在杂货店后面干活有它自己的一套规则：如果这食物是香蕉，就送到农产品区；如果是曲奇饼干，就放到零食货架；如果是孜然，就把它和香料摆在一起。
但如果那批香蕉已经烂了、曲奇饼干碎了、孜然也结块了呢？这些规则与例外支配着每一家杂货店和社区小店。起初它们是僵化的；随着时间推移，人们会发现更多的例外：那是大蕉吗？迪拜巧克力：算甜点还是烘焙原料？
编码，字面意思就是把这些规则与例外写进软件里。在 AI 出现之前，这些程序是僵化的。AI 负责处理例外情况：一次图像搜索就能把这种不熟悉的水果识别为 Musa paradisiaca，香蕉的近亲。1
但我们并不需要世界上最聪明的模型来处理“如果是大蕉就归入农产品区”这种逻辑。
最新一波 AI 是一个强大的 if-then 决策器。Jev2 和 SemIf3 能在数百毫秒内回答这类问题，相比传统 AI 成本降低了 99%。4 它们简化了现有模型：只运行一次注意力计算，然后根据少量输出 token 确定每个允许答案的概率。5
我在自己的代码里寻找那些 if-then 语句，那些我交给 AI 处理的语句。几分钟之内，我就替换掉了其中一个智能体中大约四分之一的这类调用。在 98 条经过人工核验的生产环境邮件线程上，这些专用判定器几乎将生产模型的分类准确率翻了一倍，从 47% 跃升至 80% 以上。6
软件由各种原语构成：if-then 语句就是其中之一。通过优化这单一原语，我们看到了近两个数量级的成本降低，同时准确率还更高。
这些进展引出了一个问题：还有哪些编程原语会从同样的专门化中受益。它们也凸显了 AI 经济学的分化：用于探索的前沿模型与用于生产的优化模型。我们用最大的前沿模型来训练新模型，用能力最强的模型来设计系统架构。但一旦系统被工程化并经过加固，在工作流中运行它数千次或数百万次，则受益于更窄口径的 AI。
如果这是众多为生产而专门化的原语中的第一个，那么 harness 即将捕获更多的利润空间。7
-
Musa × paradisiaca。林奈在 1753 年将大蕉命名为 Musa paradisiaca，将甜香蕉命名为 Musa sapientum；如今已知两者都是同样两个亲本——Musa acuminata 与 Musa balbisiana——的杂交种。这一分类难到他把一个类群拆成了两个。↩︎
-
Jev，由 TypeSafe 于 2026 年 9 月 15 日发布。它通过一次并行处理返回带类型、经校准的决策，而非生成文本。↩︎
-
SemIf，作者 Theodore Lee，MIT 许可，于 2026 年 9 月 18 日发布，以及 kev、Jared Palmer：该接口模式的开放复现。基准数据为 SemIf 公布的数值，在 RTX 3090 上使用冻结的 Qwen3.5-4B 测得。↩︎
-
Jev 定价：每百万输入 token 收费 $0.042，输出 $0，TypeSafe 表示这可能受到补贴。Sonnet 级别的定价为每百万输入 $3、输出 $15。在一次典型的分类调用中，输入 2,000 token、输出 60 token，这便宜了 82 倍。TypeSafe 自身的工作流评测显示，Jev 每个案例成本为 $0.0004，而两个前沿模型分别为 $0.0304 和 $0.0836，差距达 76 倍至 209 倍。↩︎
-
SemIf 机制。该开源实现并非通过完整的 Transformer 堆栈自回归地生成 token，而是在提示词上运行注意力机制，直接从输出 logits 中评估候选选项，完全跳过多层前馈网络和解码步骤。↩︎
-
在 98 个经人工核验的生产环境邮件线程上进行的评估，2026 年 9 月：在完全相同的数据集上，生产环境的生成式 LLM 分类器达到了 47% 的准确率（46/98），而 Jev 得分 80%（78/98），本地 SemIf 得分 82%（80/98）。此外，在覆盖 31 封入站邮件的实时生产日志中，本地决策器对其中 8 封采取了行动且零错误，并将其余邮件安全地交由前沿模型处理。↩︎
-
Harness 的利润空间机遇。Harness 在不牺牲准确率的前提下削减执行成本；对底层模型原语进行专门化，会进一步扩大这一毛利润空间。↩︎
Working in the back of a grocery store has its own set of rules : if the food is a banana, send it to produce ; if it is a cookie, the snack aisle ; if it is cumin, shelve it with the spices.
But what if the load of bananas has spoiled, the cookies have crumbled & the cumin is caked? These rules & exceptions govern every grocery store & neighborhood mart. At the beginning they are rigid ; over time, more exceptions are discovered : is that a plantain? Dubai chocolate : dessert or baking supply?
Coding literally means encoding these rules & exceptions into software. Pre-AI, these programs were rigid. AI handles the exceptions : an image search identifies the unfamiliar fruit as Musa paradisiaca, a brother of the banana.1
But we do not need the world’s most brilliant model to handle if-plantain-then-produce logic.
The newest wave of AI is a robust if-then decider. Jev2 & SemIf3 answer questions like these in hundreds of milliseconds at a 99% reduction in cost compared to traditional AI.4 They simplify the existing models by running the attention math once & then determining the probability of each allowed answer, based on a few tokens of output.5
I went looking for the if-then statements in my own code, the ones I had handed to AI. Within a few minutes I had replaced about a quarter of those calls in one of my agents. On 98 hand-verified production email threads, the specialized deciders nearly doubled the classification accuracy of the production model, jumping from 47% to over 80%.6
Software is composed of primitives : the if-then statement is one of them. By optimizing this single primitive, we see nearly two orders of magnitude in cost reduction alongside higher accuracy.
These advances raise the question of which other programming primitives will benefit from the same specialization. They also highlight the bifurcating economics of AI : state of the art models for discovery & optimized models for production. We use the largest frontier models to train new models, & the most capable models to architect systems. But once a system is engineered & hardened, running it thousands or millions of times through a workflow benefits from narrower AI.
If this is the first of many primitives specialized for production, then harnesses are about to capture a lot more margin.7
-
Musa × paradisiaca. Linnaeus named plantains Musa paradisiaca & dessert bananas Musa sapientum in 1753 ; both are now known to be hybrids of the same two parents, Musa acuminata & Musa balbisiana. The classification was hard enough that he split one group into two. ↩︎
-
Jev, TypeSafe, released 15 September 2026. Returns typed, calibrated decisions in one parallel pass rather than generated text. ↩︎
-
SemIf, Theodore Lee, MIT licensed, released 18 September 2026, & kev, Jared Palmer : open reproductions of the interface pattern. Benchmarks are SemIf’s published figures on an RTX 3090 with a frozen Qwen3.5-4B. ↩︎
-
Jev pricing : $0.042 per million input tokens, $0 output, which TypeSafe says may be subsidized. Sonnet-class pricing is $3 input & $15 output per million. On a typical classification call of 2,000 input & 60 output tokens that is 82 times cheaper. TypeSafe’s own workflow evals put Jev at $0.0004 a case against $0.0304 & $0.0836 for two frontier models, a 76x to 209x spread. ↩︎
-
SemIf mechanics. Rather than generating tokens autoregressively through the full transformer stack, the open source implementation runs the attention mechanism over the prompt and evaluates candidate choices directly from the output logits, skipping the multi-layer feed-forward network and decode steps entirely. ↩︎
-
Evaluation on 98 hand-verified production email threads, September 2026 : across the exact same dataset, the production generative LLM classifier achieved 47% accuracy (46/98), while Jev scored 80% (78/98) & local SemIf scored 82% (80/98). In addition, in live production logs across 31 inbound emails, the local decider acted on 8 with zero errors & safely deferred the rest to the frontier model. ↩︎
-
The Harness Margin Opportunity. A harness cuts execution cost without sacrificing accuracy ; specializing the underlying model primitives expands that gross margin wedge even further. ↩︎