Shapiro describes himself as: 'I research AI cognitive architectures based on Natural Language and LLMs. I also build automation tools and products with cutting-edge AI. Lastly, I conduct interviews with thought leaders and industry veterans.' In the video above he sets out a optimistic plan, he names a GATO framework, Global Alignment Taxonomy Omnibus. What GATO boils down to is an ambitious plan to solve the AI Alignment problem, based on heuristic imperatives.
Now, the purpose of this video is to introduce the crowning achievement of not just my work, but of the rapidly growing community that I'm building. What started around the years to comparatives as my research on alignment for individual models and agents has quickly expanded. So, this GATO Framework, Global Alignment Taxonomy Omnibus, is that comprehensive strategy that I just mentioned that is missing. It is not just for responsible AI development, but is a coherent roadmap that everyone on the planet can participate in at various levels, whatever level makes the most sense to you.
This framework has seven layers on ways to implement models, AI systems, and alignment-based regulations. We'll get into all the layers in just a moment. But basically, the whole point of this GATO Framework that we're working on is that it will unite all stakeholders, give us a common framework with which to have these discussions, to broaden the Overton window, to open the Overton window a little bit more. So, whatever part of the spectrum you're on, whether you're saying, "Eh, it's not really an issue yet," or "We're all going to die," or you don't care, or you're an optimist, whatever, this is a framework that we can all participate in, just in a decentralised, distributed, and open source manner.
So, as promised, here are the seven layers of the GATO Framework. And in the community, we started saying that it's like a seven-layer burrito, so we use like taco cat as our little avatar. So, layer one, the lowest layer, is model alignment. So, model alignment has to do with individual neural networks. So, that means GPT-2, GPT-3, GPT-4, BERT, Vicuña, Stable LM, all of these. Right? Large language models are proliferating like, well, I don't know, just like locusts, whatever is happening, right? Data sets are growing, models are growing, they're all coming out, the cat's out of the bag, right? Language technology, multimodal technology, it's all coming, you can't stop it.
So, rather than stop it, rather than call for moratoriums, what we're doing is we're focusing on, okay, let's ride this wave. I've already proposed reinforcement learning with heuristic imperatives, which is different from reinforcement learning with human feedback. Because human feedback aligns models to what humans want, which what humans want and what humans need are often very, very different. Here, it's to comparatives, which is not just what humans want, but what all life needs. We're also talking about data set curation and inner alignment problems, like Mesa optimization.
I applaud the work that Shapiro and others have put into developing this framework. It's a substantive task. I do though have concerns that it is all overly optimistic and too many major issues are skipped over in this presentation, most of which I'd guess Shapiro is aware of. The presentation of the layers of the proposed framework, for instance, gets shakier the further down the seven layers we progress.
Having developed policy at a local, national and international context I'm very aware of the significant obstacles that lay ahead, especially given the critically acute timeframes this would need to be discussed, agreed upon and enacted. One only needs to look at the failures of successive UN Conference Of the Parties to comprehend this. It is though the best set of thought through proposals I've so far come across, and the video is well worth serious consideration. Gaining sufficient traction, in time to affect meaningful change, remains a planet sized obstacle.
In a May 2022 paper by Korinek and Balwit, entitled 'Aligned with Whom? Direct and social goals for AI systems' they conclude:
As AI systems have become more powerful and their use in our world has become more widespread in recent years, we have also witnessed a growing number of cases of social alignment failures, from automated decision systems with biases against disadvantaged groups to social networks that increase polarization and undermine our political systems. Yet progress is continuing, and the powers of our AI systems are continuing to evolve. This makes it urgent to accelerate our efforts to better address the social alignment of AI. If we already have difficulty addressing the AI alignment problems we face now, how can we hope to do so in the future when the powers of our AI systems have advanced by another order of magnitude? Creating the right governance institutions to address the social AI alignment problem is therefore one of the most urgent challenges of our time.
I can only concur with their findings.
Comments
Post a Comment