Ben Hammersley‘s new (very early) startup caught my eye. It’s a writing interface called Agathonic.
Watch this short video of the Agothonic functional demo.
Here’s what happens, if you didn’t watch the video:
- Hammersley types into a text editor:
My name is Ben Hammersley and I live in New York
(along with some other prose)
- There’s a separate text box at the bottom of the screen labeled “Ask Agathonic questions.” Into this he types
When was Ben born?
(1:50)
- The system replies:
I think the answer is: 3 April 1978
(1:58)
(Hammersley has a background as a journalist so it’s worth seeing this prototype through that lens.)
What’s going on?
This tweet thread unpacks it:
[It’s] based on an NLP [natural language] interpretation of the corpus of text it builds in the background from its current set of sources. … Today it’s based off Wikidata but there are more sources to come.
So the human writes. The AI understands and builds its own mental model of what’s in the text. Based on that model, it expands its knowledge from available data sources. Then the human can have a conversation about what the AI knows.
This is a great interface.
What I like about it most is that the loop matches the creative process.
You write for a bit. Maybe finish the section, maybe get stuck. Then you go away and think, or you leaf through a book, or you do some research, or you talk through the article with your editor or a friend or another sounding board. And finally you figure out what to say, so you start writing again.
That bit in the middle is conversational. Having an AI assistant who learns from your text is a good addition. The conversational interface is perfect.
As a counterpoint, back in September I was experimenting with GPT-3, the startlingly human AI text generator. My takeaway at the time: GPT-3 is capable of original, creative ideas.
In the beta interface, GPT-3 is presented as smart autocomplete. The human writes words, then the AI tries to pick up where the text leaves off. (Applications built on top of GPT-3 can add any interface they want.)
But autocomplete makes me uncomfortable.
Autocomplete carries with it a certain kind of weight and inevitability. Red squiggle underlines to say a word is misspelt, the word count at the bottom of the window, autocompleting a word so you don’t have to type it… these feel like facts. Or if not facts then fact-adjacent.
But GPT-3, when I used it, was more like having a creative sparring partner.
What GPT-3 creates aren’t autocompletes, but instead suggestions that bump you out of your groove and take you in a new direction. It prompts, inspires.
The ideal interface for GPT-3 would be an assistant.
An assistant? Like Clippy?
Clippy was the Microsoft Office Assistant released back in 1996, now known mostly through memes. It would hang out in the corner of the screen and chip in: It looks like you’re writing a letter. Would you like help?
Etc.
People hated it.
But Clippy arrived before we had a shared understanding of software designed for realtime collaboration. It’s time to bring the assistant back.
Look: in 2020, we’re comfortable with shared Google Docs, and writing while other people are highlighting text in the same document and leaving comments. We gather together in Figma documents, checking out designs and commenting while we see a little crowd of cursors charge around. We collaborate and hang out in software built for groups and teams first, like WhatsApp and Slack.
So imagine this: you have a text editor, and your team is there too. Your colleagues are making suggestions, answering questions, filling in gaps, and being sounding boards.
But one of the team is an AI. And they appear not as a special interface element like a hovering window or a special sidebar or a squiggly underline, but in comments, chats, and suggested edits, alongside everyone else.
I think that would feel truly interactive and collaborative, and it opens the door to different styles of assistant: ones that provide creative prompts, ones that have the facts and figures at their fingertips, ones that are brilliant at wordsmithing the prose for different audiences.
The ideal interface to AIs is the team.
Ben Hammersley‘s new (very early) startup caught my eye. It’s a writing interface called Agathonic.
Watch this short video of the Agothonic functional demo.
Here’s what happens, if you didn’t watch the video:
(Hammersley has a background as a journalist so it’s worth seeing this prototype through that lens.)
What’s going on?
This tweet thread unpacks it:
So the human writes. The AI understands and builds its own mental model of what’s in the text. Based on that model, it expands its knowledge from available data sources. Then the human can have a conversation about what the AI knows.
This is a great interface.
What I like about it most is that the loop matches the creative process.
You write for a bit. Maybe finish the section, maybe get stuck. Then you go away and think, or you leaf through a book, or you do some research, or you talk through the article with your editor or a friend or another sounding board. And finally you figure out what to say, so you start writing again.
That bit in the middle is conversational. Having an AI assistant who learns from your text is a good addition. The conversational interface is perfect.
As a counterpoint, back in September I was experimenting with GPT-3, the startlingly human AI text generator. My takeaway at the time:
In the beta interface, GPT-3 is presented as smart autocomplete. The human writes words, then the AI tries to pick up where the text leaves off. (Applications built on top of GPT-3 can add any interface they want.)
But autocomplete makes me uncomfortable.
Autocomplete carries with it a certain kind of weight and inevitability. Red squiggle underlines to say a word is misspelt, the word count at the bottom of the window, autocompleting a word so you don’t have to type it… these feel like facts. Or if not facts then fact-adjacent.
But GPT-3, when I used it, was more like having a creative sparring partner.
What GPT-3 creates aren’t autocompletes, but instead suggestions that bump you out of your groove and take you in a new direction. It prompts, inspires.
The ideal interface for GPT-3 would be an assistant.
An assistant? Like Clippy?
Clippy was the Microsoft Office Assistant released back in 1996, now known mostly through memes. It would hang out in the corner of the screen and chip in:
Etc.
People hated it.
But Clippy arrived before we had a shared understanding of software designed for realtime collaboration. It’s time to bring the assistant back.
Look: in 2020, we’re comfortable with shared Google Docs, and writing while other people are highlighting text in the same document and leaving comments. We gather together in Figma documents, checking out designs and commenting while we see a little crowd of cursors charge around. We collaborate and hang out in software built for groups and teams first, like WhatsApp and Slack.
So imagine this: you have a text editor, and your team is there too. Your colleagues are making suggestions, answering questions, filling in gaps, and being sounding boards.
But one of the team is an AI. And they appear not as a special interface element like a hovering window or a special sidebar or a squiggly underline, but in comments, chats, and suggested edits, alongside everyone else.
I think that would feel truly interactive and collaborative, and it opens the door to different styles of assistant: ones that provide creative prompts, ones that have the facts and figures at their fingertips, ones that are brilliant at wordsmithing the prose for different audiences.
The ideal interface to AIs is the team.