Post #1490357

@TeflonTrout I don't use US models, and if I did, it would only be to distil their outputs. My business helps people set up open weight models on whatever GPUs they rent (whether from hyperscalers or their own), and do distributed inference. The market was much better before big bastard labs got HIPAA and other certifications, but I can't complain, it's a fun challenge still. I'm old, so my background is in training Generative-Adversarial Networks to do unsupervised anomaly detection (think anomalous radiographies). Besides that, I use LLMs for what they're good at: fuzzing. I won't name names, for fear of litigious actors, but LLMs are great at reverse-engineering proprietary blobs that get in my way, LLMs are a great 'fitness' evaluator for evolving algorithms (instead of doing exhaustive search, you use an LLM to guide param tuning). My belief is that LLMs are here to stay, and the only way forward for us laypeople to retain some semblance of power isn't rejecting LLM use, but making them so commoditised that large-scale datacentres just become economically intractable. The OAI-Nvidia-Oracle-CoreWeave circlejerk investment is the proof that capitalism has transcended labour and "voting with your dollars" is powerless as direct action.