Skip to content
JevDirectory.org
Evaluation & Training

Sycophancy

The tendency of preference-trained models to say what people want to hear, one of the reasons TypeSafe chose a calibration objective.

Category
Evaluation & Training
Also known as
—
Related terms
3
Directory entries
2
Docs
docs.typesafe.ai
Added
2026-09-24

Definition

The primer lists sycophancy among the problems with RLHF: a reward built on human preference can teach a model to please rather than to be right, and to sound confident while being wrong.

An output can be compelling to a person without being reliable enough for unattended automation. That distinction is the core argument for training on outcomes and calibration instead.

Tagstraininglimitations

From the directory

Why TypeSafe trains decision models with RLCD instead of RLHF: calibrated probabilities where 0.2 outcomes happen about 20% of the time, and the case for machine-to-machine automation.
Sites & GuidesDocs#official#docs#evaluation
Official
TypeSafe's case for machine-native composable AI: intelligence as a dependable primitive that software branches on, rather than an assistant that keeps humans in the loop.
Sites & GuidesArticle#official#article#architecture
Official

From the community

Posts from builders shipping with Jev right now.

Follow @typesafeai

The compaction plugin, in Chinese

连 JEV 的联合创始人 @CompleteSkeptic 都亲自下场转发点赞:这是 JEV 正确的打开方式! 大家在 Claude Code 里写长代码,最恶心的就是上下文一满,系统就卡住 10 秒去写总结小作文 开发者 @tamarajtran 刚刚开源了这个插件:利用 JEV Show more

tamara
tamara
@tamarajtran

found the perfect use case for @typesafeai Jev: instant compaction in 2026, why is compaction still a summarization prompt? Jev can make it instant by scoring every tool call and dropping what’s irrelevant

Reply

WTF is Jev, and 9 things people are building

TL;DR of my new article: WTF is Jev by @typesafeai, and the 9 things people are already building with it. The thesis: 𝗮 𝗰𝗼-𝗰𝗿𝗲𝗮𝘁𝗼𝗿 𝗼𝗳 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝘀𝗽𝗲𝗻𝘁 𝘁𝘄𝗼 𝘆𝗲𝗮𝗿𝘀 𝗶𝗻 𝘀𝘁𝗲𝗮𝗹𝘁𝗵 𝗼𝗻 𝗮 𝗺𝗼𝗱𝗲𝗹 𝘁𝗵𝗮𝘁 𝗰𝗮𝗻𝗻𝗼𝘁 𝘄𝗿𝗶𝘁𝗲 𝗮 Show more

Matt Van Horn
Matt Van Horn
@mvanhorn

x.com/i/article/2100…

Reply

Voice-controlled computer use on a Mac

The use cases Jev unlocks

19 open-source Jev projects

一个只会做选择题的小模型,能长出多少玩法? TypeSafe 的 Jev 就是这么个东西:给它最多 255 个选项,几百毫秒挑一个,再附一个校准过的置信度,输出 token 免费 结果开发者们已经拿它做出了 19 个开源项目,加起来 6800 多星 挑几个最有意思的 jev-ultrafast,browser-use 出品,2700 星,一句 Show more

Reply

Jev Calc: a notebook that calculates anything