Everything Is Model[S] (www.morphllm.com)

Andrej Karpathy’s Software 2.0 vision—that neural nets replace traditional code—gets a revision: the future won’t be one giant model doing everything but an ecosystem of models optimized for different roles. The piece argues that frontier models (massive, generalist models) will handle novel reasoning and creative work, while inference‑optimized specialized models and tiny transformers will take over routine, high‑throughput tasks. The author uses Fast Apply—a specialized model that merges code edits at 10,500+ tokens/sec—as a concrete example of this “models all the way down” stack, and warns against the common nerd‑snipe of endlessly finetuning local models to beat a frontier model on narrow benchmarks only to be outpaced by the next frontier release. The significance is economic and architectural: using a 5‑trillion‑parameter frontier model to apply trivial code edits is wasteful in latency, cost and reliability. Cited figures compare a frontier model at ~$20/M token and ~100 tok/s versus Fast Apply at $1.20/M and 10,500 tok/s—roughly 70x cheaper and 45x faster for that task. The takeaway for AI/ML practitioners: design hierarchies—frontier models for novel problems, specialized models for domain patterns, tiny models for simple transforms—so systems are faster, cheaper and more robust than single‑model monoliths.