TypeSafe 创始人 Diogo Almeida 在社交平台上抛出一个难题:语言模型在聊天上超越人类已经好几年了,可自动化到底在哪?
这个难题他追了四年。在 OpenAI 期间,他参与搭建了让语言模型学会遵循指令、与人对话的方法,那项工作后来成了 ChatGPT 背后的研究。当时他以为聊天模型或许能通向 AGI,但热潮之下,他越来越清楚地看到,有一样很大的东西缺失了。
经过两年隐身开发、无数工艺难题和研究突破,TypeSafe AI 发布了首个 System One 模型。这是一类新的前沿模型,目标是做出软件可以直接使用的、快速且结构化的决策。
一条完全为自动化而建的工艺栈
TypeSafe 为此搭了一整套新栈:新的模型架构、追求最高质率的并行采样器,以及一种他们称为「校准决策强化学习」(RLCD)的训练方法。
首个公开模型叫 Jev,即日起开放早期访问。在 System One 类任务上,Jev 的智能水平与现有大语言模型相当,但速度快了两个数量级、效率也高出两个数量级。
代价是它放弃了字符串生成。Jev 针对结构化输出做了优化,并且不会产生幻觉。
用 Almeida 的说法,可以把 Jev 理解成一次前沿智能的函数调用:输入非结构化状态,输出带类型的概率化决策。
RLCD 与 RLHF、RLVR 的分野
要理解 Jev 的定位,得先看它和主流训练路线的差别。TypeSafe 把三条路线摆在一起做了比较:
- RLHF(基于人类反馈的强化学习):优化的是人类评分者偏好的文字表达和聊天回复。
- RLVR(可验证奖励强化学习):优化的是可以被程序化验证的输出。
- RLCD(校准决策强化学习):优化的是在 System One 任务上给出认识论上诚实的概率答案。
输入侧同样有区别。传统路线处理的是非结构化数据(例如文本),重点在连续的消息序列;Jev 处理的也是非结构化数据,但重点落在结构化的程序状态上。
输出侧的差异更直接。传统模型产出字符串,字符串很灵活,可以是聊天回复、代码,也可以是幻觉、拒答,甚至恰好是类型安全的结构化值。但要被软件使用,这些回复还得经过解析和校验,而且始终存在 AI 跑偏的风险。
Jev 直接产出类型安全的结构化值。可能的输出和结构都是预先定义好的,模型不会犯类型错误,所有答案都附带校准过的概率和置信度分数。
并行采样,一次查询出全部结果
生成方式上,传统模型是顺序的:一次生成一个 token,每个都以上一个为条件。Jev 是并行的:一次查询就生成全部输出,效率极高,并且对硬件友好。
售价差距是这套外观最直观的结果。现有模型的输入 token 售价从每百万 token 0.20 美元到 10 美元不等,输出 token 通常比输入贵约 5 倍。
Jev 的输入 token 是每百万 0.042 美元,也就是每十亿 token 42 美元;输出 token 免费——按 TypeSafe 的说法,是便宜到没必要收。
把这些数字放在一起看,售价和速度上的差距都相当可观。Almeida 自己也知道这话听起来有多夸张,所以他在公布时补了一句:非凡的主张需要非凡的证据,下面附上了凭据。
本文来源:https://www.163.com/dy/article/L6U0EL0B05561FZX.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
0
TypeSafe founder Diogo Almeida posed a challenge on social media: language models have surpassed humans in chat for years, but where is the automation?
He spent four years chasing this answer. At OpenAI, he helped build methods for language models to follow instructions and converse with people — work that later powered ChatGPT. Back then, he thought chat models might lead to AGI, but as the hype grew, he realized something major was missing.
After two years of stealth development, countless engineering hurdles, and research breakthroughs, TypeSafe AI has released its first System One model. This new class of frontier model aims to produce fast, structured decisions that software can directly use.
An entire stack built for automation
TypeSafe built a whole new stack for this: a novel model architecture, a parallel sampler optimized for speed and fidelity, and a training method called Relabeled Calibration Decision Reinforcement Learning (RLCD).
The first public model is called Jev, now open for early access. In System One tasks, Jev matches the intelligence of existing large language models but is two orders of magnitude faster and more efficient.
The trade-off is that it doesn’t generate strings. Jev is optimized for structured output and doesn’t hallucinate.
In Almeida’s words, think of Jev as a frontier-intelligence function call: you feed it unstructured state, and it spits out a typed, probabilistic decision.
RLCD vs. RLHF and RLVR
To understand Jev’s role, it’s useful to compare it with mainstream training paths. TypeSafe lays out three approaches:
RLHF (Reinforcement Learning from Human Feedback): Optimizes for text expression and chat replies that human raters prefer.
RLVR (Verifiable Reward Reinforcement Learning): Optimizes for outputs that can be validated by automated programs.
RLCD (Relabeled Calibration Decision RL): Optimizes for epistemically honest, probabilistic answers in System One tasks.
The input side also differs. Traditional paths handle unstructured data (e.g., text) with a focus on continuous message sequences. Jev handles unstructured data too, but its focus is structured program states.
The output side is the most critical difference. Traditional models produce strings — flexible but risky. Those strings can be chat replies, code, hallucinations, refusals, or even accidentally type-safe values. But to be used by software, they must be parsed, validated, and checked for drift.
Jev directly produces type-safe structured values. The space of possible outputs and their structure are predefined; the model cannot make type errors, and all answers come with calibrated probabilities and confidence scores.
Parallel sampling: all results in one query
The generation method also shifts. Traditional models generate tokens sequentially — one token at a time, conditioned on the previous one. Jev generates outputs in parallel: all at once, in a single query, making it extremely efficient and hardware-friendly.
The pricing gap is perhaps the most immediate visual result. Current model input tokens range from $0.20 to $10 per million tokens, with output tokens usually 5x more expensive.
Jev’s input tokens are $0.042 per million — that’s $42 per billion tokens. Output tokens are free, as TypeSafe puts it, “too cheap to bother charging.”
Combining these numbers, the differences in speed and cost are stark. Almeida acknowledges how bold these claims sound, which is why he includes a follow-up note: Extraordinary claims require extraordinary evidence. See below for receipts.
