由OpenAI前研究员Diogo Almeida创办的初创公司TypeSafe AI本周发布了一款名为Jev的新型AI模型1。与传统大语言模型不同，Jev基于Transformer架构但不生成文本，而是输出校准概率，专门用于软件自动化任务1。该模型因成本低、速度快且不会产生幻觉等特点获得开发者的广泛关注，甚至因需求过高一度导致API服务中断1。
Almeida曾是OpenAI的研究员，参与了ChatGPT的构建并发明了RLHF（强化学习来自人类反馈）技术1。他在两年前离开OpenAI创办了TypeSafe AI1。Jev模型采用了创新的训练方法——"从校准决策中的强化学习"，且仅使用合成数据进行训练1。Almeida表示这一决策是公司最好的赌注，称其"比我们的发布更好，比RLHF更好"1。
从性能表现看，Vercel工程师的测试显示Jev比ChatGPT快5到18倍，准确度更高1。在成本方面，Jev采用独特的计费模式——输出token免费，输入token按十亿计而非百万计1，Bryo AI的Nikhil Mudholkar测试中发现Jev比Gemini便宜10到20倍1。该模型以19世纪经济学家威廉·斯坦利·杰文斯的名字命名1。
TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, released a new artificial intelligence model called Jev this week that represents a fundamentally different approach to AI design.1 Rather than generating text like conventional large language models, Jev outputs calibrated probabilities and is specifically engineered for software automation tasks.1 The model has generated significant developer interest, with demand becoming so intense that it temporarily disrupted the company's API services.1
Almeida, who previously helped build ChatGPT and invented the RLHF (Reinforcement Learning from Human Feedback) technique at OpenAI, departed the company two years ago to establish TypeSafe AI.1 Jev operates on a distinctive cost structure that makes it substantially more economical than existing alternatives: while output tokens are provided free of charge, input tokens are priced per billion rather than per million—a critical distinction that dramatically reduces operational expenses.1 According to independent testing, Jev demonstrates significant performance advantages; Vercel engineers found it to be between 5 and 18 times faster than ChatGPT with improved accuracy, while developers at Bryo AI reported it costs 10 to 20 times less than competing solutions like Gemini.1
The model employs a novel training methodology called "reinforcement learning from calibrated decision making" and was trained exclusively on synthetic data.1 Almeida characterized this approach as the company's most promising strategic bet, stating it is "better than our release, better than RLHF."1 The model is named after 19th-century economist William Stanley Jevons.1
评论
还没有评论，欢迎留下第一条。