What is the Current Gap Between SOTA AI and AGI?

2024 was a year of stunning advances in AI capabilities, especially in the domain of reasoning. Deepmind's AlphaProof performed at the level of a silver medalist in the International Maths Olympiad. In December, OpenAI O3 achieved human level performance in the ARC challenge and made a step jump in the Frontier Math benchmark. Reasoning is the crown of human intellect and the gap between AI and human capabilities seems to be shrinking rapidly.

We thought this would a good time to take stock and re-evaluate our assumptions and assessments regarding the gap between SOTA AI and AGI. It's also a great first topic for the new year!

If that sounds interesting, please join us for a discussion on the topic "What is the current gap between SOTA AI and AGI?" on January, 10th. The idea is that everyone comes prepared with one gap that they don't believe current methods can address fully. Then we discuss in the group whether that viewpoint is defensible. This will be a very low pressure, chilled out discussion. However a minimal level of preparation and some facts/references to support your viewpoint will be very appreciated by the others.

Here is the rough plan for the evening

🍕🍻 18:30 - 19:00: Arrival and networking with pizza and drinks
🤝 19:00 - 19:30: Introduction
🗫 19:30 - 21:00: Discussions

To get you started on looking for gaps, here is a potential starting list

1. The energy efficiency gap between AI and human brains, but see also this talk by Hinton about why this might not matter.
2. The continual learning problem
3. The problem that LLMs do not have agency. For a good essay on the connection of agency and AGI, see David Silver's opinion piece Reward is Enough.
4. Physical intelligence i.e. motor planning and co-ordination.
5. Social intelligence or mentalizing. Lack of feelings/consciousness and the limitation that imposes on art and creativity.
6. Abstraction and reasoning, but this seems close to getting solved. It is worth considering if the current approaches will get us all the way to the point where AI can make scientific discoveries.
7. Unavailability of rich enough context / embodiment. This is probably not a capabilities gap, but rather an engineering gap.

Anyway, hope to see you there and have a great discussion! 🤗

Cheers!
- Alexandra, Dibya, Mainak and Nico