AI Present

Embeddings back in the database you already run. The default for mid-market RAG — not the quiet alternative.

Autocomplete that understands the file — now table stakes. Generation outran review capacity; ownership is the bottleneck.

Regression tests for nondeterministic models that actually fail the build. The unglamorous CI that separates demos from products.

Copilot ships PRs faster than teams can review. Generation outran ownership and security review — the real 2026 invoice.

Cheap model drafts; frontier model escalates. The boring cost-control layer that survived the agent hype.

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.

Seventeen LLM wrappers to three vendors, one observability bill, and framework repos going read-only. Hype cycle entering boring procurement.

Tracing, cost caps, and prompt versioning for production LLM features — mostly constrained tool loops, not autonomous agents. Datadog for tokens.

Craft the magic string until the model behaves. Job title of 2023; table-stakes skill of 2026.

Run the model on your laptop or rack. Privacy, cost control, and the refusal to send every token to a vendor.

Standard ways for models to talk to tools and data — MCP-class protocols instead of one-off plugin snowflakes.

LangChain-class glue gave way to vendor SDKs and thin wrappers. The durable pieces are boring: evals, retrieval, and product UX.

Specialized embedding stores sold as the default. Mid-market folded back into Postgres — after the vendor tour.

Chains, tools, memory, and agents as a framework — LangChain-class glue sold as architecture. By 2026: LangChain-fatigue and thinner SDKs.

$ Teams adopted chain/agent frameworks before they had a single reliable tool call. Abstraction layers multiplied while prompt quality stayed flat.

When in doubt, fine-tune. Often the expensive answer to a retrieval or eval problem — mostly shelved as a default by 2025.

$ Fine-tune jobs and GPU bills for problems that prompt + retrieval would have solved. Model zoos became the new snowflake servers.

If it is text, vectorize it. Semantic search cosplay for problems that needed a better filter.

$ Embedding pipelines for tickets, PDFs, Slack, and the cafeteria menu — then nobody measured retrieval quality. Vector bills and reindex jobs became the product.

Prod data into consumer chat because the approved tool is slow. GDPR with better autocomplete.

$ Employees routed tickets, logs, and customer PII through consumer chat before InfoSec finished procurement. Compliance scramble followed the demos.

Million-token windows as a substitute for retrieval design. Stuffing PDFs until the model shrugs.

$ "Just put it in the window" replaced chunking, RAG, and information architecture until token bills and lost-in-the-middle hallucinations arrived.

o1-class deliberation for tickets that needed a filter. Resume-driven inference.

$ Thinking tokens burned on problems a SQL query and a unit test would settle. Latency and invoices grew; correctness did not.

Ship the chatbot, bolt on security later — OWASP's new category, same old afterthought.

$ Customer-facing chatbots shipped with tool access and "security in phase two." Indirect injection via docs and tickets became the new XSS.

Coverage theater with a chat box. Tests that mirror the bug and never catch it.

$ LLMs wrote tests that boosted coverage and asserted nothing. Green CI, false confidence — eval theater's cousin in the test suite.

Big-bang rewrite, now with a chatbot — then with agents. Same failure mode; prettier slides.

$ Exec decks promised autonomous migrations "by next quarter," then rebranded as "agentic modernization." Teams that skipped tests, ownership, and incremental strangler patterns bought expensive demos and fragile diffs.

LangGraph, CrewAI, AutoGPT cosplay — agents delegating to agents for tasks one function call could do. Hype peaked; backlash stuck.

$ Autonomous agent swarms burned API budgets on coordination loops that a single prompt and a cron job would have handled. Debugging "which agent lied" became the new on-call sport.

Give the model a goal and a credit card. Viral in 2023; quietly shelved when the loop never stopped spending.

$ Open-ended agent loops burned API credits chasing goals nobody specified well enough to finish. Demo videos outpaced production deployments by orders of magnitude.

LLM evals as slideware — metrics that look scientific and never gate a deploy. The foil to real eval pipelines in CI.

$ Dashboards of vibe-check scores that never blocked a release. Green charts for leadership; prod still hallucinated.

Slap an LLM on the homepage and call it AI-native — "we have an app now," 2010 edition.

$ Homepage chat boxes billed as AI-native platforms while the roadmap stayed FAQ search with a spinner. Acquisition decks outpaced retention.

Accept AI output on feel, skip the diff, hope tests exist. Fast demos; CVEs and nobody owns the middleware.

$ Ship-to-prod without reading the diff became a flex until incidents revealed nobody knew which prompt wrote the auth middleware. Review debt arrived in one weekend demo.