE)".

Use Cases

Having a solution that’s technologically interesting is cool for sure, but it needs to have some purpose. So let’s identify what use cases people are pitching for blockchain and generative AI/LLMs:

Blockchain is mostly pitched as a decentral way to manage ownership. Whether that is ownership of a cryptocurrency, of some ape pictures or even of property, the idea is that we can publicly keep track of it without having some third party we need to trust (such as a bank or a government). For demonstration purposes, we’ll use two examples: Currency as an example of something with rather high stakes, and cards for a trading card game as something with low stakes.

With generative AI, the use cases suggested for it are spread a bit wider, mainly because nobody knows what they actually want to do with it. I do think we can split it up into two sets of use cases: One would be asking it for information (such as the Baden-Württemberg government using it to help interpret laws (DE)), and the other would be asking it to generate something new (such as concept art). Let’s keep these two use cases as examples. I’m going to say "LLMs" from now on for brevity, but I’m including all generative AI that has a similar approach, not just text models.

Question 1: Does it fulfill that use case?

This is probably the most basic question you can ask: Does the technology in question actually offer a solution to the problems it’s trying to solve?

With blockchain and a high stakes use case like cryptocurrency, we get a "kind of yes" if we don’t look into it too deeply. Yes, we can record and transfer ownership of things, that all works great. However, if we have something important, we also better make sure that it’s secure, and that’s where the "kind of yes" turns into a "no". Seeing as the entire purpose of blockchain is that there is no single trusted third party, there can’t be such a thing as a police. If someone steals your stuff, your stuff is stolen and you aren’t getting it back, period - as seen for example with the infamous "All my apes gone" tweet.

"But wait", I hear you say, "the apes did get frozen by OpenSea and he got some of them back!". That’s correct, they got frozen, but the important part here is "by OpenSea". For the uninitiated, OpenSea is a third party trading place for NFTs, basically offering a nice front-end for accessing them. So not only could the thief still trade those NFTs outside of OpenSea, they also become a trusted third party, so now we’re back to the thing blockchains were supposed to avoid and the whole ordeal became kind of pointless.

There are additional problems with blockchain managing important things, such as the fact that you can’t change your password/key. Because all past data is immutable, it can’t be changed to now be encrypted with or linked to a new key. LastPass users are currently experiencing the consequences of that - the LastPass breach took place almost a whole year before that report, and in a normal system that’d have been plenty of time to change all your passwords and thus render the stolen data worthless. And again, they can’t get their crypto back either, because there’s nobody to give it back to them.

So in conclusion, blockchain cannot be used for high stakes situations as any sort of security breach (and security issues will come up eventually, there is no such thing as a 100% secure system, and as we’ve seen above the system doesn’t even need to be flawed for phishing and theft to occur) can’t be stopped or undone once it’s happened.

We still have the low stakes use case for blockchain though, and it doesn’t suffer from this problem as much. If you lose a few trading cards, that’s a lot less significant than losing the property your house stands on (at least it’s supposed to, I know first edition Charizard exists). Thus, I think we can give low stakes use cases a pass for this question.

Now when it comes to LLMs, in order to answer the question, we need to remember what they do: The entire concept boils down to using statistics to figure out which word is more likely to be put after the last one. A LLM doesn’t have any understanding of the underlying concepts or knowledge about what any of the things it says mean, it just crams out something that looks roughly like the data it was fed. You could cynically say that it’s basically a bullshit generator - generating some text that looks right, but whether it is correct is coincidence. Coincidence they’re trained towards, but still coincidence, and we’ve already seen hilarious cases of users making LLMs dream up cURL vulnerabilities or even precedence cases for a lawsuit that never existed, all the while believing they’re being given correct information.

So factual information is not their strong suit, but where LLMs excel is at generating something that only needs to look right, but doesn’t need to conform to any truth. An example for that would be concept art or backgrounds. There’s no such thing as "right" or "wrong" in art, the only (optional ;D) criterion is looking good, and we’ve established that that’s what they’re good at. Again, we can give this use case a pass for this question.

Question 2: Is the use case worth the cost?

There is a small second question wrapped into this one that I didn’t feel deserved its own headline, which is "Does the use case make sense in the first place?". We kinda got into the example I’m going to use in question 1, but the prime example for this one is blockchain. Seeing as the entire point of it is to be de-central and not to have a trusted authority, it’s ironic how the history of blockchain so far is riddled with examples for why having some authority can be pretty practical actually. This also shows in how quickly centralised platforms (that again got some authority) such as the aforementioned OpenSea evolved. And all this is not to speak of all the examples of people using blockchain for data where there absolutely is a central authority (I’m still salty about all the tax euros wasted on the blockchain-based school certificates (DE)).

That aside, both blockchain and LLMs are very expensive technologies. Blockchain systems depend on many different computers storing the same data (and, sometimes, using very expensive mechanisms to agree on who gets to write new data), so not only is there lots of energy being consumed, but there’s a cost linked to transactions. According to Etherscan, at the time of writing, selling something using OpenSea costs you a bit more than 1.5 dollars. That’s fine for high-stakes uses where the transaction involves much more expensive things, but when you’re using our trading card game example from earlier, that’d mean giving a card to someone else costs you 1.5 dollars. As a comparison, a similar query on a simple server with a MySQL database would cost next to nothing. So while blockchain systems could manage low-stakes systems, it absolutely wouldn’t be worth the cost.

LLMs need a lot, and I mean a lot of training data to generate results as decent as ChatGPT or Midjourney do, and then need to process all that data. There aren’t really clear numbers on the amount of energy required to train one, I’ve seen estimations such as 1.2 to 10 gigawatt-hours used to train ChatGPT. OpenAI’s Sam Altman stated that a single request costs "probably single-digits cents", which a StackExchange user has calculated to mean roughly 300 Wh per request. As a comparison, a google search costs roughly 0.3 Wh.

Seeing as LLMs could only be used to create concept art (and I’d argue art that isn’t made by humans is missing the point, but that’s a different rant) and similarly "looks-correct-but-doesn’t-need-to-be" products, I think they’ll end up having rather niche use cases, which wouldn’t justify the huge training cost required to get good results.

So by now, we’ve eliminated basically all use cases for LLMs and blockchain as either not being fulfilled or not being worth it. But for the hell of it, let’s ask one more question:

Question 3: Are there major unresolved problems?

So far, we’ve only discussed technical limitations, but problems that stop a technology from being useful go way beyond that. As an example, AI companies are being sued left and right because they used data from people’s works without their consent. LLMs depend on having loads of training data, but you realistically can’t acquire that much of it while asking all the people who made that art, software or other work for their consent. There is at least kind of an opt-out system where OpenAI/ChatGPT respects being excluded via a robots.txt file, but that can’t be sufficient, especially as most artists have zero control over the robots policy on the platforms they post their art to. Additionally, even if I consented to OpenAI collecting my data from, say, this blog, it does also contain personally identifiable information, so I need to be able to request it being taken down as per the GDPR. You can’t simply take data out of a model after it’s been trained on it, so that’d mean periodically having to retrain the model to comply with GDPR takedown requests, which would mean there’s even more of an energy cost there.

There are other problems related to generative AI, such as who owns the things it generates, but I think this example goes to show that even if it didn’t have the problems I’ve addressed before, there still are practical problems that will eventually come to bite AI companies.

Conclusion

I’ll happily admit that I’m a bit of a skeptic boomer when it comes to new hyped technologies, but I think I have good reason for that - as shown above, the technologies we’ve discussed today either hardly have any use cases or can’t justify their associated costs with their use cases. I think the questions we’ve established are a good baseline check for whether some new technology is useful - kind of like the Bechdel test, where passing it doesn’t have to mean the thing in question is great, but not passing it is a pretty strong indicator that there’s something wrong with it. So keep them in mind for the future, and maybe also tell your managers about them. Help make work meetings a better place.