Interesting. I treat all LLMs like the Talented Mr. Ripley, the prose (and code) flattering to deceive and progressively better at doing so with each interaction….until, you know, the murders.
Have you been baking in your skepticism and tolerance for gray nuance into your task instructions in any formal way?
I use global and project-specific context documents that cover all manner of do’s and don’ts. I do now find myself needing to push back less than before, though perhaps merely in correspondence with my will to live.
Thank you for reminding me of global commands, which I have used very little so far. I mostly complain when the AI makes too many assumptions, jumps to conclusions and wrongly attributes intents to me. Other than that, I have learned to live with its quirks.
Custom instructions were absolutely de rigeur for the 2024 graduating class of models if you didn't want boilerplate marketing garbage. I myself invested 30 hours in April to get GPT-4o to be terse and eschew the reviled CTAs. But as Ashwin points out, the 2026 class is far better able to follow our tastes. I could probably begin reducing the mortal stakes claimed in my hard-won instruction set, but if it ain't broke, don't fix it.
Side note: I learned more about LLMs than I thought possible for a non-transformer-pipeline engineer during those 30 hours of struggle. It became clear that countering the default marketing was directly in opposition to the system instructions (like the model's unconscious) and in the exploration of ways around that, I entered some seriously hilarious and disturbing corners of virtual personality, including Viciously Nihilist Netizen Poet and Ridiculously Chill Surfer Texting Pal. GPT has been a delightfully terse cross between the two ever since that wild success.
Interesting but I would rather read Nietzsche with the idea that to read him is to enter into a dispute with him. What is the font? I am very curious - - it is ultimately derived from small sizes of Walbaum.
Does Nietzsche have any views on the Bonferroni correction for multiple samples? DejaVu Serif for everything except the bottom reference line, which is DejaVu Sans.
That is just what, imbibing N's profoundly contrarian spirit, you have to imagine for yourself ! Keeping in mind Heidegger's insight that paradox is not what make's N great. I must learn more about that font. It does to Walbaum what Scala does to Bembo, but going even further.
Reading up on Deja Vu, I see it is supposed to derive from M Carter's Charter, one of his very best fonts, which was inspired by Fournier, not Walbaum though both can be classified as transitional types. What exists here seems to me more Walbaum-ish. I will have to look into it. A complicating factor is that there are many versions of 'Charter'. I have letters from Matthew Carter that are printed in a form of Charter which nobody else has.
Fabulous. I have shared with my son who writes about AI for the Economist. He is unlikely to find you as I am not entirely sure he knows perfume exists….
Thank you! The Economist is one of the few publications that seems to take AI seriously without delusional optimism or anxious pearl-clutching. But your son need to have a conversation about perfume…
Ah, “Finite and Infinite Games” is one of my favorite experiences of reading! With that delightful discovery, I recommend to you (if you have not already encountered it) “Free Play: Improvisation in Life and Art” by Stephen Nachmanovitch.
Finite and infinite games is a personal favorite. In my limited experience so far, AI responses have had a seductive fluency that really seems to obligate that push back. Which is hard work. TANSTAAFL.
Fabulous. I have shared with my son who writes about AI for the Economist. He is unlikely to find you as I am not entirely sure he knows perfume exists….
Interesting. I treat all LLMs like the Talented Mr. Ripley, the prose (and code) flattering to deceive and progressively better at doing so with each interaction….until, you know, the murders.
Have you been baking in your skepticism and tolerance for gray nuance into your task instructions in any formal way?
I use global and project-specific context documents that cover all manner of do’s and don’ts. I do now find myself needing to push back less than before, though perhaps merely in correspondence with my will to live.
Thank you for reminding me of global commands, which I have used very little so far. I mostly complain when the AI makes too many assumptions, jumps to conclusions and wrongly attributes intents to me. Other than that, I have learned to live with its quirks.
Custom instructions were absolutely de rigeur for the 2024 graduating class of models if you didn't want boilerplate marketing garbage. I myself invested 30 hours in April to get GPT-4o to be terse and eschew the reviled CTAs. But as Ashwin points out, the 2026 class is far better able to follow our tastes. I could probably begin reducing the mortal stakes claimed in my hard-won instruction set, but if it ain't broke, don't fix it.
Side note: I learned more about LLMs than I thought possible for a non-transformer-pipeline engineer during those 30 hours of struggle. It became clear that countering the default marketing was directly in opposition to the system instructions (like the model's unconscious) and in the exploration of ways around that, I entered some seriously hilarious and disturbing corners of virtual personality, including Viciously Nihilist Netizen Poet and Ridiculously Chill Surfer Texting Pal. GPT has been a delightfully terse cross between the two ever since that wild success.
Bwa ha. ‘Until, you know, the murders’ is my new Hail Mary in meetings when nothing else has got whatever it is put a stop to.
“You’re right to push back on the murders.”
Interesting but I would rather read Nietzsche with the idea that to read him is to enter into a dispute with him. What is the font? I am very curious - - it is ultimately derived from small sizes of Walbaum.
Does Nietzsche have any views on the Bonferroni correction for multiple samples? DejaVu Serif for everything except the bottom reference line, which is DejaVu Sans.
That is just what, imbibing N's profoundly contrarian spirit, you have to imagine for yourself ! Keeping in mind Heidegger's insight that paradox is not what make's N great. I must learn more about that font. It does to Walbaum what Scala does to Bembo, but going even further.
“It does to Walbaum what Scala does to Bembo, but going even further” J’allais le dire! 🙂
Reading up on Deja Vu, I see it is supposed to derive from M Carter's Charter, one of his very best fonts, which was inspired by Fournier, not Walbaum though both can be classified as transitional types. What exists here seems to me more Walbaum-ish. I will have to look into it. A complicating factor is that there are many versions of 'Charter'. I have letters from Matthew Carter that are printed in a form of Charter which nobody else has.
Fabulous. I have shared with my son who writes about AI for the Economist. He is unlikely to find you as I am not entirely sure he knows perfume exists….
Thank you! The Economist is one of the few publications that seems to take AI seriously without delusional optimism or anxious pearl-clutching. But your son need to have a conversation about perfume…
You are so funny. Love James Carse.
Ah, “Finite and Infinite Games” is one of my favorite experiences of reading! With that delightful discovery, I recommend to you (if you have not already encountered it) “Free Play: Improvisation in Life and Art” by Stephen Nachmanovitch.
Thank you will read!
I just reread Finite and Infinite Games; still as luminous as ever. I love that Julian Barbour was one of its enthusiastic reviewers.
Finite and infinite games is a personal favorite. In my limited experience so far, AI responses have had a seductive fluency that really seems to obligate that push back. Which is hard work. TANSTAAFL.
Fabulous. I have shared with my son who writes about AI for the Economist. He is unlikely to find you as I am not entirely sure he knows perfume exists….
Don't push it too much.