
/sesh/null
656 posts

/sesh/null
@nerdsane
VP, data-and-ai@datadoghq | peripatetic | minimalist { engineer | athlete | artist } | I opinions-my-own



Gpt 5.6 Sol for UI expands on gpt 5.5's preference for chunky designs and bold fonts. It especially now seems to prefer alternating background colors to section landing pages, and is overall heavily inspired by graphic design tradition (asymmetrical grids, bold typography, color blocking). I'd say this is a rather unique feature gpt has compared to other models.




Continuing my mech-interp experiments with the Yume world generation model. Trying out activation patching. I recorded the activations of a world generated from Monet’s poppy field (the full internal snapshot at one of the middle layers) and injected it into the same layer of a New York Manhattanhenge world mid-generation. A really cool effect, as if two memories are fighting over the same territory. This is where mechanistic interpretability meets art.





There is another idea that continuously return in my head recently: we should not stop writing code by hand. But this code will be like poetry. Small programs that are outstanding either because of style or new ideas introduced. Code you may print and frame on your wall.


Something I told 14 yo: People are going to stop reading books. I wish this wasn't so, but I fear it is. The silver lining in this cloud is that if you're one of the few people who still read, you'll have a huge advantage over everyone else.

A great write up. I’ve been sensing that Qwen 3.6 35B has a lot of potential for writing (and I actually think front end design too). This result is super encouraging and bring GRPO the top of the next steps in my writing styles/VOICE.md explorations. x.com/arni0x9053/sta…


In the age when everything sounds like Claude, I have been looking into the concept of VOICE.md (like DESIGN.md but for writing) more and more. I majored in philology and linguistics "back in the day" (Mandarin and English). I operate most of the time in a language (English) I learned later in adolescence which means I inherently rely especially heavily on pattern matching in it - a late-acquired language is less deeply entrenched than a childhood one, and thus more susceptible to whatever dominates recent input. I also work solo most of the time which means the ratio of me talking to Claude/gpt is high, and that is a dangerous combo. So paying attention to style and writing is at the top of my mind. Some might even say it is "load-bearing" (jk jk). How do you find your voice and don't lose it? How do you extract and define it? What should be included in the profile and VOICE.md? How do you get an LLM to actually produce prose in your style? We've been exploring the style adherence part with Fable as groundwork for writing styles in katagami.ai (which I will be launching soon alongside design languages), so far we're landing on this: Take 10+ samples of human writing (corpus) -> extract its measurable habits: sentence rhythm and its spread, punctuation rates, function-word profile, character-level texture, vocabulary richness -> produce VOICE.md (containing never-rules, five verbatim gold samples, signature vocabulary, the measured fingerprint, and the machine contract - every numeric limit computed from the corpus) -> ask an LLM to output text in this style -> run deterministic stylometry checks on the output (their thresholds come from the author's own out-of-sample variation, so a text only fails by deviating from the author more than the author deviates from themselves) + an LLM coach giving directional adherence feedback -> LLM-writer produces corrected output that is closer to the author's own measured habits (sentence rhythm, word frequencies, punctuation). Measured by: - a word-frequency fingerprint - an openly released authorship-embedding model (STAR) - human (my) judgement (the first two are validated to tell authors apart; whether their distance scores mean a text reads more like the voice is what I'm still validating) Here's some observations made so far:



