the spreadsheet revolution
we built our tools and they built us
Hello and sorry for the silence last week – I am at a conference, and although I had grand dreams of talking all day and posting all night, the flesh was weak.
In the meantime, though, Henry Farrell beat me at my own game – his latest post on the importance of Large Language Models in the near term future of management is something I really wish I’d written, and you should absolutely drop this post and go over and read his.
Welcome back. Before responding to Henry’s post (or potentially instead of doing so, as I am not actually back from the conference scene yet and might get tired), I think it might be interesting to expand on one of his asides:
“If LLMs are radically transformative, it will be in the apparently boring ways that the filing cabinet, and the spreadsheet were transformative”
As you know, I am someone with an unusual set of views about what’s boring, but the influence of the humble spreadsheet on the modern world really shouldn’t be understated, and might be worth considering at slightly greater length.
I joined the financial workforce in 1994, about ten years after the launch of Lotus 1-2-3 and just as the global hegemony of Excel was getting firmly established. So although there were oldsters around who remembered the days of having to set out your forecasts on a literal spread sheet of paper (and one or two cranky old Grandpa Simpson types who still did so!), I’ve always lived in the world of financial models in electronic spreadsheets.
Financial analysis is, in huge part, a game of ratios. When doing training, I usually define it as “the art of dividing one number by another, then asking if the result makes business sense”. It’s often the case that you can have much stronger intuitions about the relationship between two lines of a set of accounts than about either independently. And it’s extremely easy, when you’re making forecasts, to make reasonable-seeming assumptions which are actually ridiculous – they imply profit margins doubling in a competitive industry, or market shares that add up to more than 100%, or debt tripling without affecting interest costs, or something.
Obviously, then, it’s an absolute boon to have a computer program that calculates all the ratios for you, and updates them all every time you tweak one of your assumptions. Excel made a whole new style of working possible for the financial industry in two ways. First, it facilitated the creation of much bigger and more detailed financial models; in the days when you had to fill everything in yourself having worked it out on a pocket calculator, you would really think twice about whether you really needed to estimate staff costs separately from other costs, let alone about how many divisions and geographies you were going to model for big companies.
But much more importantly, it allowed you to work iteratively. Rather than thinking about what assumptions made the most business sense, then sitting down to project them, Excel encouraged you to just set out the forecasts, then sit around tweaking the assumptions up and down until you got an answer you could live with.
Or, for that matter, an answer that your boss could live with. Ever since the 1990s, Excel has had the “Solver” feature (and later on, its dumbed-down cousin, “Goal Seek”). Because one of the biggest use cases for it has always been the production of reasonable-looking projections to justify an outcome which had already been selected.
This hugely changed the kinds of management relationships it was possible to have, and the kinds of relationships which companies actually did have. Excel gave us financialisation, because it meant that it became possible to give answers to business questions in exactly the kind of format that finance textbooks said you should give them – something like “what will be the effect of this decision on the net present value of our future cash flows?”. In a lot of cases, and for the reasons alluded to above, these answers were the most ferocious self-serving nonsense, but it was possible to give them. And because the information was being presented just-so, other kinds of information which you couldn’t put in a spreadsheet became much less vivid and salient. (I think this may have become even more of a problem when PowerPoint came in, and the spreadsheet guys were able to draw charts).
So, anyway, go back and reread Henry’s post, it’s that good. Management is the management of information (the conversion of information into decisions), and so any significant new information technology should result in a managerial revolution. I think I’d make only one caveat.
And that’s that if you look at the use cases for LLMs in Henry’s post (which I think pretty much summarise the use cases as we are now), they are all in the category “ah yes, I can see how that might be a bit useful”. Spreadsheets were the original “killer app”, which phrase was coined to mean “a piece of software so great that people will buy a computer in order to run it and/or redesign their business processes around it”. Over the three decades of my working life, there have been lots of “Excel-killer” products launched, many of them incorporating machine learning features, and none of them really got anywhere. I do think that LLMs will be an important technology and will have consequences for management, but they haven’t found their killer application yet.
Another final admin note – quite a few people have got in touch with me about the murder of that healthcare executive, and the public reaction to it, asking whether this might be an example of the sort of thing I’m talking about in the last chapter of my book in terms of the breakdown of conventional feedback mechanisms. All I can really say is that - I definitely know what you mean, but I really don’t trust myself to address something like this without getting crass or promotional, so I probably won’t be writing anything about it.
There is another important connection between LLMs and management: the use of LLMs is itself a management task. The way I often describe them to people is that it's as if you have 100 green junior employees working for you all of a sudden. What would you do? How would you make use of their significant capabilities without wasting your own time. This is fundamentally a management question more than anything else.
When I read Henry's post this morning, my literal first thought was "I wonder if Dan has seen this yet, it seems right up his alley..."
I think I largely agree with his thoughts on the value of llm summarization as a managerial technology (and the caveats about their limitations). I think one potential risk is about incentives on both the data generation side and the assumptions about the quality/recency of the data being summarized. For instance, at my current job, there was a lot of executive enthusiasm about making it possible for sales/customer support execs to "chat with our knowledge base/documentation" in order to better respond to customer questions. Even setting aside hallucination risks, my first reaction as one of the scientists responsible for core product features was "our shipping timeline has been under such pressure for years that the literal last priority for engineering/DS has been to maintain the wiki/knowledge base - good luck to anyone treating that as a resource to answer customer questions." This is just classic garbage-in-garbage-out.
The second risk seems like the degree to which LLMs can remove friction in *generating* new "information." If it's easier to summarize large volumes of text, there could be greater perceived value/importance placed on producing large volumes of text. But if employees are producing these reports by just starting with five bullet points and then asking ChatGPT or equiv to expand them into a two (or three, or four?) page memo, (1) this is a hugely inefficient representation, as the only real information was in the five bullet points (unless we assume the LLM is going to cleverly synthesize some linkages between these points based on its compressed representation of a ton of other data), and in the game of expanding and then later summarizing, we've ultimately added a bunch of noise to the signal. Given the state of the internet these days, it seems like the arms race of "algorithmic slop generation" vs "algorithmic slop detection" generally favors the generator, not the discriminator, so it's not clear to me that we will end up with a clearer picture of the state of an organization when this technology proliferates without very clear standards/practices and incentive alignment on what generates the input data at lower levels of a complex system including lots of people with incentives to be lazy.