778 posts tagged with "llm"

A verifier that flips its own answer eight percent of the time is not a flaky model. It is a sampling configuration bug that reached production because the framework defaulted to inheritance. The planner needed temperature=0.7 to brainstorm subtask decompositions. The verifier — the role whose entire job is to give a low-variance yes-or-no on whether the answer satisfies the rubric — was instantiated through the same harness call, and silently picked up the same temperature. Nobody set it that way on purpose. Nobody set it at all.

This is the most expensive parameter in your stack that nobody owns. It compounds across the call tree: the summarizer above the verifier, the structured-output extractor below it, and the retry loop wrapping the whole thing all consume the planner's "be creative" knob as if it were a global. The bill arrives in three places at once — eval flakiness, token spend, and the half-day a senior engineer spends bisecting a regression that turns out to be no regression at all.

A PM walks into the AI feature review with a slide that reads "+12% engagement, +8% session length, retention up 3 points." The room nods. Two desks over, the support lead is staring at a different chart: tickets touching the AI surface are up 22%, and the most common resolution code is "user gave up, agent helped manually." Both numbers are real. Both come from the same product. The PM's dashboard is built on the assumption that the AI feature emits the same shape of event as the button it replaced. It doesn't. And the gap between what the dashboard counts and what the user experienced is where AI features quietly fail in plain sight.

The deterministic-feature playbook treats interaction as a click stream: user fires an event, the system reacts, the user moves on. AI features have a different event shape — a task arc with phases, retries, side trips to a human, and an offline judgment the telemetry never sees. Importing the deterministic dashboard onto that arc is the analytics equivalent of running 2018's interview loop against 2026's job. The numbers go up. The thing the numbers were supposed to predict goes down.

The on-call engineer pulls up a trace. The model returned, the span closed clean, the API call shows stop_reason: end_turn. By every signal the platform offers, this was a successful generation. Three minutes later a customer reports that the agent confidently wrote half a config file, declared the operation complete, and moved on. The trace had no warning sign because the warning sign isn't in the API contract — the provider's stop reason has four to seven buckets, and the question your incident demands an answer to lives in the gap between them.

Stop reasons are the field engineers reach for first during triage and the field that lies most cleanly when it does. The values are designed for a runtime that needs to decide what to do next: was this turn complete, did a tool get requested, did a budget get exceeded, did safety intervene. They are not designed for a human reconstructing why an answer went wrong, and the difference between those two purposes is where production teams burn entire afternoons.

The model is emitting JSON token by token. Your UI wants to render fields the moment they materialize — a confidence score before the long answer body, the arguments of a tool call as the model fills them in. Then someone wires up JSON.parse on every chunk and the whole thing falls over, because JSON.parse is all-or-nothing. It needs a balanced document to return anything. Until the model emits the closing brace, you have nothing to show.

This is not a parser problem you can fix with a try/catch. The standard JSON parser was designed against a content-length-known HTTP response. Partial input is not a state it models — it is "input error." When you treat a token stream as if it were an HTTP body, you inherit thirty years of "the document is either complete or invalid," and your UI pays the bill.

The moment your agent fans out five parallel tool calls — search across three indexes, query two databases, hit one external API — you have crossed an invisible line. You are no longer writing prompt-and-response code. You are writing a concurrent program. Most agent frameworks pretend you are not, and the bill arrives at 2 AM.

The pretense is comfortable. The planner emits a list of tool calls, the runtime fires them off, the runtime collects whatever comes back, the planner consumes the aggregate. From a thousand feet up it looks like a fan-out / fan-in pipeline, and most teams treat it that way until production teaches them otherwise. The problem is that twenty years of concurrent-programming research — partial-failure semantics, structured cancellation, backpressure, deterministic error attribution — already solved the failure modes you are about to rediscover. Your agent framework, by default, did not import any of it.

A team I talked to last quarter shipped a customer-support agent with the system prompt living in a Postgres row, one row per tenant. The pitch was sensible: enterprise customers had asked for tone customization, and "make the prompt editable" was the cheapest way to deliver it. Six months later, three things had happened. The eval suite had ballooned from 200 cases to 11,000 because every tenant's prompt now needed its own regression set. The prompt-update workflow had quietly become a write path with no review, because product owners had been given direct access to the table. And a single broken UTF-8 character in a Korean-language tenant prompt had taken that tenant's chatbot offline for two days before anyone noticed, because the deploy pipeline had no idea the prompt had changed.

None of these outcomes were forced by the requirements. They were forced by an architecture decision that nobody made deliberately: where does the system prompt live? In the code? In a config file? In a database row? The team picked "database" because it was the fastest path to a feature, and the consequences cascaded into every adjacent system over the following months.

The vendor said the upgrade was a drop-in replacement. The API contract held. The model name in your config barely changed. A week later, your context-window guard starts triggering on prompts it never tripped on before, your stop-sequence regex matches in the wrong place, and one of your few-shot examples started producing a confidently wrong answer that your eval suite happens not to cover. Nobody touched the prompt. Nobody touched the temperature. Somebody quietly retrained the tokenizer.

Tokenizer changes are the breaking change vendors don't call breaking. The API surface is byte-stable, the SDK didn't bump a major version, and the release notes mention "improved instruction following" — but the function from your input string to the integer sequence the model actually sees has been replaced. Every assumption your code made about how text becomes tokens is now subtly wrong. The cost of that invisibility is two weeks of "the model feels different" before someone re-runs a canonical prompt through count_tokens and finds the answer.

Most production AI features ship with a single confidence threshold. Above the line, the model answers. Below it, the user gets a flat "I'm not sure." That single number is doing two completely different jobs at once, and it's why your trust metric has been sliding for two quarters even though your accuracy on answered queries looks fine.

The right design has at least two cutoffs. An abstain threshold sits low: below it, the model declines because no answer is worth more than silence. An escalate threshold sits in the middle: between the two cutoffs, the system hands the case to a human reviewer instead of dropping it on the floor. Collapse them into a single dial and you ship a product that feels equally useless when it's wrong and when it's uncertain — which is the worst possible position to occupy in a market where users have a free alternative one tab away.

This isn't a new idea. The reject-option classifier literature has been arguing for split thresholds since the 1970s, distinguishing ambiguity rejects (the input is between known classes) from distance rejects (the input is far from any training data). Production AI teams keep rediscovering the same lesson the hard way, usually about six months after their first launch, when the support queue is full of people typing "is this thing broken or what."

Every team I have audited in the last six months claims to be vendor-agnostic. None of them are. The system prompt that scored highest on the eval suite did so because it leaned into a single vendor's tokenizer behavior, JSON-mode contract, refusal cadence, and stop-sequence handling — and the team that wrote it could not name which of those biases were doing the work. When the CFO asks why the cheaper model on the procurement deck cannot just be dropped in, the honest answer is two engineer-quarters of prompt re-tuning and a complete re-baseline of every eval. That is not a checkbox. It is a quarterly cost line.

The mental model that keeps biting teams is treating vendor portability as a one-time architecture decision. You add an abstraction layer, you write a model: field in your config, you congratulate yourself, and you move on. Then a year later the vendor raises prices, ships a deprecation notice, or has a bad week of refusals on a category you care about, and you discover that the abstraction was a thin wrapper around a prompt that only works on one model. The portability you bought was syntactic. The portability you needed was behavioral, and behavioral portability decays the moment you stop paying for it.

A vendor pushes a "minor refresh" to a model alias on a Tuesday afternoon. By Thursday, four customer companies are running incident response. None of them deployed code that week. None of their dashboards show a regression in latency, error rate, or any other infra-shaped metric. What changed is that the model behind their pinned alias started returning slightly different sentences, slightly different JSON, and slightly different refusals — and every prompt their team wrote against the old behavior is now a contract that nobody honored.

The asymmetry is the entire story. The provider treated the rollout as a deploy: tested internally, gated on a few aggregate evals, ramped to 100% within a maintenance window. The consumer surface received it as a semver violation: a dependency upgraded itself in production without changing its version string, and the bug reports started rolling in from end users with the cheerful subject line "nothing changed on our side."

At $50,000 a month, the "compute + tokens" line on your infra bill is rounding error. At $5,000,000 a month, it is a CFO question. The transition between those two states is not gradual — it is a phase change in how an organization talks about model spend, and most engineering orgs are unprepared for the social and political work that follows. The bill stays a single line; the conversation around it does not.

What changes is who has standing to ask "why." When three product teams share one API key and one capacity reservation, every quota argument has the same structure: someone is currently winning at the expense of someone else, and there is no neutral party to call it. The first time a team's launch is throttled because another team shipped a chatty agent, the absence of a governance body is felt by the entire engineering org at once. Calling a meeting and inventing a process under pressure is the worst time to design one.

There is a moment, six weeks into a serious LLM project, where the prompt iteration log starts to look like a therapy journal. Each tweak swaps one failure mode for another. Add a stricter "do not" clause and the model becomes evasive on cases it used to handle. Soften the tone and a different category of hallucination returns. The eval scoreboard hovers in a band three or four points wide, refusing to break out. Someone says, "let me try one more reordering," and another half day evaporates.

This is the local-maximum trap. The team is climbing a hill, but the hill does not go higher. The cruel part is that the hill is real — every prompt change does produce a measurable delta on some subset of cases, which is exactly the signal that keeps everyone tweaking. What's missing is the recognition that the ceiling above is not a prompt ceiling at all.