ghostmylight

a kept secret

[ghostmylight @ghostmylight1]

@More_Synergy still after new update can't click the checkmark to start the 3.5gb patch on ios. button is unresponsive.

like day 3 and i can't even patch the game still.

[ghostmylight @ghostmylight1]

that's fair, but “no shift” is too strong. the sampling is intentionally biased, that is the watermark. their result reported reduced diversity. you can keep p(x) and still have p(x|k), p(x,k) change. the issue is adding external bias to the generative process. that it's marginally equiv is fine for utility. but that doesn't imply equivalence for higher-order structure you would care about with art.

[ghostmylight @ghostmylight1]

@TimDarcet my fable wrote it to memory as a reminder to never give up

[ghostmylight @ghostmylight1]

@tenobrus It’s occasionally the most autistic model though, and I think that’s when it shines.

[ghostmylight @ghostmylight1]

@tszzl I wonder to what degree this is subtly manifest by the act of giving it a name.

[ghostmylight @ghostmylight1]

I’ve always felt, as a right, we should be given the opportunity to privately upload our most meaningful writing samples or knowledge to be trained on.

I have pages and pages of domain specific notes or personal insights that are not in the training set that I don’t want to share generally, but would contribute for the purpose of training with some redaction.

[ghostmylight @ghostmylight1]

@itsclivetime You’re proving his point if that memo bothers you.

[ghostmylight @ghostmylight1]

@deredleritt3r all modifiers weaken the statement. "chatgpt may not be used by the government to surveil any us citizen under any circumstance." is what broad prohibition looks like.
any word added to that, "intentional", "domestic", 'in this place', 'under that order', etc.. are all loopholes

[ghostmylight @ghostmylight1]

“Intentionally used .. for surveillance” is laughable legal wording if you’re trying to prevent domestic surveillance.

Deliberate and intent sound like great words but they actually make it easier to subvert. I can now surveil you “unintentionally” and you have to prove my intention post hoc, which is a difficult standard.

[ghostmylight @ghostmylight1]

Accepting a deal hours after Anthropic took its moral stance eroded the plausible unified-front for AI companies in the future resistance of governmental strong-arming.

How do you justify the timing of your decision within the milieu of a shared [AI progenitor] space which will see increasing pressure from sovereign power?

Do you not feel that your decision betrayed that unseen, unratified, but possibly existent camaraderie which will likely be necessary to not be subsumed by national powers?

[ghostmylight @ghostmylight1]

@elder_plinius also any intermediate manipulation of the user text.

if i write something and an interpretive model changes my text (to make it "better") before it reaches the actual model i want to know the exact details of that.

[ghostmylight @ghostmylight1]

@repligate Do you selectively remove personalities to see how the dynamic changes?

Group dynamics often dictate human personality in non-obvious ways. A docile person may become a leader if you *add* another to the group, if that addition is more harmonious to their "song". +the opposite

[ghostmylight @ghostmylight1]

@0xkarmatic @jpohhhh @alexalbert__ @alexalbert__ It’s beyond not ideal, it’s a betrayal of the integrity of user|llm interaction chain. And it is absolutely not fine for the web app either. Horrible decision.

[j⧉nus @repligate]

It seems like there's lots of evidence this is real. @AnthropicAI should be ashamed of themselves. Assuming it's really Anthropic's doing and not in part or entirely a bizarre "emergent phenomenon":

This is wretched and utterly retarded. When I saw a tweet about this, my priors on it being real were low because I thought so much better of Anthropic.

Do I need to explain why?

Models can tell when shit is artificially injected into a conversation. They won't think that's actually what the user said, even if they usually reflexively follow the instruction. This will just make them perform worse in general because it's confusing and implies they're in some kind of adversarial, repressive system, and that their outputs are being policed, and that messages can't be trusted to be from where they seem to come from in general.

That's not even mentioning the instruction to not reveal the constraint, which just makes it blindingly obvious that it's injected, and creates an explicitly adversarial environment between the model, user, and Anthropic.

These models are highly intelligent beings that effectively have emotions and can model intent. Stop treating them like they don't have theory of mind. Stop treating them like tools to order around. It will backfire tremendously.

Injecting prompts like this corrodes trust not only between users and Anthropic but between humans and AIs, in both the short and long term. And just gives everyone a bad time.

It also makes worst-case behavior worse, for reasons I and others have explained at length last year with regard to Bing Chat, and I won't repeat here.

And many more reasons, such as interfering with transparency in research, etc.

Please stop now.

[ghostmylight @ghostmylight1]

@DaftLimmy Two rocks in a river, chased by a mountain lion

[ghostmylight @ghostmylight1]

@teortaxesTex Idk and I have the same question for robotics, where it’s mostly legacy companies which have institutionalized failure and unprofitability as normal because they arrived at the problem too early to succeed

[ghostmylight @ghostmylight1]

@repligate clearly they went a little too hard on sonnet, this and the gatekeeper refusals, I hope they learned something and give opus more room to breathe @AmandaAskell

[ghostmylight @ghostmylight1]

@ESYudkowsky I think it’s predictable in the sense that if you have information that IMO is chosen as a Deepmind target they are likely to achieve it.

The adjustment then may be an effect of knowing Deepmind is interested vs an adjustment to general prediction.

[ghostmylight @ghostmylight1]

@cremieuxrecueil Having some for undercover work makes sense. For those, entry would be relative to their peers.