the authors reference to LLM's as "bullshit machines" is more true the less parameters you have trained in your model....as we scale up to trillions of parameters, add Mixture of Experts (MoE) architecture, this no longer is an accurate statement. Proof in point was yesterdays announcemnt of Mythos 5 model (10T parameters + MoE [1]) by anthropic where it seems to be so good at finding/exploiting vulnerabilities in source code that have been there for decades and only recently uncovered needs to be used to fix these critical vilnerabilities first before it gets released to the public, they even have a project called Glasswing [2] dedicated to letting people fix the thousands of vulnerabilities already found by the model before they release this model to the public, because it's so good at what it does... I think we're a little bit past the point of calling these models "bullshit machines" at this point...

[1] https://www.aimagicx.com/blog/claude-mythos-5-trillion-param...

[2] https://www.anthropic.com/glasswing