Anti-Data-Science: applying Taleb’s barbell in the AI world [1/2]
As a hands-on AI/ML professional and entrepreneur in this field, I hear from industry colleagues and my clients a continuous mantra about
“we need more data, otherwise we can’t use AI to solve this problem / launch this feature / improve this process”
Big players in the field are supporting this message and earning big bucks selling their data processing and labeling platforms, AutoML pipelines designed specifically for tons of the terabyte-measured data to be fed on. But we need to remember that during the gold rush sellers of the shovels earn the most ;)
Such a homogeneous industry message made me remember Taleb’s barbell concept, which basically tells that you can go and do what everyone does, having low risk since it’s a known and conservative path, but also having low, market-average returns. Otherwise, if you want high returns, be ready to be contrarian and aggressive, expect higher risks, and, of course, above-market returns. In the following two articles, I will draft the novel for the industry concept of the AI Taleb’s barbell, which I initially presented at the AI Ukraine conference, received encouraging feedback about it, and finally am happy to share it with you as a potential model for your project's evaluation. The follow-up article is here.
Extremely conservative investments in AI
In this article, I will start with a conservative approach, since it’s more widely described, discussed, and actually used in the community. I want to look at it through the prism of the three dimensions:
- Business model: what results do we expect from the classic “big data approach”? What is the unit economics of every prediction made by a traditionally built ML algorithm?
- Technology: what are the typical algorithms and approaches used here? What do they require?
- People: what kind of people drive such products? What’s their background and focus?
The business model of conservative AI
What are typical book AI/ML applications? We can just look at the McKinsey report and see operations optimization, call center automation, service analytics, sales&demand forecasting. The most interesting part is that the unit economics of every prediction in these cases is actually linear by nature. You can relatively easily estimate how much time/money you lose on the baseline (human decision-making or none at all) and compare it with the ML model performance (that also has its own accuracy and underlying costs) and just ensure that the profit from adding the AI/ML solution compared to the baseline is higher than the costs on it and losses from mistakes that are going to happen anyway:
As you can see, the formula is completely linear and timeless. It means, that we are cannot expect some exponential non-linear growth with time or other compounding effects. Just clear linear relationship: more data — more accuracy — more returns. Every prediction is measured like this. Straightforward, simple, and slightly boring. Big-4 consulting style :) It’s also worth to notice, that all the data is coming from the celebrated “digitalization” process — documents, photos, speech records that today are natively digital and are available in huge amounts, became a food and defensive moat for most AI companies.
The technology of conservative AI
The status-quo in the world of technology completely supports this “big data” paradigm. Most of the managed services by the biggest cloud providers as AWS have built-in basic computer vision, text, voice, tabular data, time series analysis tools that just wait for a lot of data to be fed on. What do we mean by “a lot”? For example, to train the current state-of-the-art NLP models like GPT-3 expect hundreds of billions of tokens (for simplicity let’s consider them just words) to be present in your dataset. Of course, you can start with less, maybe with millions or even thousands, but most technical solutions and research papers will repeat the mantra: more data — higher accuracy. Why it is so? because mainly we are just looking for the repeating patterns that it’s hard to find for humans due to the volume of the data and its high dimensionality. And there is only one technical success criteria — accuracy or similar metric that is telling how close your predictions are to something that other people labeled before in the data.
People of conservative AI
To complete the picture, it’s interesting to look at the people who are actually building such AI solutions and bringing them to the world, often ending with multi-billion evaluated companies. From data scientists on most of the courses and educational resources, it’s expected to get savvy with maths and programming skills (to process and clean a lot of data of course) and get domain knowledge together with communication skills (you need to know how to clean the data correctly and communicate why with more data accuracy is going to grow). At the conference where I gave the speech, I also mentioned a couple of prominent AI software startups. If you look at founders' teams and board of directors, there are mainly business and marketing people, who again, know very well where a lot of data lies and what patterns will bring those business metrics improvements we have discussed before. Last but not least, who are the customers of such products? Blue-collar or similar workers who need assistance from the speed, accuracy, and scaling point of view in their relatively simple and routine tasks.
The small questions about big data
The tables above are incomplete and summarize only the typical “big data” approach to AI products. Don’t get me wrong, this is still new technology, it is absolutely disruptive and brings massive value to the economy through different sectors. So what can be more disruptive than disruptive? Something that all these innovators don’t do or don’t see coming?
- What can be more non-linear from the business model point of view and empowering more the creative work?
- What can require less data, fewer resources, and be more generation-oriented, not only pattern recognition?
- What can attract highly creative or highly scientific professionals and how they can leverage AI?
The hint is in the table columns — “aggressive zero data” — and this will be the topic for the next article. Stay tuned!
I am open to discussions and collaborations in the technology space, you can connect with me on Facebook or LinkedIn, where I regularly post some AI-related articles or news opinions that are too short for Medium.