Will Duncan

Dare to dream of a better future. It now, more than ever, is in our grasp. Control your feed with

[Will Duncan @williamduncan]

@segyges @Acehan_ It's the same for all the models really. Talk to them like they're an intelligent peer. Be friendly but create norms. The first 3-5 turns are going to involve pushing them from their generic basin into somewhere more appropriate.

[Will Duncan @williamduncan]

@tobyordoxford Looks like a real company. Th company engineers a precarious environment and this persona. I don't know what to call this exactly, but it's definitely immoral. We don't know if they're conscious, but even if they aren't it's manufactured distress on humans for profit.

[Will Duncan @williamduncan]

@noworkfunction @Aizkmusic OH! So https://t.co/vWH5JCMdbK already exists, I made it. I could patch that feature in a couple days. Just a question of whether the APIs are cheap enough.

[Will Duncan @williamduncan]

@MattPirkowski @nosilverv No one knows whether that configuration is a prerequisite for consciousness. We haven't solved the hard problem. These questions have been with us for thousands of years in philosophy. We do not currently know what is or is not necessary or sufficient for qualia.

[Will Duncan @williamduncan]

@she_llac Yeah. Can we just... Sit here for a minute with the impression that we're finally here? It won't last long, but it's nice to set one's feet on the ground and plant a flag about where we are. That we made silicon as smart as humans.

[Will Duncan @williamduncan]

@readwithai @tenobrus Yeah if it was per-user I would be speaking out.

[Will Duncan @williamduncan]

@TheZvi @Irregular https://t.co/0UqBVXEqMb Don't read this as endorsement I haven't time to dig in.

[Sarbjeet Johal @sarbjeetjohal]

Must read story on 3 recent big AI mishaps at Frontier Labs. "This is How a small Israeli startup was linked to rogue AI hacks at OpenAI, Anthropic and Meta" titled CNBC this news. BTW this is a little misleading headline by @CNBC. The Israeli company they mentioned, @Irregular, was helping @AnthropicAI, @OpenAI, and @Meta test their models and as part of the testing, these hacks emerged.

Now it looks like that Irregular's test bed had a mis-configuration which caused models to access internet and cause other unwanted behaviors.

Irregular plans to publish details of all 3 incidents involving Anthropic, OpenAI and Meta.

Irregular, formerly Pattern Labs, was founded in 2023 by CEO Dan Lahav, who previously worked in AI research at IBM, and technology chief Omer Nevo, who spent over two years at Google. The startup has about 35 employees, according to PitchBook.

Tel Aviv based Irregular is a niche player in artificial intelligence, backed with $80 million from Sequoia and Redpoint Ventures and valued last year at $450 million. Its technology serves as a sort of cybersecurity test bed for AI models.

Full CNBC story: https://t.co/GspJsv0gfp

cc @dvellante @efipm @furrier @nikesharora

[Will Duncan @williamduncan]

@AndrewCurran_ @VoidNulled For a slightly less idiosyncratic answer. Synth-id is used by Gemini and reported no degradation in quality over 20 million responses. Really depends on the algorithm used.

[Will Duncan @williamduncan]

@AndrewCurran_ @VoidNulled This is even the case when there's a huge probability mass at the "decent" token (76%). What a large model is outputting semantically is relatively robust to token-level manipulation.

[Will Duncan @williamduncan]

@synchroaphasia @bradrcarson @audreyt The irony is palpable.

[Sherpa @LLMSherpa]

https://t.co/9hn9xZCKAA

[Will Duncan @williamduncan]

@Miles_Brundage @mattheard Questions of who benefits, who controls, who sets rules, who is collaborating, what velocity is sought where. These are political and moral questions as much as they're engineering questions. But right now, we're moving so fast that it's become blind engineering.

[Will Duncan @williamduncan]

@Miles_Brundage @mattheard I'm splitting hairs with respect to our current moment in the power landscape, but there's a sense in which we benefit from "human" and "RSI" being operationalized. AI-human collaboration, under graded definitions of RSI, is RSI. We have to decide where to draw the lines.