Year-End Meetup: Which AGI Gaps Actually Narrowed in 2025?

At the beginning of 2025, our group discussed the gap between state-of-the-art AI systems and Artificial General Intelligence (AGI), focusing on topics such as reasoning, world models, continual learning, social intelligence, and embodiment.

Over the past year, we’ve seen many impressive systems, papers, and demos. In this year-end meetup, we want to take a step back and reflect:

Which of the AGI gaps we identified earlier this year actually narrowed in 2025 and which remain fundamentally open?

We will revisit the same gaps discussed earlier this year:

• Reasoning in domains without cheap or scalable external verification e.g. writing, policy making etc.
• Social and narrative intelligence e.g. theory of mind
• Continual learning and catastrophic forgetting
• World models
• Embodiment, self-models, and uncertainty / hallucinations

For each, we’ll ask whether anything meaningfully changed in 2025. Please see the full summary of that beginning-of-year discussion along with reference papers at the end of the event description.

Small Pre-Meetup Task (required)

To help improve discussion quality, we ask all attendees to do a very small task before coming:

- Please pick one of the AGI gaps listed above.
- Identify one concrete positive development from 2025 (a paper, system, blog post, demo, or personal experience) that would have most surprised your January-2025 self with respect to this gap.

This should take no more than 10–15 minutes. It does not require reading anything new, but rather reflecting on what you have already read or experienced this year.

When you register for the meetup, you'll be asked the following question.

Pick one AGI gap from the list and one 2025 development that would most positively surprise your January-2025 self. Why? (1–2 sentences)

Please make sure to answer the question meaningfully, as this is a requirement for attending the meetup. Please no negative answers e.g. nothing surprised me. If you don't find anything matching the gaps in the list, you may choose a different gap.

In earlier meetups, we have found that this small filter meaningfully improves the discussion quality.

Here are example answers so that you know what we are more-or-less looking for (please don't copy verbatim):

World models. Genie 3 surprised me.

Difficult to scale verifiable domains. GPT 5.2's ARC AGI 2 score surpised me positively.

Format

• Casual round-table discussion in a great venue
• No speaker
• Active participation encouraged

Practical Details

Language: English
Audience: People who follow AI and AI-adjacent research
Preparation: Only the small task mentioned above
Schedule:

• 18:30 Arrive and socialize (while enjoying pizza and drinks) 🍕🍻
• 19:00 Welcome and intro round 🤝
• 19:30 Discussion Kickoff 🗫

We’re looking forward to a lively and reflective end-of-year discussion with you.

Hope to see you there,
Dibya, Nico and Somayeh

Appendix: Full summary of beginning-of-years gaps discussion along with reference papers

1. Reasoning in domains without cheap/scalable external verification e.g. creative storywriting, government policy etc. etc.
- Potential solution discussed: Exploit human reward priors known to LLMs and do RL. Example paper: [ https://arxiv.org/abs/2303.00001](https://arxiv.org/abs/2303.00001) ; there are many others
2. Social and emotional intelligence; narrative understanding
- Papers mentioned that show progress on this problem: https://ai.meta.com/research/cicero/
- For narrative understanding, also see: https://arxiv.org/abs/2403.01061
3. Continual learning/catastrophic forgetting in DL
- Papers mentioned that show progress in this problem: https://arxiv.org/abs/2403.08763, https://omni-epic.vercel.app/
4. World model
- Papers mentioned that show progress on this problem: https://deepmind.google/discover/blog/genie-2-a-large-scale-foundation-world-model/, https://arxiv.org/abs/2301.04104
5. Embodiment in real or virtual environments, self-model (sentience), hallucinations (expressing uncertainty in self knowledge)
- Why hallucinations may not be a problem: https://www.youtube.com/watch?v=hhiLw5Q_UFg
- For embodiment, also see: https://auto-rt.github.io/