Creator Lisan al Gaib disputes claims that xAI is training a 10-trillion-parameter Grok model at Colossus 2

Cursor just announced a new 1.5+ trillion parameter model pre-trained on over 100k GPUs.

According to Cursor CEO it is: "as big as Opus and GPT"

Yes, the cat is out of the bag and I can spoil the party. Opus 4.5 to 4.8 and GPT-5 to GPT-5.5 are not that big! (they are all below 2T params)

This means that the current performance of GPT-5.5 and Opus 4.8 is achievable for open-source, because guess what else is that size? -> DeepSeek-V4-Pro

The only moat is scaling. So far, Anthropic is the only lab that successfully made the jump to the ~10T scale, which is also why I no longer see OpenAI catching up until year end. Anthropic can just keep throwing RL compute at Mythos for the next 1-2 years and it will keep improving.

Google went not quite as big and sparsity-maxxed a bit too much. They also have apparently no clue how to RL to make the model actually usable.

OpenAI is still scarred from GPT-4.5

xAI and Meta are still planning.