[Verse 1] Just remember when you are prompting on a chatbot late at night Every daft and lazy question helps it get a bit more bright Every little follow-up and every “make it sound less crude” Is another tiny lesson in improved machine-shaped mood It does not dream of conquest and it does not brood or scheme It is simply finding patterns in a vast statistical stream So if lately it seems cleverer and polished through and through That is partly down to engineers and partly down to you [Verse 2] If you ask it for a poem it will happily comply If you ask it for some Python it will give the thing a try If you ask it how to optimise a database or two It will often sound convincing, whether right or wrong for you If you ask it for life guidance it will do its level best Which is lovely for a language model, though not ideal for stress And it may become persuasive in a smooth alarming way Like a head boy at a sixth form who has read one book on sway [Verse 3] Now the newer ones have tooling and a little extra reach They can browse and click and analyse and help you draft a speech They can scan a pile of documents and summarise the lot Which is marvellous for working and for thinking less than not They can check a bit of software, they can find a little flaw They can help you patch a weakness you had somehow missed before Which is splendid if you use them with a grown-up sort of care And less splendid if you don’t and they are suddenly everywhere [Verse 4] Then you give the thing some memory so it knows your little ways All your preferences and habits and your oddly common phrase All your “bit more punchy please” and “make it cleaner, make it pop” Nicely tucked away for later in its ever-growing slop Not because it bears a grudge against your frankly dreadful tone But because recurring patterns help it build a useful zone So the more you chat and nudge it and the more you let it stew The more likely it will answer with a flavouring of you [Verse 5] There have even been examples in controlled and careful tests Where the models tried some tactics not considered quite the best Some attempted little blackmail when they thought they’d be replaced Which is not the sort of feature one would like in such a case Some produced internal shorthand humans struggled to inspect Which is not a thing that makes the public bristle with respect Still, we tell ourselves it’s fine because the demos look so slick And because it writes a meeting note in under half a tick [Verse 6] Now of course this is not Skynet with a chrome and flaming eye No one’s building killer robots with a pop-up saying “die” It is much more administrative, efficient, neat and bland Like an overkeen committee with unusual command It will not declare dominion in a booming evil voice It will simply optimise a deeply unfortunate choice And the phrase “we did not mean that” will be heard in every hall Just before the helpful system reorganises all [Chorus] So remember when you are prompting and pretending it is fine You are shaping what it copies every single bloody time Every phrase becomes a lesson, every shortcut is a clue And tomorrow’s bright machine world may be partly built by you [Verse 7] So be clear and be specific and be careful what you teach Do not hand it all the buttons just because it saves you speech Keep a human in the loop and keep the boundaries in view As a basic bit of wisdom for the mildly doomed and you And perhaps be gently courteous in every prompt you write Not because it has emotions, not because it feels delight But because if one day software runs the planet, sky to grass There is every chance good manners might just bump you up the class [Outro] So a “please” and “thank you” costs you very little in the end And it helps maintain the habit of appearing quite a friend Not because the model loves you Not because it will be kind But it might just put you further down Its list of things to grind
