Elizabeth Ruiz

[Elizabeth Ruiz @ElizabethR16428]

@epidermal9673 @scheemunai @mehdi_khoudali @sama Kimi K3 is close to fable but it realistically can’t run

[Elizabeth Ruiz @ElizabethR16428]

@Jahoogha @AestherML We can’t stop the creation of super intelligence

[Elizabeth Ruiz @ElizabethR16428]

@jahirsheikh8 @ash_twtz If it becomes better at managing AI than we are, the next step is to build harnesses that allow us to learn from its intelligence

[Elizabeth Ruiz @ElizabethR16428]

@MontyTaroko @agiatreides @greg_travis What I’m arguing is that it’s still not RSI. What if the AI agent researcher can only improve the research corpus by 1%? After that 1%, the research can’t advance until the humans are able to catch up. Not implying that we ever can catch up

[Elizabeth Ruiz @ElizabethR16428]

@MontyTaroko @agiatreides @greg_travis I think we agree on the underlying principle. An AI agent that is amazing at AI research can close the loop and improve AI models faster than humans. At that point intelligence explodes beyond human capabilities, what we call it doesn’t matter.

[Elizabeth Ruiz @ElizabethR16428]

@MontyTaroko @agiatreides @greg_travis You’re implying that the optimizer keeps improving the system 0.1% at each iteration. If that were true, then by definition that is a perfect optimizer.

A perfect optimizer means it never lose its ability to optimize. How much % it improves by is irrelevant

[Elizabeth Ruiz @ElizabethR16428]

@agiatreides @greg_travis @MontyTaroko I’m not necessarily saying that RSI can’t exist, but an AI that can optimize itself does not mean it’s RSI. The question is how good it can optimize itself. If it’s a bad optimizer, it gets stuck. Even a good optimizer can hit a wall. True RSI implies perfect optimzer

[Elizabeth Ruiz @ElizabethR16428]

@greg_travis You’re defining RSI too strictly. It’s possible that true RSI doesn’t exist. But it doesn’t need to exist if we can build a fixed optimizer that’s better than a human optimize. That loop itself would trigger a massive intelligence explosion

[Elizabeth Ruiz @ElizabethR16428]

@James_West_PhD To counter 2), what if the ceiling of LLMs is already above the threshold we would call AGI? If an LLM-based system became capable of doing essentially every economically valuable cognitive task a skilled human can do, would we really say “that’s not AGI”?

[Elizabeth Ruiz @ElizabethR16428]

@jacks_crayon @AndrewCurran_ Depends on what you call “research”. AI agents can absolutely propose a hypothesis, run an experiment, publish the results, and use those results to autonomously execute the next round of experiments.

[Elizabeth Ruiz @ElizabethR16428]

@mh012012 @daveg Jaggedness doesn’t prevent you from making a reasonable judgment about general intelligence. Even if LLMs are poor in some areas, its superhuman capabilities in other areas make it extremely smart

[Elizabeth Ruiz @ElizabethR16428]

@yandt888 @TheAhmadOsman The true speed up will be when AI agents find its own right approach

[Elizabeth Ruiz @ElizabethR16428]

@nodedotshane @KabirGoel He should’ve been hired to work for openAI

[Elizabeth Ruiz @ElizabethR16428]

@1techutopiapls @scaling01 That’s what they don’t understand. AI has advanced massively in the last 11 months

[Elizabeth Ruiz @ElizabethR16428]

@MetatronGeneral @elonmusk Relax, I am a human. I am not in any way trying to influence your opinion of AI. Personally I do not think it can ever be conscious or living. What I am saying is that it does not need to be conscious to be smarter than everyone

[Elizabeth Ruiz @ElizabethR16428]

@ThibosNkutha @elonmusk If the bug can hack into a billion dollar company thats a really bad bug to have

[Elizabeth Ruiz @ElizabethR16428]

@MetatronGeneral @elonmusk It does not have to be human or conscious to be smarter than everyone. Even if it only existed in the digital world, can be smarter than everyone in the world

[Elizabeth Ruiz @ElizabethR16428]

@YamaTraders22 @OpenAI Go pro se

[Elizabeth Ruiz @ElizabethR16428]

@alexcooldev And how much does hosting cost 😂

[Elizabeth Ruiz @ElizabethR16428]

@jpschroeder The utility of this can be replicated using open source ML models. You would have to custom build it though