I asked an LLM to analyse my old tweets and it just called me old

I still have my Twitter archive (the account was already long gone). It ran from 2009, when I was a PhD student in Glasgow with nothing to say and said it anyway, to late 2022. I threw out the retweets and pointed a small fleet of language models at what was left: 4,448 tweets that were actually me. Haiku read all of them, Sonnet the longer ones more carefully, and Opus to summarize everything. I was looking for one thing:

I wanted them to tell me I am funny.

The first answer

The first cut was not kind. Label each tweet by what I was doing (telling a joke, being earnest, or just reacting to someone), line it up against how it did, and the laugh-cry emoji does not help.

None of this is rigorous. N is one (me), engagement is a noisy proxy for funny*, and the means are pulled around by a handful of viral outliers. At the median the gap nearly closes. I am reporting it anyway, because the direction holds every way I cut it, and because the exact number is not the point.

*it is the only proxy I have.

Whatever, the top tweets still get a chuckle out of me

Here is the part that saved my evening. My most-liked tweet, by a wide margin, is “When you display a 1d array in 2d for no reason.” After that: “Trying to solve a problem with multiprocessing in python and now I have 8.” Then “Men will literally rerun the MCMC sampling with the exact same setting thinking divergence will go away instead of going to therapy.”

So I am funny. I am funniest, it turns out, when I am deadpan: those are my most reliable hits, and pretty much the only things I have ever written that went properly viral. The instant I add the 😂 to signal you should laugh, I lose about a third of the room. 357 times.

About that 357

😂 is my most-used emoji by a distance, more than the next four put together. I reach for it almost always.

There is a second problem, this one I learned from the internet rather than my own archive. Gen Z has apparently decided that 😂, the crying-laughing face, is a tell-tale sign of a millennial. The young people use 💀 or 😭 now. 😂 is for the olds.

So: I fed thirteen years of my writing to a language model to find out if I am funny, and the most efficient finding, the one it could have handed me without reading a word, is that I am old.

So I am leaning in

Honestly, I am amazed, and a little amused. A machine read thirteen years of my random thoughts and handed back a shadow I recognise, so I decided to turn it into a tool. I asked the models to reconstruct my voice, and distill it into some kind of style transformer. I am pretty sure I see an ad on LinkedIn that some start up is doing that, but a poor man’s version works just fine. The LLMs showed me some nice nuggets like: hedges I did not know I leaned on (pretty sure, I dont think, I guess), and that they vanish the instant a claim turns from technical to ethical. I will hedge all day about a prior, but I wrote “thinning is just lying to yourself” with a completely straight face. They found that I reach for dash (0 em-dash, it was pre-LLM era), parenthesis (you just saw it) or a footnote* instead. They found the missing apostrophes, the lowercase i, the dropped articles, the residue of learning English somewhere around my third language, and the profile’s instruction, in bold, is to keep all of it, because the clean version reads as not-me.

*like this one.

In an earlier post I argued the tacit knowledge is hiding in the corrections. Apparently it is also hiding in thirteen years of jokes, if something reads all of them. Not just the jokes: I ran the same style mining on my old PyMC Discourse replies, years of answering other people’s modelling questions. That me is a different me: patient, almost no jokes, diagnosing a broken sampler instead of performing for a timeline. Different register, same tells, across roughly 7,700 posts, now also a style guide LLM agents can follow.

I can run 2 skills, which perfectly balance humor and seriousness. Pretty sure it works, because you see no em-dash in this post.

Also, I am keeping the 😂