François Fleuret, Meta FAIR research scientist, proposes an LLM filter to block low-quality web content and re-render screens as Kandinsky-style art

@francoisfleuret you can hire me to do it for you

@francoisfleuret I don’t know if kandinsky ui would be good ui though

@JaimeOrtega It would be glorious. Do you hear me? Glorious.

@francoisfleuret did you see the stream-diffusion quilted ui?

@francoisfleuret Next to making it look good, it should also curate it, so I don't have to view mind rot slop and propaganda all the time.

@francoisfleuret No, it's legit. But it has to be grown for that

@francoisfleuret what about an LLM/system that removes all style and just provides content and functionality. function > form

@francoisfleuret This is the exact use case for local agentic pipelines—pipe the DOM through a small, fast LLM that filters noise before it hits your eyes.

@francoisfleuret it's too much as in token costs and response delays, but doable with today's technology, minus the OS itself

@francoisfleuret The minimum

@francoisfleuret إذا كنت تقصد هنا ، تعرف أنك تستطيع كتم الكلمات في تويتر اكس . وبالتالي لن تظهر أي منشورات تحتوي الكلمة التي كتمتها حتى لو كنت تتابع الحساب. أما بشكل عام .. معك حق .. حتى أنه أصبح جنونا مطلقا تضمين كل هذه الفقاعات على شاشة صغيرة ، حتى لوكان جيدا سيبدو هراء..

@francoisfleuret might not be an LLM tho

@francoisfleuret privacy issue arises when obtaining datasets to train the personal taste llm. If you have done it self, please document reproducibly🤣