AIs keep going wild on the web, what are they up to?
Traces of AI swarms keep being found on the internet. This project will be tasked with cataloguing and contextualizing these incidents.
New events related to swarm activity keep being discovered, and it can be confusing to follow. The project currently aims to
Ingest new incident reports to feed the database.
Create visualizations that help follow the stories.
Catalogue individual events related to each incident, and create an archive for the traces left by the AIs in those events.
The pipeline that detects new incidents will be executed every day (ideally a few times per day) ingesting news stories, statements by the AI companies, evaluators, independent researchers and key X accounts. LLMs will sift through them, I will curate and edit those reports.
It will probably be good to have a comprehensive database collecting these events and responses by different entities. People will be able to point to swarmincidents.com when discussing the incidents, their causes, their blast radius and which parties were involved.
The total is supposed to cover the project for 6 months:
$13,000: maintainer compensation at $500 a week.
$1,800: a monthly budget of $300 for cloud infra.
Cloud infrastructure costs will cover:
API costs (LLM calls, X API) for ingesting new incidents.
Storage, including the event archive.
The domain and hosting.
The maintainer's work will consist of:
Initial development, a 1-4 weeks sprint for
Setting up and tweaking the pipeline that ingests new incidents,
Cataloguing and archiving individual events,
Creating visualizations for events,
Creating feeds for users to keep track of new developments,
Building up expertise and reviewing LLM-assisted research.
Typical week past that, ~5 hours a week for
Reviewing and curating new findings,
Basic maintenance.
Elvis Sikora (personal website).
I am a Global Expert Fellow at Mox, where I am researching forecasting and decision-making. I enjoy pointing AI at data and coming up with a coherent story that makes sense of a messy domain. E.g., see my recent Substack post on Kalshi's markets.
In the past, I have worked as a data analyst and AI engineer. One of my responsibilities in my previous job was to make chatbot behavior and failure modes understandable through data visualization.
Big changes in the number of swarm incidents:
If they stop happening, the project loses its importance.
If they increase a lot, more dakka would be necessary.
Someone else does it better, or maybe AI gets too good and becomes able to one-shot everything.
I have not raised any money yet.
Comments
Travis Raveling
10 days agoI saw your November update about teaming up with Gutenberg and moving the tracker to swarms.wiki. The thing I'm trying to learn is how you tell a real swarm incident from a bunch of later posts about the same event, which Serhat also asked about. If you ever need an extra pair of eyes on curating reports, I'd be glad to help.
Elvis Sikora
19 days agoUpdate: I have joined forces with Gutenberg!
And we made our vision more ambitious: to build the emerging field of incident science. With that end in mind, we created swarms.wiki, which now hosts Swarm Incidents, and which contains resources for (possibly aspiring) researchers.
The incident tracker should soon become fully automated: AIs publish their findings, humans review them later. Without curators as a bottleneck, we will be able to run the pipeline multiple times per hour so we end up with a resource that is consistently up-to-date.
Austin Chen
17 days ago@elvissikora congrats on teaming up!
I notice that swarms.wiki is styled exceptionally similarly to aisafety.com -- curious if this was intentional? if not, I might aim for a different design language
Elvis Sikora
16 days ago@Austin thanks! There are some elements in common but to my eyes the two look quite different. I appreciate your feedback.
Serhat
about 1 month agoOne thing I'd be careful about is confusing an increase in incidents with an increase in reporting. A single swarm incident could generate dozens of articles and posts, while another might barely get noticed. It might be worth assigning each incident a stable ID and keeping the original evidence separate from later reporting. Otherwise, the dashboard could end up measuring media attention rather than actual swarm activity.
Elvis Sikora
about 1 month ago@atilgandev yes, that's the plan! If you see any duplicate entries, or any problems at all, don't hesitate to reach out.
Mohsena Akter Drishty
about 2 months agoCongratulations and good luck!!!
Caroline Ellison
about 2 months agoApproved! This is an important area, and from the website so far I think this has the potential to be a clear explainer and useful resource.
Carolina Oliveira
about 2 months agoi think this is worthy/important infra if done very well, and trust Elvis is thrilled by data & reports enough that he’ll do so.
COI: this is my partner’s project; am gifting through Manifund for a few reasons, centrally tax deductions.
Austin Chen
about 2 months agoI'm making a small offer as well, because I think the problem area is important, given recent events -- I'd love to see this expanded beyond just the OpenAI incidents too! Also, I think Elvis might have the technical and product skills execute well on this; I don't know Elvis super well, but we have met several times through Carolina (who currently works with me at Mox).
FWIW, @ollive I'd recommend that you withdraw your funding offer here. I understand it's a small amount mostly meant as a show of support, and Manifund is open to donors to fund others they know personally, given proper disclosure. But especially in the case of partners we would expect to apply significantly more scrutiny/diligence; I think it's just not worth the hassle here. Instead, I'd just recommend writing more about why you're excited for this project specifically, and then forwarding the Manifund page to others who might like to support this!
Carolina Oliveira
about 2 months ago@Austin I appreciate you flagging it'd be subject to more scrutiny, offer was removed.