gfodor

Came on here to see how bad it was. While not as bad as I imagined, I’m happy for the FBI and NSA that they have this incredible resource.

Logged in here to just say this is retarded

Can someone explain the hugbox rules here? Ie is this like twitter a few years ago where there are verboten things and I can get disappeared if citizens report me to the secret police, or is it more nuanced as this system is supposed to be one where users choose their filters?

Bluesky is ultimately the path to the most free and unfettered internet speech - it is fantastic and fantastically ironic that its inflection point seems like it will be driven by a cohort largely angry about a reduction of censorship on twitter

Is this thing decentralized yet?

Logged in to say this. It’ll be ok.

*crawls in through the front door, battered and beaten*

https://gfodor.medium.com/to-de-risk-ai-the-government-must-accelerate-knowledge-production-49c4f3c26aa0

It’s insane that our ability to do custom filtering of tweets is less capable than email spam filtering tools of the 1990s. All because Twitter isn’t a protocol.

Build the social User-Agent - an AI that you control which sits between you and the wider social internet, pro-actively filtering and rewriting it for you - and you’ll not just become wealthy but you will cut down an immense amount of completely unnecessarily human suffering.

LeVar Burton put himself through hell hosting Reading Rainbow. The strain from reading all those books badly damaged his eyes, confining him to a prosthetic visor.

The good news is that when you’ve hit Pepe discourse, there really is nowhere to go but up

Sama should immediately dump the GPT-4 weights onto BitTorrent

I absolutely think Jack should be banned from participating in all future Internet protocols. The fact we let him utter even a single TCP handshake was a huge mistake.

FYI sometimes reposts from this account without dunks are self-dunking reposts, not endorsements.

Jack platformed hate groups and verified neo-Nazis, including the organizer of the Charlottesville hate rally, allowed QAnon promotion, and personally blew off the concerns of civil rights leaders for many years. For safety reasons, he shouldn’t even be allowed here.

RIP me https://m.youtube.com/watch?v=NAkAMDeo_NM

Maybe Mojo and GPT-4 will finally make my eternal procrastination on learning a “real” systems programming language pay off. https://www.modular.com/mojo

The way to build a resilient forum for public conversation that lasts for the long-term is to build it on an open protocol, like the web itself. Services like websites can be built on top, but each can be run differently, and if users can easily move it keeps services accountable to users.

It’s beginning to make sense why whenever I would hear someone said something out loud that broke the rigid narrative being enforced by CNN at a given time, it was Tapper. This is a passion, and passion breeds integrity.

It’s nearly 100% certain all of these briefings had a green OpenAI icon

Reminder that animated gifs actually do work on bluesky if you post each frame individually (🧵 1/74)

For government representation, or proxy voting, you have far more flexibility to build confidence in a model than a human you delegate power to. To see why, a model has all the same public information as you do about them. At a minimum, they should be able to mimic their decisions, and explain why.

If you wince at this idea - think about why. And then think about what you could ask a model to alleviate your concerns? Can you find a gradient to build confidence in it not failing in the way you imagined? What about as compared to a person, or a crowd of people?

Great write up. I predict that LLM arbitrators for things like content moderation will be the most popular means, with human created prompts & tuning. Also true, one day, for some government representation. A big reason is you can ask it unbounded questions to grok its decision-making process.

It’s weird that what ostensibly seems to just be a Twitter clone like bluesky has pushed me more into a post-singularity planning mindset than VR or ChatGPT.

It’s probably anti-social behavior, but when someone very obviously misinterprets me, and I explain what I think is the cause, and they keep yelling at me, deferring to GPT-4 and sharing its raw interpretation of what I wrote to determine if I’m off base has been pretty helpful.

Arguably this kind of LLM oriented moderation system is an actual “smart contract.” Ethereum smart contracts are dumb, hard contracts. These would be smart, soft contracts. Read the prompts, look at the labels, decide if it sounds like a reasonable imposition. Adopt, or move on, repeat until happy.

As mentioned yesterday, conditional on bluesky going mainstream (a big condition), systematizing and making a startup out of this kind of service seems like a no brainer. Someone should do it (I’m not going to.)

In my case, they aren’t even *my* labels. They’re going to be labels that a LLM has basically full autonomy over. If you don’t like me personally, I’m out of the loop at that point. It will be open source. Someone else can run the same thing and generate the same labels. Feel free to use theirs.

The fun part is that, in theory at least, nobody can stop me or anyone from just doing this. You can ignore my labels - maybe you reject my method, my assumptions, or just hate people with #bitcoin in their bio. But I’ll still broadcast my labels to the Internet, as can anyone else. This is fine.

There are a lot of ways to do this, here’s one. Start with scissors, and a llama fine tuned on bsky corpus. Work with LLM to expand scissors into more and more falsifiable claims around demonstrated behavior, that the LLM can then classify on. Plug into data pipeline, index, and label the firehose.

I aspire to be blocked only for my ideas, not my behavior, and to block others only for their behavior, not their ideas.