567 posts tagged with "llm"

Here is a pattern that appears in almost every production AI deployment: the team ships a feature that handles 90% of queries well. Then they start getting complaints. A user asked something outside the training distribution; the model confidently produced a wrong answer. A RAG pipeline retrieved a stale document; the model answered as though it were current. A legal query hit an edge case the prompt didn't cover; the model speculated its way through it. The fix, in each case, wasn't a better model. It was teaching the system to say "I don't know."

Abstention — the principled decision to not answer — is one of the hardest and most undervalued capabilities in AI system design. Virtually all product effort goes toward making answers better. Almost none goes toward making the system reliably know when to withhold one. That asymmetry is a design debt that compounds in production.

By 2025, every major LLM provider had shipped constrained decoding for structured outputs. OpenAI, Anthropic, Gemini, Mistral — they all let you hand the model a JSON schema and guarantee it comes back structurally intact. Teams adopted this and breathed a collective sigh of relief. Parsing errors disappeared. Retry loops shrank. Dashboards turned green.

Then the subtle failures started.

A sentiment classifier locked in at 0.99 confidence on every input — gibberish included — for two weeks before anyone noticed. A credit risk agent returned valid JSON approving a loan application that should have been declined, with a risk score fifty points too high. A financial pipeline coerced "$500,000" (a string, technically schema-valid) down to zero in an integer field, corrupting six weeks of risk calculations. Every one of these failures passed schema validation cleanly.

The lesson: structural validity is necessary, not sufficient. You need a semantic validation layer, and most teams don't have one.

Your team spent three days iterating on a system prompt. The eval score went from 82% to 85%. You ship it. Three weeks later, production metrics are flat. What happened?

The short answer: your eval lied to you. Not through malice, but through insufficient sample size and ignored variance. A 3-point accuracy lift on a 100-example test set is well within the noise floor of most LLM systems. You cannot tell signal from randomness at that scale — but almost no one does the math to verify this before acting on results.

This is the statistical power problem in LLM evaluation, and it is quietly corrupting the iteration loops of most teams building AI products.

Every fine-tuning effort eventually hits the same intuition: better data means better models, and better data means higher-quality examples. So teams build elaborate annotation pipelines to filter out the mediocre outputs, keep only the gold-standard responses, and train on a dataset they're proud of. The resulting model then underperforms on the exact use cases that motivated the project. This failure is so common it deserves a name: the curriculum trap.

The trap is this — curating only your best, most confident, most authoritative outputs doesn't teach the model to be better. It teaches the model to perform confidence regardless of whether confidence is warranted. You produce something that looks impressive in demos and falls apart in production, because production is full of the messy edge cases your curation process systematically excluded.

Most AI product post-mortems focus on the same story: the model was wrong, users noticed, trust eroded. The fix is obvious — improve accuracy. But there is a more insidious failure mode that post-mortems rarely capture because standard accuracy metrics don't surface it: the model was right, but for the wrong reasons, and the power users who checked the reasoning never came back.

Call it the overclaiming trap. It is the failure mode where correct final answers are backed by fabricated, retrofitted, or structurally unsound reasoning chains. It is more dangerous than ordinary wrongness because it looks like success until your most sophisticated users start quietly leaving.

A team ships a JSON extraction pipeline. It works perfectly in development: 98% accuracy, clean structured output, predictable token counts. They push to production. The model starts hallucinating extra whitespace, the JSON parser chokes on malformed keys, and the API bill is 2.3x what the prototype suggested. The model hasn't changed. The prompts haven't changed.

The tokenizer changed — or more precisely, their assumptions about it were wrong from the start.

Tokenization is the first transformation your input undergoes and the last one engineers think about when debugging. Most teams treat it as a solved problem: text goes in, tokens come out, the model does its thing. But Byte Pair Encoding (BPE), the tokenization algorithm behind most production LLMs, makes decisions that cascade through structured output generation, prefix caching, cost estimation, and multilingual deployment in ways that are entirely predictable once you know to look.

Six months after your best prompt engineer rotates off to a new project, a customer-facing AI feature starts misbehaving. Response quality has degraded, the output format occasionally breaks, and there's a subtle but persistent tone problem you can't quite name. You open the prompt file. It's 800 words of natural language. There's no changelog, no comments, no test cases. The person who wrote it knew exactly why every phrase was there. That knowledge is gone.

This is the prompt archaeology problem, and it's already costing teams real money. A national mortgage lender recently traced an 18% accuracy drop in document classification to a single sentence added to a prompt three weeks earlier during what someone labeled "routine workflow optimization." Two weeks of investigation, approximately $340,000 in operational losses. The author of that change had already moved on.

Fine-tuning a language model gives you a competitive edge until the provider updates the base model underneath your adapter. At that point, one of two things happens: your service crashes with a shape mismatch error, or — far more dangerously — it silently starts returning degraded outputs while your monitoring shows nothing unusual. Most teams discover the second scenario only when users start complaining that "the AI got dumber."

This is the adapter compatibility cliff. You trained a LoRA adapter on model version N. The provider shipped version N+1. Your adapter is now running on a foundation it was never designed for, and there is no migration path.

There's a belief embedded in most RAG architectures that goes something like this: if retrieval returns the right chunks, the LLM will produce correct answers. Teams invest heavily in embedding model selection, hybrid retrieval strategies, and reranking pipelines. Then, three months after deploying to production, answer quality quietly degrades — not because the model changed, not because query patterns shifted dramatically, but because the underlying corpus rotted.

Enterprise RAG implementations fail at a roughly 40% rate, and the failure mode that practitioners underestimate most isn't hallucination or poor retrieval recall. It's document quality. One analysis found that a single implementation improved search accuracy from 62% to 89% by introducing document quality scoring — with no changes to the embedding model or retrieval algorithm. The corpus was the variable. The corpus was always the variable.

Andrej Karpathy put it bluntly: AI labs were "overfitting" to Arena rankings. One major lab privately evaluated 27 model variants before their public release, publishing only the top performer. Researchers estimated that selective submission alone could artificially inflate leaderboard scores by up to 112%. The crowdsourced evaluation system that everyone pointed to as ground truth had become a target — and once it became a target, it stopped being a useful measure.

This is Goodhart's Law in action: when a measure becomes a target, it ceases to be a good measure. It's been well-understood in economics and policy for decades. In LLM engineering, it's actively destroying eval suites right now, often without the teams building them realizing it.

Most GPU clusters running LLM inference are wasting between 30% and 50% of their available compute. Not because engineers are careless, but because the scheduling problem is genuinely hard—and the tools most teams reach for first were never designed for it.

The standard approach is to stand up Kubernetes, request whole GPUs per pod, and let the scheduler figure it out. This works fine for training jobs. For inference across a heterogeneous set of models, it quietly destroys utilization. A cluster running three different 7B models with sporadic traffic will find each GPU busy less than 15% of the time, while remaining fully "allocated" and refusing to schedule new work.

The root cause is a mismatch between how Kubernetes thinks about GPUs and what LLM inference actually requires.

Your agent passes every unit test, handles the happy path beautifully, and then one Tuesday afternoon it tries to call get_user_preferences_v2 — a function that has never existed in your codebase. The call looks syntactically perfect. The parameters are reasonable. The only problem: your agent fabricated the entire thing.

This is the phantom tool call — a hallucination that doesn't manifest as wrong text but as a wrong action. Unlike a hallucinated fact that a human might catch during review, a phantom tool call hits your runtime, throws a cryptic ToolNotFoundError, and derails a multi-step workflow that was otherwise running fine.