Mmmm frosting anttwo.github.io/frosting/ [Holodeck on the Quest is a-coming with the rate of current progress ...🚂 🚂 🚂 ] [Also love the examples! Makes me smile.]

“These classes are a waste of time. They either require too much coursework or are simply too boring,” on.ft.com/3zverH5 [Sometimes I love the natural human attitude - it’s often the reality when it comes to forced ideology.]

Has any startup figured out how to make coding w/ LLMs cope w/ directory/project complexity & interconnectivity? It seems possible (having a DB of functions/classes, directory structure>txt>prompt context, using doc strings etc.). It's been 1 yr & haven't seen any1 nail it.

Ah the joys of deregulated bus services. I feel the headteacher’s pain - parents are blaming the school when they should be blaming Thatcher + their own desire to pay less tax + capitalism (no one wants to run busy loss making routes full of teenagers) + underpaying drivers.

LLMs teach us that most of the answers to our problems are large-scale actions across many actors rather than just more words.

Liking this daily tip from Anthropic's Claude: Embrace a paradox: Intentionally hold two contradictory ideas in your mind about a situation you face today. For example, "This task is both important and unimportant." Sit with that contradiction without trying to resolve it.

Nice pattern for daily mental exercise: * Write 750+ words a day (new.750words.com) * Get your writing & paste into a Claude project for daily review * Get Claude to review & critique + provide daily actions * Summarise + add to project

VLM-Floorplan - arxiv.org/abs/2409.12842 - sites.google.com/view/vlm-floor… [Love xLMs - swiss army knives - a day's worth of hacking and you can get a navigation engine that has previously taken a PhD project (& not really worked)]

Using Generative AI to Enable Robots to Reason and Act with ReMEmbR | NVIDIA Technical Blog developer.nvidia.com/blog/using-gen… [More and more stuff like this is coming into my timelines - exciting! - transformers are great because it's easy to iterate big improvements with external LxMs]

LLMs show a clear drop off with complexity, but you can counter that with chain-of-thought: breaking down into sub-problems, solving individually, then integrating (e.g. looking at room navigation). Interesting underlying learnings about problem solving in general.

Miranda July Called Before Congress To Explain Exactly What Her Whole Thing Is - The Onion theonion.com/miranda-july-c… [Just finished All Fours, which was very enjoyable, so feels time to repost this.]

Network Rail: Twenty railway stations affected by cyber attack - bbc.co.uk/news/articles/… [Funny how we’ve normalised & minimise these attacks. Quasi-state backed. On screen “European terror attack” - sow fear/panic, classic psych-ops. Or a launch of a new Adam Curtis series.]

The end of the popular politician on.ft.com/3zuX6ya [The problem, dear reader, is us and our demands as filtered through the global desire kaleidoscope of the Internet.]

LLMs and LVMs are going to go like all technology aren’t they? Splitting edge/central servers with low/high capacity functions. It’ll just be the next cycle of the same process, hopefully with less CORBA & Java.

Only six businesses sign up to customs sites in UK freeports on.ft.com/3BjAzF2 [How did that flagship Tory policy turn out? 💩]

China’s accelerating green transition on.ft.com/4eeo9Na [When China does something it really does something rather than a load of empty pronouncements. Self interest & being green can align?]

The Empusium — Olga Tokarczuk’s clever, carnivalesque homage to Thomas Mann on.ft.com/3XItZzz [Never a dud from Olga Tokarczuk - the kind of playful grown up literary novelist tackling big themes deftly I love. This goes on the post-Jacob reading list.]

Some AI regulations create more headaches than they solve... Ban on emotion & face recognition means models can't react to visual cues from users. Meaning all solutions just ignore user emotion. Which just becomes Victorian repression all over again.