difficult to find people who are being sincere, seeking coherence, and building collective knowledge in public.
I’m interested in enabling productive discourse and community building on at least some parts of the web.
semi-private spaces like newsletters and personal websites
retreat further into gatekept private chat apps like Slack, Discord, and WhatsApp.
express our ideas, with things we say taken in good faith and opportunities for real discussions.
none of this is indexed or searchable, and we’re hiding collective knowledge in private databases that we don’t own.
They’re trained on a huge volume of text scraped primarily from the English-speaking web.
Jasper, Copy.ai, Moonbeam
more sophisticated methods of prompting language models, such as “prompt chaining” or composition.
Ought has been researching this
libraries like LangChain
Prompt chaining is a way of setting up a language model to mimic a reasoning loop in combination with external tools.
It can pick from a set of tools to help solve the problem, such as searching the web, writing and running code, querying a database, using a calculator, hitting an API, connecting to Zapier or IFTTT, etc.
“generative agents”.
Just over two weeks ago, this paper **“Generative Agents
These language-model-powered sims had some key features, such as a long-term memory database they could read and write to, the ability to reflect on their experiences, planning what to do next, and interacting with other sim agents in the game.
There’s a new library called AgentGPT
It’s now relatively easy to spin up similar agents that can interact with the web.
we’re about to drown in a sea of informational garbage.
absolutely swamped by masses of mediocre content.
We’ll need to find more robust ways to filter our feeds and curate good-quality work.
Such as facilitating genuine human connections, pursuing collective sense-making and building knowledge together, and ideally grounding our knowledge of the world in reality.
about digital gardening which is essentially having your own personal wiki on the web.
make the web a space for collective understanding and knowledge-building,
Why does it matter if a generative model made something rather than a human?
differences between content generated by models versus content made by humans.
First is its connection to reality. Second, the social context they live within. And finally their potential for human relationships.
generated content is different because it has a different relationship to reality than us.
This is the core of all science, art, and literature. We are trying to understand and teach each other things through writing.
In some sense, it’s fully UNHINGED. The model cannot check its claims against reality because it can’t access reality.
They’re confused about who they are and where they are, but they’re still super knowledgeable.
So simulated humans that can only deal with language are missing a big part of what we perceive as human “reality.”
Everything we say is contextual and relies on a shared social world.
They know nothing about the cultural context of who they’re talking to.
represent a very particular way of seeing the world.
“Every way of life represents a communal experiment in living. The world itself is never settled in its structure and composition. It is continually coming into being.”
Generating a mass of content from a very particular way of viewing the world funnels us down into a monoculture.
When you read someone else’s writing online, it’s an invitation to connect with them.
A lot of this talk is based on an essay called The Expanding Dark Forest and Generative AI
how we might prove we’re human on a web filled with fairly sophisticated generated content and agents.
On the new web, we’re the ones under scrutiny. Everyone is assumed to be a model until they can prove they’re human.
This raises both the floor and the ceiling for the quality of writing.
They will try to outsource too much cognitive work to the language model and end up replacing their critical thinking and insights with boring, predictable work.
they shouldn’t be letting language models literally write words for them. Instead, they’ll strategically use them as part of their process to become even better writers.
using them as sounding boards while developing ideas, research helpers, organisers, debate partners, and Socratic questioners.
enter a phase of human centipede epistemology.
going to use the text generated by these models to train new models. That tenuous link to the real world becomes completely divorced from
We will begin to preference offline-first interactions.
the only way to confirm humanity is to meet offline over coffee or a drink.
Two people who know each of these people can confirm each other’s humanity because of this trust network.
create on-chain authenticity checks for human-created content on the web.
reasonable to assume we’ll each have a set of personal language models helping us filter and manage information on the web.
The product decisions that expand the dark forestness of the web are the problem.
if you are working on a tool that enables people to churn out large volumes of text without fact-checking, reflection, and critical thinking. And then publish it to every platform in parallel… please god, stop.
First, protect human agency. Second, treat models as reasoning engines, not sources of truth And third, augment cognitive abilities rather than replace them.
more ideal form of this is the human and the AI agent are collaborative partners doing things together. These are often called human-in-the-loop systems.
locus of agency remains with the human.
treat models as tiny reasoning engines, not sources of truth.
One alternate approach is to start with our own curated datasets we trust.
We can then run many small specialised model tasks over them. We can do things like: * Summarise * Extract structured data * Find contradictions * Compare and contrast * Group by different variables * Stage a debate * Surface causal reasoning chains * Generate research questions
These outputs aren’t final, publishable material. They’re just interim artefacts in our thinking and research process.
This paper on **“Sparks of Artificial General Intelligence
we should be augmenting our cognitive abilities rather than trying to replace them.
Note: Good picture too
Language models are very good at some things humans are not good at, such as search and discovery, role-playing identities/characters, rapidly organising and synthesising huge amounts of data, and turning fuzzy natural language inputs into structured computational outputs.
And humans are good at many things models are bad at, such as checking claims against physical reality, long-term memory and coherence, embodied knowledge, understanding social contexts, and having emotional intelligence.
we should think of robots as animals – as a companion species who compliments our skills.
