I tried to find some baselines for each of the 4 criteria but disregarded them because it left me feeling like all of them are already achievable in late 2022, i.e. even if there are no breakthroughs in AI and no hardware progress and we will give the goal to make this question resolve positively to DeepMind, they will surely do it in 2 years, probably much less. The easiest way to achieve this would be to make not truly generalized agents but just squash 2-4 different agents together. However, Gato shows that generalized agents are not so much worse. ...

Transnistria is mentioned on the map Lukashenko showed at the security council meeting recently. There are red arrows leading to it.

[A Generalist Neural Algorithmic Learner](https://arxiv.org/pdf/2209.11142.pdf) > The cornerstone of neural algorithmic reasoning is the ability to solve algorithmic tasks, especially in a way that generalises out of distribution. While recent years have seen a surge in methodological improvements in this area, they mostly focused on building specialist models. Specialist models are capable of learning to neurally execute either only one algorithm or a collection of algorithms with identical control-flow backbone. Here, instead, we focus on constructing...

New Google paper reaches 75.2.

Deputies in 18 districts of Moscow and St Petersburg have signed a statement demanding that Putin resign.

I think now it's around 30.

I think the community somewhat overupdated on Gato article. It is very likely that it is soon will be possible to create such a model (if not possible now), but I think most such models will not be tested on all 4 conditions in the question.

So it is quite possible that the question will be resolved by some public model which will be available significantly later (my median is 2029-2030).

— edited by qumidium

@Joker that's good data but 1% is clearly overconfident. You mentioned 16 cases. It would be 17 with Kiev. 1/17 is more like 6%, not 1%. Perhaps my math is not quite correct but I am pretty sure correct math would give more than 1%.

Also, I am not sure about other situations, but there are some signs which may point to faster action here, for example, Putin clearly wants it to be captured as fast as possible, and also Russian army is objectively much stronger than Ukranian (despite its poor performance).

— edited by qumidium

I wonder why it moved from 60% to 32% in 2021?

New SotA neural prover from Meta solves 10 IMO problems and reaches 67% accuracy on minif2f.

"The AI must be open-source, publicly released before the IMO begins, and be easily reproducable." -- this is a very strong criterion which can delay resolution for years.

Human-level Atari 200x faster -- seems like we moved from "billions" to "hundreds of millions".

@PhilippSchoenegger I am a) Russian speaker b) Have seen a full 20 minute video of Lukashenko with this map

I confirm it is true

Ukrainian interior ministry reports that the Russians have taken Gostomel

@fianxu When creating the question, I meant only government restrictions, not YouTube's own actions. The wording might be improved though.

@kokotajlod Maybe not 100% trustworthy but it's generally a reliable source https://twitter.com/FirstSquawk/status/158164…

40% -> 50% because of the recent Transnistria foreign ministry statement (Google Translate): UPD: this was 2 days ago and was already mentioned here (I didn't notice at first), still leaving full text here Statement of the MFA of Transnistria in connection with the application of the Republic of Moldova to join the EU We regard the submission by the leadership of the Republic of Moldova of an application for membership in the European Union as a geopolitical decision leading to a change in international borders and spheres of influence in the regional...

Flan-PaLM 540B reaches 75.2 on MMLU.

— edited by qumidium

From [OpenAI's plans according to Sam Altman](https://humanloop.com/blog/openai-plans): # The scaling laws still hold Recently many articles have claimed that “the age of giant AI Models is already over”. This wasn’t an accurate representation of what was meant. OpenAI’s internal data suggests the scaling laws for model performance continue to hold and making models larger will continue to yield performance. The rate of scaling can’t be maintained because OpenAI had made models millions of times bigger in just a few years and doing that going forward w...

Ukraine’s Defense Ministry: Russia has sustained an estimated 800 fatalities in combat.

YouTube blocked channels of one of the main Russian propagandists, Soloviev.