Basaltlabs Monolith-1.0 – #1 on Last Exam, AIME, GPQA Diamond, MMLU-Pro

Generally, I am very sceptical when it comes to public eval results, as I have in the past seen labs verifiably train for specific benchmarks [0].

Having read the technical report [1] however, I am absolutely certain that this is not the case here as they took such great lengths to prevent any training data contamination and due to this I bestow upon Monolith-1.0 the honour of being the first model I have such faith in that I do not see a reason to run my own benchmark suite. It is clearly Super AGI.

Here is a Pelican on a Bicycle to assess that aspect of the model, seems slightly ahead of Geminis output to me: https://imgur.com/a/9brUIow

/s for clarity.

Here the video [2] that covers how they did this along with the insanity that is the uncritical social media reporting and ease of creating hype via benchmark numbers, something repeatedly and provably used to great effect by some actual labs. Labs have done this and they will continue doing this as long as the incentives outweigh those trying to hold them to account.

[0] https://news.ycombinator.com/item?id=48951229

[1] https://basaltlabs.org/Monolith-1.0-2606.pdf

[2] https://www.youtube.com/watch?v=enk4w5mRjQY