ML in the Middle
Basically it’s about how and why it’s tricky to introduce ML particularly to a medium-sized org, with accompanying caveats and why that actually might be wrong at the end.
Note one: when I’m talking about “ML”, I’m referring to the construction of an ML system that’s integrated into the existing business and technical infrastructure. Not some models in a NB. Not using ChatGPT to write SQL query, or Perplexity to do some research.
Note two: emphasis on introducing. If DS/ML/AI is a core component of your business proposition, it doesn’t matter what size you are. It’s definitely going to be there, duh.
Size Matters
Small companies, I’m talking really here about start-ups, these days have to answer this question overtly and immediately: is ML going to be a core component of what we do, or not?
Predicting the outcome of cricket matches? ML is intrinsic. Creating an app for companies to check employee mental health? ML not needed.
Large companies are in a surprisingly similar position: an ML system already exists (and is probably a core part of what they do), or it doesn’t. In the case of the latter, they can easily ignore it (for now) and continue with the business activities that are clearly working (we know they’re working to some extent, otherwise they wouldn’t be large). Alternatively they can throw lots and lots of money and time at the problem.
Medium-sized companies are in a trickier situation (I’m talking mostly about contemporary, tech-enabled, scale-ups). They usually have something that’s kind of working, that has been scaled to a certain extent. But this success can be a curse: the preceding period of growth/funding/building stuff that worked means that now systems and processes throughout the business may not be a good fit for the new scale.
Specific Problems
I will now go through some general challenges associated with building an ML system, and then tell you how these problems are more acute at medium-sized orgs.
Costs
Building an ML system requires resources. Medium-sized companies are often unable to commit these resources to something new that may or may not be financially beneficial.
Compare this to large companies, who make a lot of money. They can afford to throw these large piles of money at problems. This includes the problem of building an ML system.
Small companies aren’t expected to be profitable.
Time horizon
The reason MSCs are often looking for some round of funding or exit in the near future. Hence a reluctance to invest in long-term projects.
Both small and large companies have an advantage here. The big corps have a theoretically infinite time horizon to play with. They can afford to invest in something now that will have benefits in the long term. Short-term thinking is incentivised, sure, but at least they can justify long-term investment.
Start-ups have a definitively long-term view. It takes 10 years to build a really successful company, so although they are strapped for cash, they sort of have to make long-term investments.
Data
MSCs have some data, but it’s unclear if they have enough for the investment required to properly utilise this data to be justified.
Large companies obviously have reams and reams of data so this isn’t a problem. And small companies have a plan to leverage the small amount (and usually niche) of data they have, use publicly-available data, or not use ML at all.
Platform
It can be hard to justify committing resources to work on platform issues given the potential tech debt work required to re-build/improve current infrastructure. People can be reluctant to introduce new components to these systems.
Small companies, by contrast, are building from scratch so can design their systems with ML in mind.
At large companies, although their own systems are often antiquated, specific teams of people are tasked with dealing with specific issues, so there’s no risk of the data science team doing (all the) infrastructural work.
Expertise
They (large corps) can afford to hire these specific teams and experts to design and implement a system. Because although the mission might not be as attractive, the technology not as cutting-edge, and the use cases not as exciting, large corporations have multiple advantages, many of which I have mentioned.
Start-ups have the advantage of working on exciting problems, with equity in a potentially-very-valuable business, and a chance to shape ML at the company in all the ways you want to.
Scale-ups can have trouble attracting top talent. They probably have some already at the company, but it can be fairly raw/junior, and hence will take time to properly deploy. Their big selling point is joining a fast-growing company.
Other Thoughts
Let’s all work together
You can kind of see how all these things work in conjunction with each other (against MSCs).
Large corporations have something that’s working, and, they assume, will continue to work in the future, which gives them money and a long-term view. They are able to hire a specific team to implement a system that has obvious (because of their huge amounts of data) medium-to-long-term benefits.
Small corporations can attract top talent to build a system from scratch to solve a novel problem. The system doesn’t have to be immediately scalable or profitable, but has the potential to be so in the future, allowing the founding participants to be the lead of a substantial ML team at a successful company.
Medium-sized corporations are faced with a dilemma. They have a product/service that’s kind of working, although maybe not profitably. Fixing outstanding issues with the organisation, platform, and product, as well as focusing on reducing costs and proving (at least potential) profitability can take priority over introducing new components, that have questionable value-add because of a lack of 1) big data and 2) expertise to leverage this data effectively, that will have to be integrated with, and will add extra load to, the already-stressed infrastructure. They have a medium-term view: they are driven by doing what’s necessary to earn their next funding round, or exit.
It’s not all bad
MSCs do have some advantages over other companies.
Start-ups have no data or existing infrastructure. They are also poor, so can’t afford to make non-critical investments in just about anything. Data science employees may end up doing a lot of things that aren’t data science.
Large companies (some) have old infrastructure that few people in the global job market understand. They also have big risk teams, and compliance, and many board members.
It’s also not a binary problem. It’s not like companies have a choice between building a full ML system, or not. In reality, one (one as in literally one person. Me, for example) can get a basic system operational fairly quickly. Which can be iterated upon. And there’s always the option to of course not create a system and use ML for isolated projects.
So maybe it’s not so bad for scale-ups after all?