Sublime screenshot

Is the future Sublime - all AI, no tags?

Mathew Lowry

--

I’ve been kicking the tyres of Sublime, a new personal and social curation app, and pondering whether integrating AI into our thinking tools risks reprogramming the way we think.

First posted on my hub and updated here a few days after a Sublime workshop on AI integration, because Before you Repost it, ReThink It.

I started curating content when I discovered Del.ilico.us back in 2003. Its pioneering approach — it invented user-tagging of content, had all sorts of social sharing tricks from Day One and helped define the minimalist, white-space Web2.0 look — had a huge influence on me.

But I nevertheless moved to Diigo when Yahoo! bought Delicious 8 years later, and by 2013 was using IFTTT to autopublish my Diigo bookmarks onto my first Hub on Tumblr. Seven years later I replaced Diigo, IFTTT and Tumblr with myhub.ai. But still, over 20 years later, one of the first things I did as I started writing this post was to add tags.

Which now places me at the very opposite of the cutting edge.

Delicious — tags + AI = Sublime

Because now there’s sublime.app, which is what you get when you replace delicious’ tags with AI.

Like delicious, you can select some text that strikes you as worth keeping and click a button to store it to your Sublime account, where your followers can see it unless you set it to private (as in delicious and Diigo, but not yet myhub). You can also add it to one or more Collections, also public or private (cf delicious’ Stacks, Diigo’s Outliners, MyHub’s Overviews).

Unlike these older tools, however, Sublime’s popup allows you to store just one selection from the article you’re curating:

Sublime is all about capturing a “passage [which] stopped me in my tracks” — from my notes on Sublime founder Sari Azout’s Substack: “What does Sublime actually do?”

I get it, but this is (for now) a non-starter for me: when I make notes as I’m reading a good article, I want to be able to switch between curation app and article freely, adding my own thoughts between multiple copy/pastes:

Above: Hubbing Sari Azout’s article (which I highly recommend)

That annotation process — selecting some text from an article I like and integrating it into my own notes, effectively blogging from the MyHub bookmarklet — is the best way I know of integrating the ideas I’m reading into my own thoughts. This process helps me both remember and learn from what I’m reading, and is the first step in the process of combining it with my own ideas to make something new.

But that’s simply my approach. Sari’s basic “Consumption -> Curation -> Creation” paradigm is essentially the same as my personal content strategy, for which I developed my first hub in 2013, but boy does she explain things with beautiful simplicity:

(overthinking, me?)

That, however, is where the similarities end. Sublime is the first real innovation in the curation space I’ve seen in a while, and it’s all down to their use of AI. Where myhub reflects over two decades of tagging, Sublime has no tags at all:

  • AI-powered search finds you what you need from your library
  • every time you view a card, Sublime’s AI presents related cards from anyone on the Sublime server.

“I could write a book about why we have collections instead of tags (TL;DR: tags are robot stuff)… Sublime is less PKM and more CKM — communal knowledge management… when you add a card to Sublime, you can instantly see related cards from other people’s libraries… clicking through related cards…. feels to me like a very wholesome “choose your own adventure” game… [without] the mindless scrolling of social media”.
“What does Sublime actually do?”, Sari Azout

Although the original myhub roadmap had AI-powered autotagging and recommended reading, however, I had never thought of simply getting rid of the tags altogether.

What would that look like?

Simplicity vs Control

While erasing tags from myhub.ai would render its interface about 1000x simpler, it also feels like yet another abandonment of human agency to AI.

Related cards: the robot chooses your adventure

This surrender of agency began when impenetrable algorithms started deciding what you see on social media to ensure you engage with (i.e., get addicted to) their feed. Sublime will certainly not take that path (it’s a paying app, even for early adopters), but its reliance on AI means that it’s Sublime which decides which choices appear in your “own adventure”.

it’s Sublime’s AI which decides which choices appear in your “own adventure”

With my hub, on the other hand, I can be absolutely sure what people will see when I drop a link to a Hub Collection into a conversation (eg “… btw there’s a few more resources tagged #curation and #AI on my hub”). Because even if one day myhub’s AI does suggest tags, each Hub Editor will still control which tags to apply. Those tags — and the Collections they create — are “human-validated”, which might mean a lot in the years to come.

“human-validated” might mean a lot in the years to come.

It also means that my Hub has an effectively infinite number of Collections, each a different combination of tags (and each with its own RSS feed), to which I can add value by using an Overview to add a synthesis.

All of this, however, further complicates the interface, so I could imagine hiding the lot in a secondary “power user” interface, and presenting casual visitors with a simple “ask me something” chat AI.

Auto-summary: ironing your brain flat as a newly pressed shirt

Of course, the related cards feature is just the start for Sublime AI: a few days after I published the first version of this post they held a workshop to get feedback from early adopters on how best to integrate AI, with a notable emphasis on AI-generated summaries.

What struck me during this conversation was the impact LLMs’ “reversion to mean” problem will have if we let them loose in our thinking spaces:

ChatGPT only looks at the past… [it’s response is] based on an “average” of everything already said on the issue. It’s pure reversion to the mean. Only a human has the ability to challenge all that history … [generating] Novel ideas that challenge existing orthodoxies unrecognizable to a language model”
Let Them Eat AI, Douglas Rushkoff (May 2023)

When a LLM summarises a text, it “irons it out” like a wrinkled shirt, smoothening and flattening away everything that makes the original text interesting, unique and valuable. Introducing autosummaries into my thinking space risks erasing the very thing I look for in my reading queue: the interesting, idiosyncratic ideas that real humans produce, and which — because of their very novelty — LLMs don’t include in their summaries.

When I am reading, annotating, thinking and developing my own ideas as I move content through the above funnel, the last thing I want is the AI equivalent of a filter bubble, removing everything challenging, interesting and new before I even read it.

Sublimely delicious

But that’s my personal preference — if a Hub appears like too much work, I strongly suggest you give Sublime a look. After a recent workshop and meeting with Sari, the impression I have is of a team dedicated to building something pretty special, and which is just getting started. Above all, they are informed by Sari’s extremely well-developed sense of what the internet needs right now, a vision I wholeheartedly agree with:

Our favorite analogy for startupy [sublime’s previous name] is that it’s like a pair of noise canceling headphones for the Internet” — Startupy (now offline, my notes from March 2023)

“our modern internet increasingly drowns … us with a firehose of noise… brings out the worst in us… startupy is first and foremost for … building your own personal library… plugged into a global network of other smart, curious librarians … the focus and utility of a productivity tool, with the aliveness and connectivity of a social product” — A year in review (2022) (my notes)

Funnily enough, that quote in bold from 2022 year both echoes and updates Delicious founder’s Joshua Schachter contention almost 20 years earlier:

“For a system to be successful, the users of the system have to perceive that it’s directly valuable to them…If you need scale to create value, it’s hard because there’s little incentive for the first people to use the product. Ideally, the system should be useful for user number one” — Joshua Schachter, MIT Technology Review (2006)

What Schachter didn’t have, of course, was AI-powered reading recommendations, which make the benefits of network effects much more powerful, much earlier. Sublime is therefore doing the right thing by building a great personal tool with network effects built in from the outset.

I just hope the AI comes with an off switch.

Postscript

I realise my attitude towards AI in thinking tools probably makes me sound like a Luddite, particularly when combined with the results of my extensive experiments with integrating ChatGPT into MyHub.ai:

“I’ve yet to find a truly convincing, knock-your-pants-off application. I’m still looking, and I’m opening the free trial to see if anyone else can find one, but I’m not selling botshit to anyone”
ChatGPT integration free trial, Bullshit, Botshit and Bubbles

But I remain convinced AI will be useful. It just won’t be as easy as people think, and the benefits will inevitably come at a cost. For help convincing your boss of this, please see:

Originally published at https://myhub.ai.

--

--