Unlike simple classification (one pass), generative AI performs recursive inference. Each new token (word, pixel) requires a full pass through the model, turning a single prompt into a series of demanding computations. This makes inference a major, ongoing driver of GPU demand, rivaling training.

A core challenge in physical AI is the tension between large, powerful models (offboard, in a data center) and the need for low-latency models (onboard, on the machine). The key is using techniques like distillation to create smaller derivatives that run in milliseconds for safety-critical decisions.

Top inference frameworks separate the prefill stage (ingesting the prompt, often compute-bound) from the decode stage (generating tokens, often memory-bound). This disaggregation allows for specialized hardware pools and scheduling for each phase, boosting overall efficiency and throughput.

The pace of AI development is so rapid that a complex inference task assigned to a model could take longer to complete than the time it takes to train and release the next, more powerful version of that same model. This highlights an emerging paradox in the deployment of large-scale AI.

For physical design, simulation shouldn't just be a final verification step. Instead, it should be a tool used during model training to build the AI's intuition or "taste." This allows the model to generate high-quality designs quickly at inference time, mirroring how expert human engineers develop their skills.

Optimizing transformer inference, specifically the separation of pre-fill (KV cache building) and decode (token generation), is becoming a foundational skill. Chris Fregly predicts this complex topic, known as disaggregated pre-fill decode, will be a core component of AI engineering interviews at top labs within two years.

AI development is inefficiently split into pre-training (optimizing for compression) and RL (optimizing for tasks), where RL often invalidates pre-training metrics. Combining these into a unified, end-to-end learning algorithm focused on final outcomes could yield an order-of-magnitude improvement in training efficiency.

Inference engineering is not a monolith. Data center teams focus on making models "less slow" for massive throughput. Local AI teams focus on making models "less dumb" on constrained hardware, using methods like advanced quantization to fit models in memory.

A fundamental constraint today is that the model architecture used for training must be the same as the one used for inference. Future breakthroughs could come from lifting this constraint. This would allow for specialized models: one optimized for compute-intensive training and another for memory-intensive serving.