Lightnews — Scholar-powered news

Max Andrews @thatsjustlikeyouropinionman.com · 2h

We’re building kardashev type II modular fusion synthesis reactor that can produce any desired element on demand, provided an energy conduit of at least 0.7 sols.

1 2

Max Andrews @thatsjustlikeyouropinionman.com · 5h

interestingly looks like embedding approaches are less susceptible to gaslighting

Max Andrews @thatsjustlikeyouropinionman.com · 5h

1 1

Max Andrews @thatsjustlikeyouropinionman.com · 6h

Used to work at one of those; No, they’re just bad. It’s unavoidable when you’re an integration company rather than an engineering company, especially in a low-margin commodity business that competes mainly on price.

1 1

Max Andrews @thatsjustlikeyouropinionman.com · 7h

interestingly, if I omit your incorrectly paraphrased generalization, it's good faith. But if I include it, it just sneaks over the line into bad faith. Pretty accurate! Thanks for the live test case.

1 6

Max Andrews @thatsjustlikeyouropinionman.com · 7h

ATproto moderation works via labeling, which you as a user have to specifically subscribe to. It only affects what you see in your own view of the app, different people can choose their own filters according to their own needs.

1 1

Max Andrews @thatsjustlikeyouropinionman.com · 7h

I am not a member of the bluesky team; this is an independent concept I shared. And this is a user-directed framework, allowing anyone to define examples of things they do and do not want to see in their replies.

2 4

Max Andrews @thatsjustlikeyouropinionman.com · 7h

a close up of a man 's face with his eyes closed

ALT: a close up of a man 's face with his eyes closed

media.tenor.com

1

Max Andrews @thatsjustlikeyouropinionman.com · 8h

Done. @hailey.at @pfrazee.com github.com/madmaxbr5/Hy...

1 17

Max Andrews @thatsjustlikeyouropinionman.com · 16h

going to port this to instruction-aware embeddings tomorrow which should be a step up in accuracy

2 5

Max Andrews @thatsjustlikeyouropinionman.com · 16h

Effectiveness improves with number of examples added into the corpus. Nice thing about this is A) it's very fast compared to LLM and B) you can really add any arbitrary examples you want to filter out over time to create a fully personalized filter, which is really just a dataset.

1 7

Max Andrews @thatsjustlikeyouropinionman.com · 17h

You can also try skipping the LLM and using HYDE embedding similarity. Each time you flag something that you don't want to see, it's added to the example document corpus. Use LLM to generate examples, not at test time. I'll do a quick POC

1 6

Max Andrews @thatsjustlikeyouropinionman.com · 18h

1 8

Max Andrews @thatsjustlikeyouropinionman.com · 18h

a man in a star wars costume is dancing in front of a sign that says this is the way .

ALT: a man in a star wars costume is dancing in front of a sign that says this is the way .

media.tenor.com

1

Max Andrews @thatsjustlikeyouropinionman.com · 18h

They see the pollin'...
They hatin'...
Deployin' interstate militia resources
Tryn'a provoke Antifa, tryn'a provoke Antifa...

1

Max Andrews @thatsjustlikeyouropinionman.com · 19h

top signal lmao

1

Max Andrews @thatsjustlikeyouropinionman.com · 1d

Everyone but the people in the animal costumes, who should do dance parties and play card games

1

Max Andrews @thatsjustlikeyouropinionman.com · 1d

What’s the intended use case here? Can PDQ be used for content classification generally? I thought it was more for detecting near-duplicates. So like… label one image and auto-apply it to all near-matches as well?

1 2

Max Andrews @thatsjustlikeyouropinionman.com · 2d

I have ideas on pushing some embedding workloads to the edge to help this scale better

Max Andrews @thatsjustlikeyouropinionman.com · 2d

proper framing, thank you

Max Andrews @thatsjustlikeyouropinionman.com · 3d

LMNOP

1

Max Andrews @thatsjustlikeyouropinionman.com · 3d

Big "old man argues with the insurance company call center person" energy.

1

Max Andrews @thatsjustlikeyouropinionman.com · 3d

Something like this will become the new standard for evidenciary sources i.e. security cameras, police bodycams, forensic captures, etc.

2

Max Andrews @thatsjustlikeyouropinionman.com · 3d

Authentic images and video need to be hashed by a signed sensor at the A/DC interface and published contemporaneously to create a point-in-time verifiable capture event.

1 1

Max Andrews @thatsjustlikeyouropinionman.com · 3d

Chicago should deploy its guard into Texas and start abducting random white people off the street, if interstate deployments operating without warrants are just fine and dandy.

2 13