It's 2024, and based on what you can read on the Internet, everyone is a Gen AI expert already 😄. Well, honestly, I am not. I mean, yes - I do use a bunch of Gen AI tools daily, and as I work for a company that is very involved in this gold fever (we "sell the shovels & pickaxes", if you know what I refer to ...), I do have hands-on practical experience with building LLM-powered apps, but that's lightyears away (IMHO) from real-life expertise that comes from training & fine-tuning new models, building entirely new categories of tools (e.g., to identify AI-generated stuff or detect hallucinations). In short, I don't even play in the same league as real experts who do such things.

Nevertheless, I enjoy observing the Gen AI "landscape" from a product perspective: value-added speculations, unit economics forecasts, stages of the adoption cycle, and early-stage use cases. The last 18 months have already provided a lot of food for thought. Some early predictions have come true, but there were also many surprises. Today, we have far more data to work with. That's why I've decided to create this post - my predictions on what will happen in the Gen AI space in the forthcoming months (until the end of 2024). Yes, there's a significant chance I'll make a fool of myself 🤷, as we're still talking about a very innovative field of computer science. Still, I believe there are some visible patterns & early signs of future events one could use to attempt some forecasting ...

Observations

I'll start with some observations I find important at this point.

  1. This "hype" is not like all the other hypes before it. It's pretty clear that Gen AI has practical applications. We don't have to hope we'll discover them one day.
  2. The "hype cycle" (as defined by Gartner), in this case, apparently advances significantly faster than in the majority of cases before - it seems we've already started getting into the "plateau of productivity" (❗).
  3. In early 2023, when new models kept appearing left & right nearly every day, I was sure that we'd have a highly competitive race period. Every few weeks, a new model (trained by a different company) would take the lead as a new reigning king of capability. That didn't happen: GPT-4 has yet to be dethroned, and no one has beaten Midjourney so far.
  4. The quality of the trained model is more (than we initially thought) dependent on the fine-tuning process. Input data & raw computing power are insufficient to gain an advantage in this field.
  5. It's one of the few industries where there's no prize for second place: why would anyone use a subpar model when it's so easy to jump ships?

These observations lead to some important conclusions:

LLMs will stay with us, either as a "neutral" layer on top of other technologies or as a "gel" that integrates them. We'll get used to their convenience quickly but also stop noticing them. That is a bit of a problem as Gen AI is an expensive technology, and it will most likely get even more expensive pretty soon. Why?

  • To make profits, you need to be number one, which requires massive investments (In doubt? Check who you're competing against).
  • The owners of the data (that is fed to LLMs) have realized their lifetime opportunity and started to openly ask for their share of the prize (vide recent NYP trial)
  • Gen AI solutions will have to be accompanied by run-time guardrails (potentially also powered by Gen AI) to detect hallucinations, breaches of IP laws, or applying unwanted bias - which will inevitably increase the cost of usage.

OK, but this post was supposed to be about predictions. What will happen with Gen AI in 2024? Let's try to play an oracle ...

Unit economics kick back

The following 12 months will be a year of unit economics in Gen AI. This business is super-expensive to operate, and many companies (who have raised mountains of gold 🪙) have yet to prove even a future perspective of profitability. So, I expect buy-outs, takeovers, mergers, and companies going underwater or pivoting into uninhabited niches (with highly specialized LLMs).

The issue of trust

The testing of LLMs will rise to be a primary blocker in Gen AI adoption (instead of data privacy). Why so? By nature, LLMs are probabilistic (instead of deterministic). With NL (Natural Language) as an interface, it's impossible to test all the cases/combinations of cases. One can't rule out subjectivity from the process of preparing the LLM - the borderlines between what is acceptable and what is not are extremely hard to define & assess (objectively).

The testing issue is also related to the biggest lie about LLMs - the one about "open source" models. In fact, there's nothing "open source" about them. What you get with such a model (in the optimistic case) is the model itself, weights, user guide, license, use policy, and some metadata. You don't get access to (or details about) the actual data used in training or the fine-tuning process details and its guardrails - how "open" (/transparent) is that? 😆

The adequate description (for such models) would be "open to free usage (under certain conditions)".

Last but not least, with the top LLMs available as endpoints/APIs, there's always an issue of release transparency. You have barely any idea what has changed in the new version of the model (forget any "changelog" ...), and in fact - you may not even know that there was a release after all (as obviously there's no breaking change in NL-based "API").

The chatbot disillusionment

I find it quite funny that 18 months after ChatGPT has shaken the Internet, many people are still fixed on the single use case for LLMs: chatbots. Their disappointment is just around the corner ... Yes, Gen AI-powered chatbots may be nice to use, but:

  • They are expensive 💰.
  • In the majority of popular cases (1st line of support, internal knowledge base), they are not THAT MUCH better (when compared to "traditional" automata) to create a "wow effect" (that would justify the cost).
  • Sadly, they are much more unpredictable (especially when exposed to the open Internet)—we have already had many proofs of that (e.g., the Air Canada case).

The disruption is quietly happening. Somewhere else.

The true potential is in the industries/organizations that rely heavily on knowledge/creative workers whose main value-added is based upon memorizing a lot of information (lawyers, doctors, "experts" of all kinds) or using a replicable skill(s), which is not about physical labor (all kinds of artists).

LLMs can't replace these roles completely. There'll be a new wave of Gen AI-powered assistant tools that will perform analysis / create artifacts, but it will be the humans' role to feed them with data and assess/validate the output before it is used. Why will that be so beneficial?

  1. One person will be able to perform the work previously completed by ten people (because Gen AI does the tedious/time-consuming parts).
  2. The cognitive load of an LLM-powered bot (context) may extend to a much higher level - w/o increasing the probability of a "human" mistake. In other words: "Gen AI doctor" may analyze much more information than a "protein" one (& hence provide a more accurate diagnosis).

What will be the practical outcome for these industries? Many professions will evolve / fork (e.g., medical technician). In some professions (where the demand is stable), the unit cost of work will drop (making them less profitable), but in some other ones (where the demand was never met), we may see an unexpected breakthrough of some sort (e.g., because the scale of service availability increases exponentially).

Future of Gen AI solutions

The true power of Gen AI is not just in LLMs themselves but in creative combining LLMs with agents (traditional systems' endpoints, up-to-date sources of real-time data, etc.). That kind of architecture not only helps reduce the hallucination effects but also expands LLM capabilities w/o skyrocketing the costs (of continuous fine-tuning) by specializing the models & their RAG knowledge bases.

The "wow effect" (of Gen AI) will not grow exponentially or even linearly with the number of tokens or length of the context of forthcoming models. The law of diminishing returns kicks in here. Some events in the public space indicate that the next serious progression step in the Gen AI space will be related to a significant revamp of hardware architectures (optimized for transformer processing):

Parting words

LLMs are not going anywhere, so it's time to get used to them (if you haven't so far). They will NOT replace 1:1 anything we've used so far. Still, as they introduce NL as a primary interaction interface, they are a perfect "gelling" layer to combine information from various sources and automate simple processing (that can be precisely described with LP). That's a significant leap forward (for humanity), but it's NOT an artificial intelligence yet. It's not even close to real AI.

However, keep in mind that the Gen AI-powered solutions have a high chance of being more intuitive, more productive & far simpler to comprehend (for an average Joe) than "traditional" applications - that's why the reasonable application of LLMs will give you a competitive edge in the foreseeable future. Use this advantage wisely.

Share this post