Magic

Magic is building frontier code models and large-scale compute infrastructure, but public evidence of broad production customer deployment remains limited.

Magic's 2024 research update describes HashHop: random incompressible hash-pair retrieval with shuffled order and multi-hop chains to test long-context reasoning without semantic needle cues. This specific evaluation design adds technical evidence beyond generic long-context and pretraining claims.

Magic combines frontier pretraining, domain-specific reinforcement learning, ultra-long context and inference-time compute, and reports more than 10x compute-efficient pretraining versus leading open-weight baselines.

Magic combines frontier pretraining, domain-specific reinforcement learning, ultra-long context and inference-time compute, and reports more than 10x compute-efficient pretraining versus leading open-weight baselines.

Magic's August 2024 technical update identified a 23-person team and systems/supercomputing hiring alongside Ben Chess, described as former OpenAI Supercomputing Lead, showing a specific technical organization beyond its compute and fundraising announcements.

Magic reports thousands of GB200s and $768M raised from major technology investors to support frontier-scale model research.

Magic reports thousands of GB200s and $768M raised from major technology investors to support frontier-scale model research.

Magic reports building its own training and inference stack with custom CUDA and without PyTorch autograd to support ultra-long contexts. This concrete systems work supports an analyst inference of portable kernel, training-systems and inference engineering skills.

Frontier model research, coding agents, RL, pretraining and long-context skills transfer exceptionally broadly across frontier AI labs and developer AI companies.

Frontier model research, coding agents, RL, pretraining and long-context skills transfer exceptionally broadly across frontier AI labs and developer AI companies.