The 12T Models and the 'Northern Myanmar Compound': This Silicon Valley Leak Has Commenters Skeptical First
The leak claims Google trained multiple 12T models at the same time, while Anthropic uses a high-cost approach to prevent distillation. The comment section argues from parameter scale to source reliability; almost every detail is disputed by someone.
A tweet from @lee_xeo spilled details about Google's and Anthropic's training. It carries a lot of information and no small amount of controversy. The original link is at the end.
First, the leak itself.
Gemini Argon and Gemini Carbon are made by two separate teams; the model architecture is exactly the same, but the data and training methods differ. The original writes 12TA555B without explaining it further. Two teams building models with the same architecture is similar to the 'horse racing' approach used domestically in China. Google very likely trained multiple 12T models at the same time, and two ultimately emerged. The tweet also judges that the compute gap between China and the U.S. is far greater than two orders of magnitude.
Google currently has no stronger model internally, but Anthropic does.
Anthropic's mode of operation was described as being like the KK Park in northern Myanmar. It hires many people, who go in and participate in a project, but the participants don't know whether the project they are working on ultimately generates data for the model or is part of the model itself. R&D costs are therefore very high, and Anthropic also doesn't want others to distill it. Distillation usually refers to using the outputs of a strong model to train a small model, reducing training costs.
Anthropic's model should be smaller than Google's, and Opus is also smaller than Gemini 4. The tweet says it didn't find out exactly how large.
The comment section did not follow the leak and run with it; it started by taking the numbers apart.
Some commenters did not believe the difference between the two 12T models would be as large as the original text said. Others focused on the amount of knowledge: if Carbon really is 12T, why doesn't its world-knowledge performance look like that of a 12T model, and instead it gets crushed by Fable 5.2. According to this commenter's back-calculation, Fable 5.2 should be at the 20T level. Another commenter added that Hydrogen was distilled from Carbon.
Parameter scale was also repeatedly questioned. '12T? How could it be that big?' 'km3.1's parameters are only 3 trillion; a 4x parameter increase could not possibly produce only these results.' Someone else said outright that 1, 2, and 4 must be wrong.
The source did not escape scrutiny either. Someone wrote, 'Classic: first, “I have a friend,” and then everything after that is actually made up by the person themselves.' Someone asked, 'Is this melon guaranteed ripe?'—a sarcastic way of asking whether the tip can be trusted. Others used sarcasm, saying that the Google lead and the chief engineer of Anso Pika had both sent the leaker WeChat messages.
One reply provided another detail: Google has another internal model that has not been released, positioned between Argon and Barium. Another reply explained the horse-racing approach more concretely: if the source is reliable, horse-racing may catch things that small-scale experiments miss; a training recipe that is effective at small scale may not continue to lead after compute is scaled up.
None of these claims has official confirmation yet. What can be confirmed is that the information gap in the large-model race has become so large that leaks, counter-leaks, and parameter speculation are all mixed together. When reading this kind of news, read the main text and the comment section together for a more complete picture.
Source: Original tweet
发布时间: 2026-11-16 04:26