AI: Medically trained models, consciousness

Most people can perhaps name a handful of LLM models, maybe a dozen if they follow the tech news. However, meanwhile about 3 million different LLM models are freely available from Hugging Face  . And not all of them are low-quality models. E.g. Google’s Open Source Gemma models are also hosted there; they share a common genesis with Google’s Gemini models, but have been optimized for local and research use. Importantly for biomedical researchers, there are also many open, freely available, and locally usable, medically trained LLMs. I have not tested them, but there is a curated list of specialty models on GitHub: https://github.com/FreedomIntelligence/Awesome-Specialized-Medical-LLMs  . The list also contains several freely available general medical models like

• Google’s MedGemma: https://huggingface.co/google/medgemma-27b-it 
• Me-LLaMA & PMC-LLaMA: https://github.com/BIDS-Xu-Lab/Me-LLaMA 

Alex Bunker and myself presented at a recent lunchon seminar how we use AI in our work. It was clear that the audience was very interested in the topic, because seldomly were there so many questions asked and the discussion about the many different aspects of AI continued beyond the event.

If you have ever done any serious work with AI, you know that it is almost unbelievanĺe that such sophistication can be achieved what is essentially merely a statistical prediction of the next word. Hence, it is not surprising that some people (who do not know the technological underpinnings) are convinced that LLMs are conscious. While Alex was very clearly denying that LLMs could be conscious (which is in line with the overwhelming majority of AI researchers), I was a bit more controversial:

I defended Richard Dawkins, but not because I agree with him that current AI is close to anything resembling human consciousness. But I agree with him that in the discussion about AI and consciousness, nuance is missing. Even my favorite podcast  was treating the topic of AI and consciousness in an overly simple manner. Here are the points that IMHO have not been receiving enough attention:

• Consciousness is not binary. Most people would agree that humans are conscious. There is also ample evidence for consciousness in animals ( https://doi.org/10.1016/j.tics.2020.07.007  ). But animals are a good example that consciousness exists on a spectrum. It is difficult or impossible to define an arbitrary point on that spectrum where consciousness starts or ends. Thus, I argue, it is difficult to conclude that a complex computational model cannot have ANY consciousness at all. Especially at the low levels, consciousness would be very difficult to reliably detect or to rule out.

• The second aspect that I am missing is the acknowledgment that we have only one datapoint for consciousness. If we stretch it, we could count octopuses and birds as separate, but all three brains (invertebrate, avian and hominid) have a common origin with common biochemistry that probably restricts possible developmental trajectories from a very large number to a still large but significantly smaller number. We implicitly assume that consciousness must be human-like and manifest itself in similar ways as human consciousness does. This is IMHO too simplistic and overly anthropocentric. Most people would agree that octopuses show behavior consistent with consciousness. Octopuses have diverged long time ago from from the lineage leading to vertebrate consciousness, and that makes it arguably more difficult to interpret their behavior in terms of assigning consiousness. But behavior is all we can measure. E.g. we might assume that if a being (a biological organism or a complex computer program) is conscious, it would want to communicate this fact to us. But that’s mostly a reflection of our own need to communicate, most likely not something inherent to consciousness. We have the same problem with alien live forms. We have only one data point and therefore all that most of us can imagine how aliens would look like looks rather similar to us or another life form on our planet.

• Another big problem is that our knowledge about consciousness is still rudimentary. We know much more than we did 100 years ago, but experts are still arguing about many of its aspects and very nature. We know some of its prerequisites (e.g. what brain areas’ activities are needed to create it), but all of that that is limited to humans. If we knew exactly what consciousness is and how it emerges from more simple forms of information processing, we would have created conscious machines long ago.