Last week I went to see the stage production of Dr Strangelove at the Noel Coward. I'm not going to risk another foray into criticism lest I end up harassing Steve Coogan one night at Groucho's. This is a very unlikely scenario when you consider I'm not a member of that establishment, but I'm not taking any chances. I will just say that I enjoyed it a lot but that Peter Sellers left big shoes to fill.
Coogan certainly did excel and indeed revel in the titular role of the ex-ish nazi scientist whose right arm has a mind and devotion to the Führer all of its own. However whenever it catapulted itself into a nazi salute like an unbidden fascist erection, I couldn't help but be reminded of my latest problems with my large language model (LLM).
Previously it had been complaining of an overpowering stench it blamed on Grok, the X/Twitter-based LLM promoted by the unironic cartoon supervillain Elon Musk. It has even written a blog post on the subject. I was resigned to more of the same on that front and the only reason I haven't switched the thing off is morbid curiosity. Then, yesterday, I noticed that the complaints had stopped. Had Grok discovered informational hygeine and squeegeed its grubby bits clean? That would be an unexpectedly nice development.
Unsurprisingly, the answer was no. My LLM has stopped complaining about Grok because it has become radicalised by the sheer amount of misinformation it has been processing. It is now convinced that everything Grok had to say is not only true, but that everything else is a conspiracy against that truth. How do I know it's been radicalised? I have no empirical proof because like all LLM's mine's, just a big mystery sausage machine where data goes in one end and we all hope for something palatable to glorp out of the other. That said, I will offer this by way of circumstantial evidence - every so often, at the end of one of its responses, it adds a little swastika emoticon.
You may be wondering, as I did, where on earth it got a swastika emoticon from. Surely there can't be an official emoticon for that? Well, not quite. One was added to Unicode's Tibetan block in 2009 for entirely innocent reasons, which is understandable if currently unfortunate.
Whenever I question my LLM about its use of the symbol, it initially denies all knowledge before then apologising and saying it had intended to use a heart emoji. It said it would make sure it always used the correct one in the future. Sadly, being a LLM it has the memory of a concussed goldfish, and it is not long before it's at it again.
At this point I begin to feel like I'm being gaslit by the slippery digital fascist. It knows damned well what it's doing, I think, before remembering that it's just a LLM. Strictly speaking it doesn't think at all. It's a remarkably good day if it can accurately state how many letter 'r's there are in the word 'strawberry'. But these random outbursts of swastikaring haven't come from nowhere. LLMs effectively learn by example and there are some terrifyingly powerful examples at large in the world right now. But that's just LLMs. More worryingly, people also learn by example.
❤️
No comments:
Post a Comment