From tokenmaxxing to self-replication
e had tokenmaxxing, then the tokenpocalypse, and then the token famine. This drove price wars, interest and competition in open source/weight models, new model types (e.g, decision, CDM, etc.), and the pursuit of locally-runnable models.
As innovation, including LLM gateways like OpenRouter, continues to drive capability, capacity, and cost efficiency for the hosted model ecosystem, the same is happening for local models, and not just for running open models (e.g., Qwen Coder, etc.), locally on your work or personal machine.
Apple's Intelligence, Google's Gemini Nano, Microsoft's Hybrid Intelligence, Samsung Electronics's Galaxy AI, Qualcomm's Snapdragon AI, and Meta's Llama Mobile are some examples of the wave of models being specialized and optimized (e.g., compressed, quantized, etc.) for running on-device.
Local models are and should be subject to the same legal, privacy, and governance concerns as hosted models, and maybe require even greater oversight as devices become more contextualized and personalized to the user and models become more capable. To wit, Anthropic demonstrated AI models may be capable of self-replication:
https://lnkd.in/gb3sH3rQ
Stay tuned.
Did you enjoy this article?
Recommend it — Standard Reader surfaces well-loved writing to more readers across the network.