- May 2023
-
maggieappleton.com maggieappleton.com
-
We will have to design this very carefully, or it'll give a whole new meaning to filter bubbles.
Not just bubble, it will be the FB timeline. Key here is agency, and design for human biases. A model is likely much better than I to manage the diversity of sources for me, if I give it a starting point myself, or to see which outliers to include etc. Again I think it also means moving away from single artefacts. Often I'm not interested in what everyone is saying about X, but am interested in who is talking about X. Patterns not singular artefacts. See [[Mijn ideale feedreader 20180703063626]]
-
I expect these to be baked into browsers or at the OS level.These specialised models will help us identify generated content (if possible), debunk claims, flag misinformation, hunt down sources for us, curate and suggest content, and ideally solve our discovery and search problems.
Appleton suggests agents to fact check / filter / summarise / curate and suggest (those last two are more personal than the others, which are the grunt work of infostrats) would become part of your browser. Only if I can myself strongly influence what it does (otherwise it is the FB timeline all over again!)
If these models become part of the browser, do we still need the browser as a metaphor for a window on the web, or surfing the net? Why wouldn't those models come up with whatever they grabbed from the web/net/darkweb in the right spot in my own infostrats? The browser is itself not a part of my infostrats, it's the starting point of it, the viewer on the raw material. Whatever I keep from browsing is when PKM starts. When the model filters / curates why not put that in the right spots for me to start working with it / on it / processing it? The model not as part of the browser, but doing the actual browsing, an active agent going out there to flag patterns of interest (based on my prefs/current issues etc) and organising it for me for my next steps? [[Individuele software agents 20200402151419]]
-
Those were all a bit negative but there is some hope in this future.We can certainly fight fire with fire.I think it’s reasonable to assume we’ll each have a set of personal language models helping us filter and manage information on the web
Yes, agency at the edges. Ppl running their own agents. Have your agents talk to my agents to arrange a meeting etc. That actually frees up time. Have my agent check out the context and background of a text to judge whether it's a human author or not etc. [[Persoonlijke algoritmes als agents 20180417200200]] [[Individuele software agents 20200402151419]]
-
And lastly, we can push ourselves to do higher quality writing, research, and critical thinking. At the moment models still can't do sophisticated long-form writing full of legitimate citations and original insights.
Is this not merely entering an 'arms race' against our own tools? With the rat race effect of higher demands over time?
What about moving sideways not up? Bringing in the richness of the layering of our (internal) reality and lives? The entire fabric that makes up our lives, work, communities, societies, indicately more richly in our artefacts. Which is where my sense of beauty is [[Schoonheidsbegrip 20151023132920]] as [[Making sense is deeply emotional 20181217130024]]
-
On the new web, we’re the ones under scrutiny. Everyone is assumed to be a model until they can prove they're human.
On a web with many generative agents, all actors are going to be assumed models until it is clear they're really human.
Maggie Appleton calls this 'passing the reverse Turing test'. She suggests using different languages than English, insider jargon etc, may delay this effect by a few months at most (and she's right, I've had conversations with LLMs in several languages now, and there's no real difference anymore with English as there was last fall.)
-
I think we’re about to enter a stage of sharing the web with lots of non-human agents that are very different to our current bots – they have a lot more data on how behave like realistic humans and are rapidly going to get more and more capable.Soon we won’t be able to tell the difference between generative agents and real humans on the web.Sharing the web with agents isn’t inherently bad and could have good use cases such as automated moderators and search assistants, but it’s going to get complicated.
Having the internet swarmed by generative agents is unlike current bots and scripts. It will be harder to see diff between humans and machines online. This may be problematic for those of us who treat the web as a space for human interaction.
-
There's a new library called AgentGPT that's making it easier to build these kind of agents. It's not as sophisticated as the sim character version, but follows the same idea of autonomous agents with memory, reflection, and tools available. It's now relatively easy to spin up similar agents that can interact with the web.
AgentGPT https://agentgpt.reworkd.ai/nl is a version of such Generative Agents. It can be run locally or in your own cloud space. https://github.com/reworkd/AgentGPT
-
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
Generative agents have a database for long term memory, and can do internal prompting/outputs
-
Recently, people have taken this idea further and developed what are being called “generative agents”.Just over two weeks ago, this paper "Generative Agents: Interactive Simulacra of Human Behavior" came out outlining an experiment where they made a sim-like game (as in, The Sims) filled with little people, each controlled by a language-model agent.
Generative agents are a sort of indefinite prompt chaining: an NPC or interactive thing can be LLM controlled. https://www.youtube.com/watch?v=Gz6mAX41fs0 shows this for Skyrim. Appleton mentions a paper https://arxiv.org/abs/2304.03442 which does it for simlike stuff. See Zotero copy Vgl [[Stealing Worlds by Karl Schroeder]] where NPC were a mix of such agents and real people taking on an NPC role.
-