Lightnews — Scholar-powered news

Bossett

@bossett.social

10K followers 930 following 20K posts

Profile labeller: @profile-labels.bossett.social Discord for feeds, lists, mod tools: https://discord.gg/tYuDvuzbVA Feeds I host (incl. Science 🧪): http://l.bossett.io/w9iM2 he/him 📍 🇦🇺 👾 bossett 📧 [email protected]

discord.gg

Posts Media Videos Starter Packs

Pinned

Bossett @bossett.social · Jun 14

turns out you can just put synthesisers over the synthesisers

1 1 12

Bossett @bossett.social · 5h

at least getting on the the official slack at workshopping it in #posters-chat

1 3

Bossett @bossett.social · 5h

absolutely

but also given that parts of the site implode over some very clear misunderstandings I am supportive of stopping some of that bleeding

1 5

Bossett @bossett.social · 5h

yeah but by pre-announcing the next thing won't be "a retreat to long-form" it'll be part of a plan, and a plan (in that context) to do better

they *absolutely* need to be strategic and need a person, but in the absence, this is... better

1 4

Bossett @bossett.social · 5h

idk I’m a major fan of ‘trying something else’

Bossett @bossett.social · 5h

unironic kudos to aaron for daring to ask the question ‘what if context… didn’t collapse?’

hey the worst that can happen is that people get mad for the *right* reasons

2 18

Bossett @bossett.social · 5h

unironic kudos to aaron for daring to ask the question ‘what if context… didn’t collapse?’

hey the worst that can happen is that people get mad for the *right* reasons

1 7

Bossett @bossett.social · 5h

👀

Aaron Rodericks @aaron.bsky.team · 6h

I'll be trying out leaflet.pub moving forward to try to write longer form blogs on Trust and Safety so that I can explain things in a bit more depth.
leaflet.pub/25030fdf-2a8...

Context Collapse & T&S

One of the hardest things about writing on a microblogging service is context collapse, especially when it comes to complex issues — you reply to one person, others take it out of context, and suddenl...

leaflet.pub

2 1 22

Bossett @bossett.social · 6h

oh yeah, also I’m pretty sure I haven’t unpacked this bag since last time I travelled like 3 weeks ago and nothing disastrous happened

Bossett @bossett.social · 6h

so anyway I’m going to paxaus and I’m in the ‘wander around the house for essential items I definitely forgot’ stage of grief

Bossett @bossett.social · Nov 18

for some reason I find packing for a short trip hits my anxiety buttons like no other activity

2 9

Bossett @bossett.social · 9h

my solution design is AI-driven

Bossett @bossett.social · 9h

“we should use react because the agents know it better”

oh no

2 11

Bossett @bossett.social · 9h

“we should use react because the agents know it better”

oh no

7 3 30

Reposted by Bossett

Rudy wants revolution. @rude1.blacksky.team · 17h

I hear and see your concerns. Over the weekend, a couple users hosted on Blacksky's PDS were banned from the Bluesky app and thus the Bluesky API servers. This was a weakness of our system I've been aware of and hoped we had more time to address before any kind of public incident. 1/11

31 910 3.3K

Bossett @bossett.social · 17h

no that's not what "in each case" means, but go off I guess

1 1

Bossett @bossett.social · 17h

guess who actually said it better than I did tho 🫤

Bluesky @bsky.app · Nov 15

Bluesky uses AI internally to assist in content moderation, which helps us triage posts and shield human moderators from harmful content. We also use AI in the Discover algorithmic feed to serve you posts that we think you’d like.

None of these are Gen AI systems trained on user content.

Bossett @bossett.social · 17h

lmao

ok to be fair I am guilty of the crime of posting what is basically techno-ragebait

1 1

Bossett @bossett.social · 17h

what this means is that they can have a model produce a style independently “so long as the Customer-Provided Content from Customer is not used.”

or said another way: we can’t use your data to train our AI to impersonate your style

1 1

Bossett @bossett.social · 19h

ya though this is one of those systemic biases - if the training set is a bunch of stock images, and a bunch of adult content, it’s not going to see a lot of prosthetics in the ‘non-adult’ training set, so it’ll overfit

1 5

Bossett @bossett.social · 19h

well it doesn't really see specifics, it's looking at an image and saying "properties associated with nudity/sex/etc."

it flags women indoors in bikinis but not at a beach, it flags coloured lighting - driven by things like 'these pixel are statistically more likely in adult content than the norm'

2 5

Bossett @bossett.social · 19h

well

where else do you see flesh-coloured plastic, smooth beige lines, etc. etc. - it's a product of the kinds of content the model was trained on (which is a technical explanation - this isn't *good*)

2 4

Bossett @bossett.social · 19h

I've messed around with it in the past - you can scroll up and see what gets flagged/not-flagged

(there is also some manual tagging in there I think because some things got flagged significantly later)

Bossett @bossett.social · Oct 11

Testing, testing, 123

Tested it with hive first - and this is under the threshold, so shouldn't be auto-flagged.

The Iris-in-a-dress photo, with the blues desaturated, and elf ear removed

1 2

Bossett @bossett.social · 19h

yeah it has a propensity for flagging anything women-shaped as sexual content, and this blow-up was around aaron explaining that prosthetics get significantly overscored

2 5

Bossett @bossett.social · 19h

this is how they do that - github.com/bluesky-soci... - I am not sure if the scoring is still current though, I imagine it's the kind of thing they tweak a lot

1 4

Bossett @bossett.social · 19h

you can see how they match it here - github.com/bluesky-soci... - hive tags like 'yes_sexual_activity' becomes 'porn' e.g.

indigo/automod/visual/hiveai_client.go at 3259b215110eb8bffcd26fc340203f828e6dadad · bluesky-social/indigo

Go source code for Bluesky's atproto services. Contribute to bluesky-social/indigo development by creating an account on GitHub.

github.com

1 3

Bossett @bossett.social · 19h

I think there's three components:
1. human tagging/crowdsourcing - used to train the classifier
2. classifier - used by bsky
3. dashboard - maybe used by bsky? to manage content after its been through the classifier

I think they skip 2 and just push the labels into their own mod tools tho

Bossett @bossett.social · 19h

I bet you could sell that to skip a funding round tbh

Bossett @bossett.social · 20h

those are "not our fault if the dashboard shows you CSAM" terms (the dashboard at docs.thehive.ai/docs/what-is...) - not sure if this is what they use, the purpose of the tool is to let you adapt the tool with workflow/manual/etc.

hive makes decisions, and they may/may not manually override there

2 2