I am very curious about Anthropic's strategy right now; they launched with "More progressive and careful about AI risks," which in my opinion is a good AI worker recruiting strategy, but not a good commercial strategy, or even a good quality strategy, based on Claude's current state.
They have done some outstanding research this year -- in particular, I loved their work on tracking conceptual bundles through different layers, and I'd say their paper on embedding adversarial behavior in LLMs was my favorite paper of the year -- But, I'd like to see that research turn into more capable models.
As I write this, I wonder what I'd do if I were in charge there. I think I'd probably try and one-up Mistral, and release larger more capable models more quickly. I'd try and buy Groq and build a custom vertically integrated stack that's 5x faster in tokens/s; worst case Amazon could help with the acquisition and deploy Grok into AWS. Basically the moat would be deployed tokens/s at or above certain levels of quality. I'd stop running anything but my own models on Groq hardware.
There are two next-gen directions for inference in the cloud, high quality multi-modal foundation models, which are definitely coming, but also clearly need some UI work to get into a compelling product format, and very long context length models, so I guess I'd possibly be trying to leapfrog there. It's highly competitive in the model space right now, so a model-only strategy isn't good.
Alternatively, I could sit back and relax and let AWS pimp Claude while we just scaled it up to 250k tokens and used MoE to get non-exponential scaling / possibly explore an SSM / Mamba architecture, and not worry too much since we have 4bn in the bank. I wouldn't be surprised if that latter path is the one they're on, sadly.