Practice · 2024–now · AI Present
Reasoning models for grep tasks
o1-class deliberation for tickets that needed a filter. Resume-driven inference — and, in September 2026, a "decision model" that was a classifier.
Why Reasoning models for grep tasks became a costly fad
Reasoning-model overkill failed as a default because expensive deliberation is not a substitute for clear specs and cheap tools. Frontier reasoning sticks on hard planning and math; it fades when every autocomplete call escalates to a thinking model. In September 2026 TypeSafe put Jev in early access and called it a new "System One" model: it does not write text, it scores a list of options you already supplied and returns a choice plus a probability. That is supervised classification, probability calibration — Platt scaling, temperature scaling, a Brier score, expected calibration error — and zero-shot label scoring. A constrained label cannot invent a sentence. It can still be the wrong one. The training name RLCD had no public paper, and the weights were unpublished in the launch week. Twelve days of attention is not an invoice. The fad was status; the stuck layer is model routing.
Cost of the fad
What Reasoning models for grep tasks cost
Thinking tokens burned on problems a SQL query and a unit test would settle. Latency and invoices grew; correctness did not.
Patterns
Sources
Compare with
Related
Architecture · 2024–now
Model routing / tiered inference
Cheap model drafts; frontier model escalates. The boring cost-control layer that survived the agent hype.
Framework · 2023–now
LLM app frameworks
LangChain-class glue mutated into MCP clients and thin wrappers around vendor APIs. The durable pieces are still boring: evals, retrieval, and product UX.
Practice · 2021–now
AI pair programming
Coding agents that edit the repo — now table stakes. Generation outran review capacity; ownership is the bottleneck.
Practice · 2023–now
RAG as default architecture
Retrieval-augmented generation as the answer to every knowledge problem — then "just stuff the window" as the counter-fad. Often right; often Postgres with pgvector would suffice.
$ Teams stood up vector pipelines, chunking strategies, and rerankers before asking if fine-tuning or a SQL query would answer the question. Retrieval infra became the product.