Thoughts on AI

August 10, 2023

I wanted to just put some of my thoughts down since playing with ChatGPT and exercise my brain to deliver a shortened version of some of the highlights that has been happening in the AI field as of late. To anyone working in the field this may feel like a magazine article or over simplified and non-exhaustive as it’s not meant to go into technical details or provide a full recap. Stanford has a decent State of AI 2023 available for reference as well as the IEEE.

Current Advancements

The past 5 years have seen a lot of advancements in the field of AI with hype from my media. The headlines have been nonstop with Tesla updates, and especially since ChatGPT rolled out in the past year. Large Language Model’s (LLMs) are all the rage now where much of the advancement has revolved around scaling up the transformer architecture. Many researchers are also working on multi modal models (like Gato) that assimilate text, video, audio etc data into one model. Doing a full recap of all the happenings in the field is beyond this article but a general timeline is below.

2018 Google BERT
2020 Deepmind’s AlphaFold wins CASP14 for protein folding problem¹
2020 OpenAI releases GPT3
2020 DARPA holds AlphaDogfight and AI system defeats F-16 Pilot²
2022 DALL-E released³
2022 Google Engineer claims LaMDA is sentient⁴
2022 Github releases Copilot to Developers
2022 OpenAI releases ChatGPT
2023 GPT4 Outperforms Medical Students on USMLE⁵

Most of the above advancements were with NLP and LLMs, but Computer Vision tasks have also advanced. CV has been a primary technology in the driver-less car product. Not only can we do image classification and object detection but now we can distinguish background/foreground objects with segmentation techniques. Machine learning in the driver-less cars like Tesla allow it to “see” the road and all the objects in the screen and perceive the environment to a greater extent. These were the rage before LLMs 5-10 years ago, but despite being promised they are right around the corner we still don’t have them on the road. Albeit a certain few designated routes from companies like Waymo. Wide spread adoption IMO is still decades away due to driving outliers/edge cases, perception in different conditions, safety, etc. We also face huge infrastructure and regulatory challenges if we want wide spread adoption. With all the hype in LLMs and Driver-less tech, I think we must keep in mind Moravec’s paradox, that is in general those human skills we find easy are extremely hard to reverse engineer since we are largely unconscious of them. I do not know what the future holds, but it seems to be that as we progress further into this territory the capabilities of the technology increases at a non-linear rate. Is there a physical limit to this advancement? One could extrapolate that thought and you end up with a machine that could potentially have more capabilities than a human.

AGI

The term Artificial General Intelligence has been used to distinguish between the AI of old & recent that is built for a very specific/narrow task versus a future AI that could be used to do many general tasks. I’m not partial to a lot of the terms used in this field because they bring about a line of thinking that anthropomorphize the technology. I think this makes believe certain things about the machine that are not necessarily true. E.g. because ChatGPT formulates very good answers we believe there is reasoning behind it, but actually there doesn’t appear to be any reasoning behind the transformer architecture, it’s just predicting the next best word. Like it or not, this term has stuck and people are using it in everyday talk. Despite the short comings of ChatGPT the models today are stronger versions of AI that we created 20 years ago and do display some generalization power. This will continue and I believe the models we create in the future will be able to generalize to new tasks better than what we have now. To me this seems inevitable. Whether or not the transformer architecture brings us to AGI I am not sure. It could be one component in a bigger architecture and lead us to AGI or could be a red herring and lead us down a wrong path.

AI Safety

With all the talk of an AGI or Super-intelligence in the future one cannot help but think of the implications of this kind of technology. It would be immensely powerful (analogous to atomic fission) and could be harnessed for good or if misused, could be used for more nefarious purposes. With this potential risk there is more talk around existential risk in regards to AI. It seems hard to imagine that 20-30 years ago this talk was probably just viewed as sci-fi nonsense talk, but now it seems to be gaining ground. The x-risk camp believe that AI will bring about an existential risk for humans and we must solve the problem now given the rate of progress before it is too late. Many people have made bold claims that Super-intelligence is near, however, I’m a bit more reserved and my general consensus is that the field has once again been swept up in hysteria and the outrageous claims continue to feed the fire. The mechanism about how AI leads to existential risk is the key issue for me. Luckily there are folks in the growing field of AI Safety that are doing a lot of work on multiple fronts that would ensure that this technology is safe from a technical standpoint. My general consensus around this is that AI could pose an existential risk but not in the near term. I don’t see strong arguments on the side that AGI or some form of super-intelligence will be with us in < 5 years, which to me is the near term. I think we should focus on immediate concerns in current technology but also prioritize research on AI Safety for the medium and longer term risks. The IEEE posted a nice scorecard for various leaders in the field and their opinions on the AI risk.

Governance

As well as the potential long term risk that AI poses it also poses some challenging short and medium term risks (algorithmic bias, manipulation, autonomous weapons, adversarial attacks, etc). Many believe that Government needs to play a role in regulating these types of risks and there has recently a big push to get an agency up and running. In the past 6 months or so we have had multiple senate hearings detailing that the industry needs Government regulation to help steer the technology in the right direction. The hearing in May included Sam Altman from OpenAI and Gary Marcus from NYU, and Christina Montgomery from IBM. The latest hearing in July included Stuart Russell, Yoshua Bengio and Dario Amodei. I listened to the latter and it sounded like both sides agreed that regulation should be done it was really about how they should go about doing that. I think these are all important steps to take and should be done. I also believe that there must be international cooperation as to avoid an AI arms race.

Closing Thoughts

In closing, I think this field will continue to grow as more and more people get involved and more and more people utilize the technology. There is already billions of dollars spent on R&D and it seems too big to fail at this point. We must minimize the hype and be clear about the benefits and harms this technology could bring though so humanity can adapt. I think we need to keep in mind Amara’s law which states,

“We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”

I also think we need more fundamental research in multiple areas like Neuroscience, Psychology, Cognitive Science, and AI and have interdisciplinary teams working on all the different issues. We also shouldn’t be dismissive and should listen to what others are saying to really understand their point of view. Hopefully then some consensus can be reached among the community.

Sources

Lee