https://t.co/rvw7uIvrqy

The Rone exhibit at Flinders St station was absolutely stunning - the imagery, space design, details, light, and sound were all impeccable

So control over user agents will be an interesting battle - the big players will be bigger targets for publisher litigation; server-side agents will be much easier to block. It'll be game on to win consumer opinion quickly, and smaller players have an advantage here

A massive thing to think through is the ecosystem impact - abstracting away developer control has echos of the incentive issues that lead to no-one buying into the semantic web, but this time the the user agent is in control

And so many fun challenges - how to introduce these capabilities while they're only useful part of the time - if you leave it to the user pressing a button, most will forget. If you prompt, you'll drive people bananas*

And that's what's so exciting here - people have been having ideas like the above for decades now, but the capability was just never there, yet now it all feels within reach

Now most of the UI you're dealing with is indirect - perhaps a dashboard of things to read and confirm - I have long believed that one of the "final UIs" is a series of prompts (agency and choice are important), and it feels like it's getting close

Then it's anothing small leap to start doing this proactively in combination with the personalization built up over time

And then it's a small leap from that request/response model to doing this for ongoing tasks - "get me tickets to Carly Rae Jepsen when available"

A set of powerful next steps is being able to act on your behalf (Action Transformers): - accessibility-style automation of the screen you are looking at (send email) - multi-step actions like "complete this purchase flow"

Then there's re-presenting apps - take content transformation and apply it to app UI, like https://t.co/VVqzAaEOmy but automatic and at scale - have your apps re-presented and reduced to the essentials you care about, customized to you

A first step is input assistance - form fields autocomplete or your keyboard helping you create better content. This can extend as far as text expansion, image creation (GIFs++), tone analysis, or bringing in some of the content tools to increase veracity (fact checkers)

Ok, so that's content, but a huge part of computing is apps - which you might say are just lists of content with actions attached

The next jump is going from that to ongoing monitoring - "I am interested in F1" / "let me know about new APIs in Unity" and summarizing the news for you. The relationship between publishers, ads, and your user agent will be complex, especially if its client side.

By now we have replaced content search and browsing - in may ways you can think of GPT4 as something that has pre-browsed the web and done all the above already. So you can skip even loading the page - "just give me answers and fall back to a webpage if I have to"

Then allowing queries on top - "what's a gluten free substitute for corn flour?", "explain this like you would to a ten year old", "what are possible counterpoints?" These are all queries GPT4 handles stunningly well today (it has transformed reading papers for me)

As a first step - content transformation - reader mode but good - turn text, video, audio, foreign languages into something better for you: "make this korean cooking video into a list of instructions and ingredients" (personally I'd love every recipe as gantt chart)

Some quick thoughts on how GPT/LLM/AI features could be added to browsers, phones, and operating systems, going from where we are today through to platforms-as-we-know-them-don't-exist:

ps you can't link to specific nodes because then it wouldn't be a roguelike

i made this weird thing the haikuniverse awaits not sure why i did https://t.co/oxGp1l58cZ