Why we avoid the term AI on our website and in our messaging
TL;DR
- A successful business addresses a market need that is unmet, over-, or underserved, with a solution that costs less than the user assigns value to it
- Successful business is about providing solutions then — not (necessarily) technologies. Nonetheless, there is growing number of businesses (starting up and established) that market their offering with “AI” — a technology at best, not a solution
- We think the AI term is overused and we ourselves avoid using it for our own business mainly for two reasons:
1) the definition and meaning of it are far from clear & agreed upon, nobody is really doing AI
2) It does not serve your business in the long run — people don’t care as much about technologies, as they do about solutions
Read on to learn more about the meaning and state of AI — and why businesses should avoid fueling the AI hype by inserting the term wherever possible.
Reason one: we don’t know what we are talking about — nobody is really doing “AI”
By “we”, in the above statement, I mean everybody. To clarify, we, as Varia, provide solutions that enable efficient journalistic research and we are specialized in natural language processing (NLP), by leveraging latest machine learning (ML) algorithms and methods.
Many would call what we do at Varia “AI”, especially those who see the world in AI/ML/DL (you know the famous three circles). We are even featured in reports on “AI in Media”. Companies and startups that (develop and) apply far simpler algorithms than ours, are selling their services with phrase like “latest AI technologies” or “AI supported”. We don’t. Here is why:
“AI” as such is a misnomer for pretty much everything that is going on in the field to date. To properly define AI, would require that we fully understand “I”, intelligence, as in humans, or other biological beings first. We are far from that, as beautifully discussed in this podcast episode of the Lex Fridman podcast, with Michael Jordan.
If you want to jump to the discussion about the AI term, use this link, that brings you directly to minute 15.
We have an extremely limited understanding of the intelligence embodied in the human brain and in our reasoning processes. As Jordan puts it, all the efforts that are done in the field today are not towards that intelligence — as we cannot work towards something we don’t yet understand. Instead, we try to produce stable systems that help us automate one-dimensional tasks. And these systems are of course of many different forms (RPA, macros, rule-based systems, ML, etc.).
However, one must think of intelligence as multidimensional. And many research efforts today go towards that, but its baby steps (e.g. symbolists), and we still don’t know how the end result should look like. Only a combined processing and meaningful reacting to a broad type of inputs (visual, audio, text, further sensory data) by one model will approximate what we observe as truly intelligent behaviour. And even if we can apply such systems to multiple domains, there is absolutely no guarantee that reasoning will result as an emergent property.
Most companies today work specialized in one field, hence at best they work on one sub-dimension of AI. This applies also to us; as mentioned above, we work with text — convert text into numerical data (vectorization) that allows us to derive sentiments, semantics, entities and other textual relations. To do that, we apply ML methods and algorithms — which are not bound to the NLP domain, but the ones we use (e.g. transformer architecture based models) are particularly useful for the contextual mapping and understanding, which is required when dealing with language.
This is what everybody does, on an abstract level: trying to optimize an automated pattern recognition process via closed feedback learning. Whether the patterns that you are trying to recognize are coming from text, audio, or video data is almost a non-relevant question at that level. All is converted into numerical values and fed to learning systems in (mostly) vector form. That’s it, machine learning. If you want, you can go down the definition road to differentiate between ML and DL (deep learning), where you separate simple regression based models and neural network architectures with multiple hidden layers. I am ok with that debate, but it does not really merit anyone.
To recap, what we do so far could be framed as “narrow intelligence”, as it (only) excels at processing one type of input — in an environment of controlled inflow of data. However, as specified above, intelligence must be thought of as multidimensional, which this is not, as it generalizes barely on any new form of data. To put “narrow” in front of a term that we don’t fully understand, to make it feel better — feels only one thing: wrong.
Many researchers (as Francois Chollet) of the field, use generalization as a way to categorize intelligent systems. All the “AI” done today is on the level of local generalization — and trying to make that truly robust, which, to Gary Marcus is one of the key challenges facing the filed. The commentary in the above graphic mentions “human-level driving” as one example of broad generalization. We are not there yet. Even with millions of real and generated data samples, the current level of intelligent driving systems does not generalize broad enough to handle traffic cones in rather common cases.
The discussion about generalization, intelligence and the limits of applicability of machine learning in bounded vs. unbounded, discrete vs. non-discrete problems could go much further. But this is not the point of this article. If you are interested in the debate, this talk/podcast by ML-Street Talk with Francois Chollet could be a great way forward:
We believe it serves no one to talk of AI, while on the other hand, if everybody would name names — use precise language and say what they are actually doing — some light could be shed into the dark alley of confusion that leads on to AI hype-way.
Journalists, VCs, consultancies, businesses all play a big role in this. What if for every “AI” term in coverage of what a company or startup does, we just switch to talking about the real solution provided & methods applied?
“We are solving the X problem of Y (by applying Z)”. Z, if stated at all, in that case should not be “AI”, but the actual machine learning or automation method applied.
Let me close off this first section with a differentiation between AI and Machine Learning, by David Kriesel:
“Machine Learning is usually written in a programming language, AI is usually written in PowerPoint”.
Reason 2: At the end of the day, customers and users do not care about your approach, they care about solutions, they care about the result
Here is a story to exemplify what is meant by this: In the acquisition process for our first big customer, we were given a trial dataset, on which we had to prove that we actually can do, what we promise on our slides and homepage. We were given a dataset and some instructions about target results — along with a timeline: two weeks.
We alerted the customer that two weeks for the sort and scope of trial-problem that we were given, might be too much — and that other companies might decide to do all the work manually. This could be done by the teams themselves, or the work could be offloaded to third parties in a low cost country of your choice (AI here, as in it’s “All Indians”). But here comes the point of the story, the customer’s response to our alert: they could not care less. “Even if some companies are doing this, we would not care. As long as the result, in terms of performance and price is right, why would we care if the applied technology is actually manual labor?”.
In all fairness, this test did not go towards eventual scalability of the service, but for many tasks, this could also be achieved with manual labor. The point of the story, is the result orientation of our customer. I would love to see that kind of result-oriented thinking in the companies offering “AI” services.
As a business owner, ask yourself: What do you deliver? What do you do? What problem do you solve? Sell your results (customer success), your solution — not your technology, not your means to an end! Technologies tend to evolve much faster than customer needs, hence companies should focus on solving a customer need, and sell a solution that uniquely addresses that need — not a hyped technology.
There is a reason the innovation framework by Clayton Christensen et al. is called “jobs to be done” and not “technology to be used”.
Furthermore, be aware that the automation component (through whatever ML or other) is only a small part of a working solution. In the greater product scope, many more mission-critical elements (throughout the entire customer journey) contribute to the UX that will win or lose a customer.
For many online forms, of events, accelerators, VCs, etc. we already had to declare a “key technology” or the like, that we employ at Varia. Not seldom, there is a checkbox for “AI”, while there is none for “NLP” or “Machine Learning”. How is that tick at “AI” helpful? It does not tell you anything.
If you think that this emphasis on the right terms and against buzzwordery is exaggerated, let me wake you up to the reality that “40% of European AI startups don’t actually apply AI”.
While the study on which the above article bases was initiated by a VC, it has to be mentioned that in most cases VCs themselves are not helping the cause here. There are even VCs that crawl the web for startup portfolio candidates, by explicitly searching for “AI”. And I can tell you, we have gone full circle: I know founders who purposely distribute the AI term over their homepage — admitting and knowing that they do absolutely nothing in that realm — just to lure in the attention of hype-minded capital.
Further testament to this are the now omnipresent “top AI startups/ companies” posters and lists. If you treat AI as synonymous with all sorts of automation, then you will find AI in all industries. Then what makes a firm “top”? The criteria and selection here tell you that there is not much science involved in this, rather, that these lists and posters are mostly a marketing exercise.
I am pretty sure, when we look back to this point in time, say in 20 years, we will laugh at all the businesses that called themselves or their domain “something.ai”. We already today laugh at efforts that go by titles such as “Design Thinking for AI”. This indicates a deep lack of understanding of both Design Thinking and AI.
At the same time, there are of course a few companies who do actual research oriented work and try to push boundaries of inference and generalization. But even OpenAI or DeepMind are at this point far from reasoning systems and broad generalization. Still, if you, as a company are focused on research oriented work on the generalization of inference and automated systems — towards artificial (general) intelligence, then I guess calling yourself xyz.ai seems even reasonable.
“But what about GPT3?” you ask? This is not reasoning either, and no “first sign of AGI” as some news outlets called it. To dismantle the myth around the language model by OpenAI, there is no one better than Walid Saba.
Keep in mind, even if I might come across as a luddite here (at least for some), I am very much a technology enthusiast and try to explore, understand and exploit all the possibilities that new advances in the field of machine learning provide. All I am asking for is prudence with the “intelligence” term and avoidance of the “AI” term, thank you.
I also want to stress that these are my own opinions and do not represent those of our entire company. While we probably agree on most of what is written here (as I remain rather abstract), at Varia we regularly have intense debates (e.g. over a remote after work beer) about exactly these topics. What is intelligence? What is inference? And which requires the other?
Thank you for your reading time & feedback !
This article was written and researched using Varia Research.
