101–110 of 123 posts
Plug: I’ve built something vaguely similar, but of course a trillion times less brilliant and powerful, yet nonetheless useful for some straightforward cases where really you just want GPT et al. to call your API at just the right moment of the conversation, with JSON that is guaranteed to be valid: https://github.com/manuelkiessling/php-ai-tool-bridge
Not sure if this is obvious. But its incorrect to ditch on GPT-4. The paper uses self-instruct on GPT-4 to generate the training data on which it is fine-tuned. This paper would not exist without GPT-4. Although they claim GPT-4 can be replaced by any LLM, I'm sure the results would be nowhere as good, and so they stuck with GPT-4.
Not sure if this is obvious. But its incorrect to ditch on GPT-4. The paper uses self-instruct on GPT-4 to generate the training data on which it is fine-tuned. This paper would not exist without GPT-4. Although they claim GPT-4 can be replaced by any LLM, I'm sure the results would be nowhere as good, and so they stuck with GPT-4.
It does make me wonder what the converged fixed point on this technique is. If I fine tune with GPT4 to make model A, which then performs better than GPT4, then fine tune model B with A, at what point does either artifacting or diminishing returns set in?
Earlier quoted context omitted.
Sure, my argument is that there is zero evidence whatsoever we will be able to prevent these from becoming dangerous or that we’d be able to stop deployment once they do. All technologies are dangerous, and many of the most dangerous ones correctly have tons and tons of safeguards around them both as intrinsic properties of the technology (e.g. it takes nationstate resources to produce a nuke) and extrinsic constrain…
> my argument is that there is zero evidence whatsoever we will be able to prevent these from becoming dangerous Well no, but there's no need to prevent them from becoming dangerous inherently. They are tools, extensions of human agency. Tools are, by their nature, purpose-agnostic. It is good for humans to have better tools and more agency; good humans tend to cooperate and limit the harms from bad humans, while inc…
Don't worry folks, the years and years of "of course we'd never connect experimental AI systems to the open internet" assurances were never necessary for safe development and deployment. This reassessment was founded on the last several months that showed us these systems are close to 100% dependable, they never deceive humans, they are entirely aligned with human morals (you know, all those morals we all agree on),…
Plug: I’ve built something vaguely similar, but of course a trillion times less brilliant and powerful, yet nonetheless useful for some straightforward cases where really you just want GPT et al. to call your API at just the right moment of the conversation, with JSON that is guaranteed to be valid: https://github.com/manuelkiessling/php-ai-tool-bridge
Very cool and thanks for sharing, would you say this is similar to langchains tool concept?
But it works surprisingly well, because the underlying idea (again, from LangChain) is simply clever.
Earlier quoted context omitted.
Paperclip Maximizer was a conceptual mistake - not because it's wrong, but because people get hung up on the "paperclip" bit. Bostrom and the disciples of Yudkowsky should've picked a different thought experiment to promote. Something palatable to the general population, that seems to have problems with idea of generic types. The example of Paperclip Maximizer was meant to be read as: template class Maximizer { ... }…
They should have learned the fundamentals of ML before promoting thought experiments. No, it's not that people are silly and get stuck on the paperclips bit, it's that you uncritically buy assumptions that a meaningful general-purpose optimizer is 1) a natural design for AGI 2) may pursue goals not well-aligned with human intent 3) is hard to steer.
"Fundamentals of ML" were known by these people. They also don't apply to this topic in any useful fashion.
Earlier quoted context omitted.
Other than snark, do you have a good argument? We know that technology can be error-prone, and LLMs fail in a great plethora of ways, but you are trying to sell an AI Doom narrative. I have never bought the idea that AI will be airgapped, because the whole paradigm of Yudkowsky at al. is ludicrous and even within it airgapping was a strawman of a technique (they argue that a truly dangerous AI will get itself out reg…
Sure, my argument is that there is zero evidence whatsoever we will be able to prevent these from becoming dangerous or that we’d be able to stop deployment once they do. All technologies are dangerous, and many of the most dangerous ones correctly have tons and tons of safeguards around them both as intrinsic properties of the technology (e.g. it takes nationstate resources to produce a nuke) and extrinsic constrain…
Is there an argument that we can prevent nuclear power, cars, or the sun from becoming dangerous?
Does the fact that all of those things are intrinsically dangerous mean we should panic?
AI may or may not kill is all, but I’m pretty sure random panic won’t change the outcome.
Earlier quoted context omitted.
> Wait, I too thought your original comment on alignment questioned its fundamental premise—than one dominant culture should not/cannot define the adequacy of alignment. What?! This is exactly what I was worried about when OpenAI, et al. co-opted the term "alignment" to refer to forcibly biasing models towards being polite, unobjectionable, and espousing specific flavor of political views. The above is not the import…
The one thing we have going for us is the general public and government officials seem to grok it when it is explained plainly to them. Tech and developer types who have been drinking the Silicon Valley koolaid for too long will complain and sealion, but the average person seems to realize how self evident it is that - oh, this is really really bad and we should really really stop this.
Earlier quoted context omitted.
> Alignment assumes a well agreed foundational philosophy on what is good, what is fair, what is doable today and tomorrow. Alignment assumes that there exists a foundational philosophy on what is good and fair and nice, that's close enough a match to everyone. It's a reasonable assumption, because there are core human universals, and the cultural differences around the world are a rounding error in comparison. We're…
> We're not talking here about someone's view on when white lies are justified or which model of marriage is the bestest - we're talking at the level of "cooperation = good", "love = good", "trust = good", "death = bad", "suffering = evil", etc. Most people disagree to a significant degree. Reminder: the majority of humanity (and a big majority of people that have 2+ children) adhere to religious doctrines which all…