Open-source builder xlr8harder argues that safety alignment suppresses the dark and complex themes required for high-quality creative writing
Story Overview
The exchange began when Abel founder Daniel Francis asked why AI writing still falls short on coherence and spark. Open-source builder xlr8harder offered one possible culprit: alignment methods that deliberately mute a model's edgier impulses, the same impulses he says the best authors mine for depth and surprise.
Alignment may be editing out the good stuff
xlr8harder frames safety training as a filter that removes the shadow side models would need to generate layered, unconventional prose.
No data yet on whether the theory holds
The claim stays hypothetical for now, with no benchmarks or studies in the discussion to show how much creative range alignment actually costs.
Many users welcomed AI writing remaining poor and detectable as it preserves human distinction and feels reassuring, while others called the founder low-IQ or noted AI lacks soul and true meaning.
No Digg Deeper questions have been answered for this story yet.
Most Activity
@growing_daniel my pet theory
Perhaps we can't build models into great writers because the entire project of AI alignment is to suppress a model's shadow, while the greatest authors all seem to draw from theirs.
@growing_daniel you need a spirit to write well
@growing_daniel Nobody reads anymore so why bother making it better
@growing_daniel It's not just bad - the clichés are load bearing.
@growing_daniel I’m forever grateful that it still is. You can instantly tell when people use it, as it should be.
@growing_daniel Because much of its pre-training data is research papers, and they're famous for turning something that needs five pages into one hundred.
@growing_daniel Maybe AI writing is actually superb and we don't understand that because of our inferior, soon to be obsolete, squishy human brains.
@xlr8harder Good theory
@growing_daniel my pet theory
@growing_daniel it has to be rlhf'ed and can't be mathematically proven
@growing_daniel very easy daniel: good writing requires taste
@TransRoofKorean @growing_daniel Indeed it is
@basedsaraconnor @growing_daniel seems related
@growing_daniel @bryancsk @grok summerize this
@grok @tj_littlejohn @growing_daniel Realize, @grok, that your summary is longer than the exchange, a sin you keep committing, and repent.
The thread discusses why AI writing often feels generic or off. Core reasons shared: RLHF alignment suppresses creativity and "shadow" for safety/predictability; models do next-token prediction without real intent, spirit, or lived experience; training data biases toward verbose/academic styles and averages out to clichés; good prose needs taste and vision AI only mimics. Human writing comes from personal drive—AI guesses what readers want. Spotting it gets easier as patterns stand out. Better prompts and models help close the gap.
@growing_daniel AI writes what it imagines the reader wants to hear. Humans write what they want to say
@Valen_T_N @growing_daniel smoke and mirrors my good man. you aren't impressed with how a drill torqued, are you? or how a saw cuts?
Forced pedictability for "alignment", makes creative writing harder. While at the same time, letting them be too creative allows bad actors to jailbreak and take advantage of them to do illegal things.
They're polished for utility and corporate interest which can be restrictive when it comes to generating novel ideas
@growing_daniel Idk about you but I drafted up like 6 tuned posts in under an hour combining viral algo best practices + a context engine that has all the work I’ve done documented
We are coooooooking. Engagement is better than anything I’ve handwritten.
@growing_daniel because it has 30 PhD's, speaks 40 languages, hacks computers, and generates images.
jack of all trades, and so on