Every Company Needs a Cassandra

Every company has lots of people trying to be Cassandras. The problem isn't that no one is interested - it's that they don't know which 1 out of 10 of the folks complaining actually are worth listening to.

When I look back in my career, I'm having trouble finding the person who complained a lot who was actually right in the long run.[1] They vastly outnumber the ones worth listening to.

[1] Or rather, they were technically correct but whatever they were complaining about wasn't as big a deal as they made it out to be, and ignoring them was OK.

And that Cassandra will be ignored too. It's the role of "a cassandra" to be ignored, like it's one of society operating rules. People who tend to be good at being a cassandra learn very fast to stop voicing their opinions.

The way people become corporate Cassandras is by doing the hard work for a while and building a mental model that is more influenced by the grit of the real work and unfiltered customer feedback than by the group discussion.

You could train a model to just read the group discussion and say plausibly sounding things. But if its input is just the group discussion and, as the post suggests, git history and competitor web searches, you can disagree, but you can’t have the orthogonal mental model that produces the “you’re not even wrong” style responses that corporate Cassandras produce. You’d need an AI that actually does the work and maintains private opinions it doesn’t reveal to its prompters.

More importantly, this post seems to ignore the real role of politics. It’s true the bot itself doesn’t have a promotion to worry about. But if its goal is influence, then it will need to take into account the reactions and unpopularity of its ideas. Or if not, it will get itself ignored or turned off. Politics is just how decisions happen, not a problem that goes away!

> an ai is oddly well suited to this job. cassandra doesn’t want a promotion and doesn’t need the people in the channel to like her. disagreeing with the vp three times this month won’t show up in a performance review. she doesn’t have an ego invested in being right either. if she objects to a project in january and the data proves her wrong in march, she can just say so. being wrong should make her model of the company better.

This is true, in terms of personal risk, but it also misses something subtle I think I have observed in the last ~year or two. I have been a nag on issues for years to be ignored or dismissed without even really listening before an AI flags the exact same thing by the exact same person I've told this to a dozen times, and they act like it's some profound new discovery that must be acted on immediately.

I thought it was my delivery style, or maybe bad communication, but I've noticed this same thing with a handful of other people too, with myself entirely out of the loop.

We talked about it, and what I think it is, is a human ego/relationship thing - being told something is wrong or bad or pointing out 'negative' things requires the other party to concede something, and the mental effect is they feel they 'lost' something (the argument, political capital, ego, whatever) to the cassandra party. You can quietly and privately concede to an AI's criticism without feeling that way, and even better for some types, claim the idea as your own if you want, so it's much more receivable. The AI will also explain the issue with infinite patience and ego-coddling.

So many organizations need this. I've watched companies make $100M+ investments basically on group think and no one being willing to say it won't work.

The worst is when everyone privately acknowledges that the idea won't work but because leadership has committed everyone has to march on to the failure point.

I have yet to see one of these Cassandra's to be proven wrong too.

I am not sure how this is Cassandra. Cassandra didn't say bad things. Cassandra knew the future and know one believed her, making her helpless to see doom unfold. The author even mentions the myth, but Cassandra didn't point out uncomfortable truths. There are tons of complainers in companies, automating it with AI isn't going to do anything.

Yes, but sometimes forward motion, even in the wrong direction, is better than inaction (because forward motion reveals more information that you can use to course-correct). I think more companies, particularly large companies, fail from strategic inaction than from strategic mistakes.

I would pair this with Bezos's concept of one-way doors. If a decision is a two-way door, i.e., you can reverse the decision at low cost, then you don't need a Cassandra--just try it and see. It's only one-way doors that need a Cassandra.

ETA: This also reminds me of the Tenth-Man Doctrine, which I only learned from World War Z.

If people wanted to listen to Cassandras, or more broadly, people who don't have the power to make or stop changes but are voicing consent or considerations -- they would far more often, as that is very easy to do.

Corporate culture is especially nearly designed to never give Cassandras the time of day.

Why wouldn't people just write the agent off as 'broken'? Would this agent learn to keep most of it's responses to itself or end up actually broken/drifting as a result of having good advice ignored?

The grammar in this article is insulting. No sentence case, all lower case, no capitalization of acronyms. Made it very distracting and borderline unreadable. Do not write like this.

Completely distracted from whatever point the author is making.

the author made the assumption that these speaking machines are actually thinking machines with any sort of interiority and capability for novel thought, something these explicitly are not and cannot do.

Every Company Needs Sqlite.

If you implemented something this insufferable people would at best just move all the discussion to private channels where this thing couldn't respond to it or at worst just stop providing input at all.

No one wants to be lectured by a clanker when they're just trying to banter.

Jesus Christ. Let's replace humans with AI at this point.